2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
42 #include "types/commrec.h"
44 #include "gromacs/math/utilities.h"
45 #include "gromacs/math/vec.h"
46 #include "nbnxn_consts.h"
47 /* nbnxn_internal.h included gromacs/simd/macros.h */
48 #include "nbnxn_internal.h"
50 #include "gromacs/simd/vector_operations.h"
52 #include "nbnxn_atomdata.h"
53 #include "nbnxn_search.h"
54 #include "gmx_omp_nthreads.h"
58 #include "gromacs/pbcutil/ishift.h"
59 #include "gromacs/pbcutil/pbc.h"
60 #include "gromacs/utility/smalloc.h"
62 #ifdef NBNXN_SEARCH_BB_SIMD4
63 /* Always use 4-wide SIMD for bounding box calculations */
66 /* Single precision BBs + coordinates, we can also load coordinates with SIMD */
67 # define NBNXN_SEARCH_SIMD4_FLOAT_X_BB
70 # if defined NBNXN_SEARCH_SIMD4_FLOAT_X_BB && (GPU_NSUBCELL == 4 || GPU_NSUBCELL == 8)
71 /* Store bounding boxes with x, y and z coordinates in packs of 4 */
72 # define NBNXN_PBB_SIMD4
75 /* The packed bounding box coordinate stride is always set to 4.
76 * With AVX we could use 8, but that turns out not to be faster.
79 # define STRIDE_PBB_2LOG 2
81 #endif /* NBNXN_SEARCH_BB_SIMD4 */
85 /* The functions below are macros as they are performance sensitive */
87 /* 4x4 list, pack=4: no complex conversion required */
88 /* i-cluster to j-cluster conversion */
89 #define CI_TO_CJ_J4(ci) (ci)
90 /* cluster index to coordinate array index conversion */
91 #define X_IND_CI_J4(ci) ((ci)*STRIDE_P4)
92 #define X_IND_CJ_J4(cj) ((cj)*STRIDE_P4)
94 /* 4x2 list, pack=4: j-cluster size is half the packing width */
95 /* i-cluster to j-cluster conversion */
96 #define CI_TO_CJ_J2(ci) ((ci)<<1)
97 /* cluster index to coordinate array index conversion */
98 #define X_IND_CI_J2(ci) ((ci)*STRIDE_P4)
99 #define X_IND_CJ_J2(cj) (((cj)>>1)*STRIDE_P4 + ((cj) & 1)*(PACK_X4>>1))
101 /* 4x8 list, pack=8: i-cluster size is half the packing width */
102 /* i-cluster to j-cluster conversion */
103 #define CI_TO_CJ_J8(ci) ((ci)>>1)
104 /* cluster index to coordinate array index conversion */
105 #define X_IND_CI_J8(ci) (((ci)>>1)*STRIDE_P8 + ((ci) & 1)*(PACK_X8>>1))
106 #define X_IND_CJ_J8(cj) ((cj)*STRIDE_P8)
108 /* The j-cluster size is matched to the SIMD width */
109 #if GMX_SIMD_REAL_WIDTH == 2
110 #define CI_TO_CJ_SIMD_4XN(ci) CI_TO_CJ_J2(ci)
111 #define X_IND_CI_SIMD_4XN(ci) X_IND_CI_J2(ci)
112 #define X_IND_CJ_SIMD_4XN(cj) X_IND_CJ_J2(cj)
114 #if GMX_SIMD_REAL_WIDTH == 4
115 #define CI_TO_CJ_SIMD_4XN(ci) CI_TO_CJ_J4(ci)
116 #define X_IND_CI_SIMD_4XN(ci) X_IND_CI_J4(ci)
117 #define X_IND_CJ_SIMD_4XN(cj) X_IND_CJ_J4(cj)
119 #if GMX_SIMD_REAL_WIDTH == 8
120 #define CI_TO_CJ_SIMD_4XN(ci) CI_TO_CJ_J8(ci)
121 #define X_IND_CI_SIMD_4XN(ci) X_IND_CI_J8(ci)
122 #define X_IND_CJ_SIMD_4XN(cj) X_IND_CJ_J8(cj)
123 /* Half SIMD with j-cluster size */
124 #define CI_TO_CJ_SIMD_2XNN(ci) CI_TO_CJ_J4(ci)
125 #define X_IND_CI_SIMD_2XNN(ci) X_IND_CI_J4(ci)
126 #define X_IND_CJ_SIMD_2XNN(cj) X_IND_CJ_J4(cj)
128 #if GMX_SIMD_REAL_WIDTH == 16
129 #define CI_TO_CJ_SIMD_2XNN(ci) CI_TO_CJ_J8(ci)
130 #define X_IND_CI_SIMD_2XNN(ci) X_IND_CI_J8(ci)
131 #define X_IND_CJ_SIMD_2XNN(cj) X_IND_CJ_J8(cj)
133 #error "unsupported GMX_SIMD_REAL_WIDTH"
139 #endif /* GMX_NBNXN_SIMD */
142 #ifdef NBNXN_SEARCH_BB_SIMD4
143 /* Store bounding boxes corners as quadruplets: xxxxyyyyzzzz */
145 /* Size of bounding box corners quadruplet */
146 #define NNBSBB_XXXX (NNBSBB_D*DIM*STRIDE_PBB)
149 /* We shift the i-particles backward for PBC.
150 * This leads to more conditionals than shifting forward.
151 * We do this to get more balanced pair lists.
153 #define NBNXN_SHIFT_BACKWARD
156 /* This define is a lazy way to avoid interdependence of the grid
157 * and searching data structures.
159 #define NBNXN_NA_SC_MAX (GPU_NSUBCELL*NBNXN_GPU_CLUSTER_SIZE)
162 static void nbs_cycle_clear(nbnxn_cycle_t *cc)
166 for (i = 0; i < enbsCCnr; i++)
173 static double Mcyc_av(const nbnxn_cycle_t *cc)
175 return (double)cc->c*1e-6/cc->count;
178 static void nbs_cycle_print(FILE *fp, const nbnxn_search_t nbs)
184 fprintf(fp, "ns %4d grid %4.1f search %4.1f red.f %5.3f",
185 nbs->cc[enbsCCgrid].count,
186 Mcyc_av(&nbs->cc[enbsCCgrid]),
187 Mcyc_av(&nbs->cc[enbsCCsearch]),
188 Mcyc_av(&nbs->cc[enbsCCreducef]));
190 if (nbs->nthread_max > 1)
192 if (nbs->cc[enbsCCcombine].count > 0)
194 fprintf(fp, " comb %5.2f",
195 Mcyc_av(&nbs->cc[enbsCCcombine]));
197 fprintf(fp, " s. th");
198 for (t = 0; t < nbs->nthread_max; t++)
200 fprintf(fp, " %4.1f",
201 Mcyc_av(&nbs->work[t].cc[enbsCCsearch]));
207 static void nbnxn_grid_init(nbnxn_grid_t * grid)
210 grid->cxy_ind = NULL;
211 grid->cxy_nalloc = 0;
217 static int get_2log(int n)
222 while ((1<<log2) < n)
228 gmx_fatal(FARGS, "nbnxn na_c (%d) is not a power of 2", n);
234 static int nbnxn_kernel_to_ci_size(int nb_kernel_type)
236 switch (nb_kernel_type)
238 case nbnxnk4x4_PlainC:
239 case nbnxnk4xN_SIMD_4xN:
240 case nbnxnk4xN_SIMD_2xNN:
241 return NBNXN_CPU_CLUSTER_I_SIZE;
242 case nbnxnk8x8x8_CUDA:
243 case nbnxnk8x8x8_PlainC:
244 /* The cluster size for super/sub lists is only set here.
245 * Any value should work for the pair-search and atomdata code.
246 * The kernels, of course, might require a particular value.
248 return NBNXN_GPU_CLUSTER_SIZE;
250 gmx_incons("unknown kernel type");
256 int nbnxn_kernel_to_cj_size(int nb_kernel_type)
258 int nbnxn_simd_width = 0;
261 #ifdef GMX_NBNXN_SIMD
262 nbnxn_simd_width = GMX_SIMD_REAL_WIDTH;
265 switch (nb_kernel_type)
267 case nbnxnk4x4_PlainC:
268 cj_size = NBNXN_CPU_CLUSTER_I_SIZE;
270 case nbnxnk4xN_SIMD_4xN:
271 cj_size = nbnxn_simd_width;
273 case nbnxnk4xN_SIMD_2xNN:
274 cj_size = nbnxn_simd_width/2;
276 case nbnxnk8x8x8_CUDA:
277 case nbnxnk8x8x8_PlainC:
278 cj_size = nbnxn_kernel_to_ci_size(nb_kernel_type);
281 gmx_incons("unknown kernel type");
287 static int ci_to_cj(int na_cj_2log, int ci)
291 case 2: return ci; break;
292 case 1: return (ci<<1); break;
293 case 3: return (ci>>1); break;
299 gmx_bool nbnxn_kernel_pairlist_simple(int nb_kernel_type)
301 if (nb_kernel_type == nbnxnkNotSet)
303 gmx_fatal(FARGS, "Non-bonded kernel type not set for Verlet-style pair-list.");
306 switch (nb_kernel_type)
308 case nbnxnk8x8x8_CUDA:
309 case nbnxnk8x8x8_PlainC:
312 case nbnxnk4x4_PlainC:
313 case nbnxnk4xN_SIMD_4xN:
314 case nbnxnk4xN_SIMD_2xNN:
318 gmx_incons("Invalid nonbonded kernel type passed!");
323 /* Initializes a single nbnxn_pairlist_t data structure */
324 static void nbnxn_init_pairlist_fep(t_nblist *nl)
326 nl->type = GMX_NBLIST_INTERACTION_FREE_ENERGY;
327 nl->igeometry = GMX_NBLIST_GEOMETRY_PARTICLE_PARTICLE;
328 /* The interaction functions are set in the free energy kernel fuction */
347 void nbnxn_init_search(nbnxn_search_t * nbs_ptr,
349 gmx_domdec_zones_t *zones,
361 nbs->DomDec = (n_dd_cells != NULL);
363 clear_ivec(nbs->dd_dim);
369 for (d = 0; d < DIM; d++)
371 if ((*n_dd_cells)[d] > 1)
374 /* Each grid matches a DD zone */
380 snew(nbs->grid, nbs->ngrid);
381 for (g = 0; g < nbs->ngrid; g++)
383 nbnxn_grid_init(&nbs->grid[g]);
386 nbs->cell_nalloc = 0;
390 nbs->nthread_max = nthread_max;
392 /* Initialize the work data structures for each thread */
393 snew(nbs->work, nbs->nthread_max);
394 for (t = 0; t < nbs->nthread_max; t++)
396 nbs->work[t].cxy_na = NULL;
397 nbs->work[t].cxy_na_nalloc = 0;
398 nbs->work[t].sort_work = NULL;
399 nbs->work[t].sort_work_nalloc = 0;
401 snew(nbs->work[t].nbl_fep, 1);
402 nbnxn_init_pairlist_fep(nbs->work[t].nbl_fep);
405 /* Initialize detailed nbsearch cycle counting */
406 nbs->print_cycles = (getenv("GMX_NBNXN_CYCLE") != 0);
407 nbs->search_count = 0;
408 nbs_cycle_clear(nbs->cc);
409 for (t = 0; t < nbs->nthread_max; t++)
411 nbs_cycle_clear(nbs->work[t].cc);
415 static real grid_atom_density(int n, rvec corner0, rvec corner1)
421 /* To avoid zero density we use a minimum of 1 atom */
425 rvec_sub(corner1, corner0, size);
427 return n/(size[XX]*size[YY]*size[ZZ]);
430 static int set_grid_size_xy(const nbnxn_search_t nbs,
433 int n, rvec corner0, rvec corner1,
438 real adens, tlen, tlen_x, tlen_y, nc_max;
441 rvec_sub(corner1, corner0, size);
445 assert(atom_density > 0);
447 /* target cell length */
450 /* To minimize the zero interactions, we should make
451 * the largest of the i/j cell cubic.
453 na_c = max(grid->na_c, grid->na_cj);
455 /* Approximately cubic cells */
456 tlen = pow(na_c/atom_density, 1.0/3.0);
462 /* Approximately cubic sub cells */
463 tlen = pow(grid->na_c/atom_density, 1.0/3.0);
464 tlen_x = tlen*GPU_NSUBCELL_X;
465 tlen_y = tlen*GPU_NSUBCELL_Y;
467 /* We round ncx and ncy down, because we get less cell pairs
468 * in the nbsist when the fixed cell dimensions (x,y) are
469 * larger than the variable one (z) than the other way around.
471 grid->ncx = max(1, (int)(size[XX]/tlen_x));
472 grid->ncy = max(1, (int)(size[YY]/tlen_y));
480 grid->sx = size[XX]/grid->ncx;
481 grid->sy = size[YY]/grid->ncy;
482 grid->inv_sx = 1/grid->sx;
483 grid->inv_sy = 1/grid->sy;
487 /* This is a non-home zone, add an extra row of cells
488 * for particles communicated for bonded interactions.
489 * These can be beyond the cut-off. It doesn't matter where
490 * they end up on the grid, but for performance it's better
491 * if they don't end up in cells that can be within cut-off range.
497 /* We need one additional cell entry for particles moved by DD */
498 if (grid->ncx*grid->ncy+1 > grid->cxy_nalloc)
500 grid->cxy_nalloc = over_alloc_large(grid->ncx*grid->ncy+1);
501 srenew(grid->cxy_na, grid->cxy_nalloc);
502 srenew(grid->cxy_ind, grid->cxy_nalloc+1);
504 for (t = 0; t < nbs->nthread_max; t++)
506 if (grid->ncx*grid->ncy+1 > nbs->work[t].cxy_na_nalloc)
508 nbs->work[t].cxy_na_nalloc = over_alloc_large(grid->ncx*grid->ncy+1);
509 srenew(nbs->work[t].cxy_na, nbs->work[t].cxy_na_nalloc);
513 /* Worst case scenario of 1 atom in each last cell */
514 if (grid->na_cj <= grid->na_c)
516 nc_max = n/grid->na_sc + grid->ncx*grid->ncy;
520 nc_max = n/grid->na_sc + grid->ncx*grid->ncy*grid->na_cj/grid->na_c;
523 if (nc_max > grid->nc_nalloc)
525 grid->nc_nalloc = over_alloc_large(nc_max);
526 srenew(grid->nsubc, grid->nc_nalloc);
527 srenew(grid->bbcz, grid->nc_nalloc*NNBSBB_D);
529 sfree_aligned(grid->bb);
530 /* This snew also zeros the contents, this avoid possible
531 * floating exceptions in SIMD with the unused bb elements.
535 snew_aligned(grid->bb, grid->nc_nalloc, 16);
542 pbb_nalloc = grid->nc_nalloc*GPU_NSUBCELL/STRIDE_PBB*NNBSBB_XXXX;
543 snew_aligned(grid->pbb, pbb_nalloc, 16);
545 snew_aligned(grid->bb, grid->nc_nalloc*GPU_NSUBCELL, 16);
551 if (grid->na_cj == grid->na_c)
553 grid->bbj = grid->bb;
557 sfree_aligned(grid->bbj);
558 snew_aligned(grid->bbj, grid->nc_nalloc*grid->na_c/grid->na_cj, 16);
562 srenew(grid->flags, grid->nc_nalloc);
565 srenew(grid->fep, grid->nc_nalloc*grid->na_sc/grid->na_c);
569 copy_rvec(corner0, grid->c0);
570 copy_rvec(corner1, grid->c1);
575 /* We need to sort paricles in grid columns on z-coordinate.
576 * As particle are very often distributed homogeneously, we a sorting
577 * algorithm similar to pigeonhole sort. We multiply the z-coordinate
578 * by a factor, cast to an int and try to store in that hole. If the hole
579 * is full, we move this or another particle. A second pass is needed to make
580 * contiguous elements. SORT_GRID_OVERSIZE is the ratio of holes to particles.
581 * 4 is the optimal value for homogeneous particle distribution and allows
582 * for an O(#particles) sort up till distributions were all particles are
583 * concentrated in 1/4 of the space. No NlogN fallback is implemented,
584 * as it can be expensive to detect imhomogeneous particle distributions.
585 * SGSF is the maximum ratio of holes used, in the worst case all particles
586 * end up in the last hole and we need #particles extra holes at the end.
588 #define SORT_GRID_OVERSIZE 4
589 #define SGSF (SORT_GRID_OVERSIZE + 1)
591 /* Sort particle index a on coordinates x along dim.
592 * Backwards tells if we want decreasing iso increasing coordinates.
593 * h0 is the minimum of the coordinate range.
594 * invh is the 1/length of the sorting range.
595 * n_per_h (>=n) is the expected average number of particles per 1/invh
596 * sort is the sorting work array.
597 * sort should have a size of at least n_per_h*SORT_GRID_OVERSIZE + n,
598 * or easier, allocate at least n*SGSF elements.
600 static void sort_atoms(int dim, gmx_bool Backwards,
601 int gmx_unused dd_zone,
602 int *a, int n, rvec *x,
603 real h0, real invh, int n_per_h,
607 int zi, zim, zi_min, zi_max;
619 gmx_incons("n > n_per_h");
623 /* Transform the inverse range height into the inverse hole height */
624 invh *= n_per_h*SORT_GRID_OVERSIZE;
626 /* Set nsort to the maximum possible number of holes used.
627 * In worst case all n elements end up in the last bin.
629 nsort = n_per_h*SORT_GRID_OVERSIZE + n;
631 /* Determine the index range used, so we can limit it for the second pass */
635 /* Sort the particles using a simple index sort */
636 for (i = 0; i < n; i++)
638 /* The cast takes care of float-point rounding effects below zero.
639 * This code assumes particles are less than 1/SORT_GRID_OVERSIZE
640 * times the box height out of the box.
642 zi = (int)((x[a[i]][dim] - h0)*invh);
645 /* As we can have rounding effect, we use > iso >= here */
646 if (zi < 0 || (dd_zone == 0 && zi > n_per_h*SORT_GRID_OVERSIZE))
648 gmx_fatal(FARGS, "(int)((x[%d][%c]=%f - %f)*%f) = %d, not in 0 - %d*%d\n",
649 a[i], 'x'+dim, x[a[i]][dim], h0, invh, zi,
650 n_per_h, SORT_GRID_OVERSIZE);
654 /* In a non-local domain, particles communcated for bonded interactions
655 * can be far beyond the grid size, which is set by the non-bonded
656 * cut-off distance. We sort such particles into the last cell.
658 if (zi > n_per_h*SORT_GRID_OVERSIZE)
660 zi = n_per_h*SORT_GRID_OVERSIZE;
663 /* Ideally this particle should go in sort cell zi,
664 * but that might already be in use,
665 * in that case find the first empty cell higher up
670 zi_min = min(zi_min, zi);
671 zi_max = max(zi_max, zi);
675 /* We have multiple atoms in the same sorting slot.
676 * Sort on real z for minimal bounding box size.
677 * There is an extra check for identical z to ensure
678 * well-defined output order, independent of input order
679 * to ensure binary reproducibility after restarts.
681 while (sort[zi] >= 0 && ( x[a[i]][dim] > x[sort[zi]][dim] ||
682 (x[a[i]][dim] == x[sort[zi]][dim] &&
690 /* Shift all elements by one slot until we find an empty slot */
693 while (sort[zim] >= 0)
701 zi_max = max(zi_max, zim);
704 zi_max = max(zi_max, zi);
711 for (zi = 0; zi < nsort; zi++)
722 for (zi = zi_max; zi >= zi_min; zi--)
733 gmx_incons("Lost particles while sorting");
738 #define R2F_D(x) ((float)((x) >= 0 ? ((1-GMX_FLOAT_EPS)*(x)) : ((1+GMX_FLOAT_EPS)*(x))))
739 #define R2F_U(x) ((float)((x) >= 0 ? ((1+GMX_FLOAT_EPS)*(x)) : ((1-GMX_FLOAT_EPS)*(x))))
745 /* Coordinate order x,y,z, bb order xyz0 */
746 static void calc_bounding_box(int na, int stride, const real *x, nbnxn_bb_t *bb)
749 real xl, xh, yl, yh, zl, zh;
759 for (j = 1; j < na; j++)
761 xl = min(xl, x[i+XX]);
762 xh = max(xh, x[i+XX]);
763 yl = min(yl, x[i+YY]);
764 yh = max(yh, x[i+YY]);
765 zl = min(zl, x[i+ZZ]);
766 zh = max(zh, x[i+ZZ]);
769 /* Note: possible double to float conversion here */
770 bb->lower[BB_X] = R2F_D(xl);
771 bb->lower[BB_Y] = R2F_D(yl);
772 bb->lower[BB_Z] = R2F_D(zl);
773 bb->upper[BB_X] = R2F_U(xh);
774 bb->upper[BB_Y] = R2F_U(yh);
775 bb->upper[BB_Z] = R2F_U(zh);
778 /* Packed coordinates, bb order xyz0 */
779 static void calc_bounding_box_x_x4(int na, const real *x, nbnxn_bb_t *bb)
782 real xl, xh, yl, yh, zl, zh;
790 for (j = 1; j < na; j++)
792 xl = min(xl, x[j+XX*PACK_X4]);
793 xh = max(xh, x[j+XX*PACK_X4]);
794 yl = min(yl, x[j+YY*PACK_X4]);
795 yh = max(yh, x[j+YY*PACK_X4]);
796 zl = min(zl, x[j+ZZ*PACK_X4]);
797 zh = max(zh, x[j+ZZ*PACK_X4]);
799 /* Note: possible double to float conversion here */
800 bb->lower[BB_X] = R2F_D(xl);
801 bb->lower[BB_Y] = R2F_D(yl);
802 bb->lower[BB_Z] = R2F_D(zl);
803 bb->upper[BB_X] = R2F_U(xh);
804 bb->upper[BB_Y] = R2F_U(yh);
805 bb->upper[BB_Z] = R2F_U(zh);
808 /* Packed coordinates, bb order xyz0 */
809 static void calc_bounding_box_x_x8(int na, const real *x, nbnxn_bb_t *bb)
812 real xl, xh, yl, yh, zl, zh;
820 for (j = 1; j < na; j++)
822 xl = min(xl, x[j+XX*PACK_X8]);
823 xh = max(xh, x[j+XX*PACK_X8]);
824 yl = min(yl, x[j+YY*PACK_X8]);
825 yh = max(yh, x[j+YY*PACK_X8]);
826 zl = min(zl, x[j+ZZ*PACK_X8]);
827 zh = max(zh, x[j+ZZ*PACK_X8]);
829 /* Note: possible double to float conversion here */
830 bb->lower[BB_X] = R2F_D(xl);
831 bb->lower[BB_Y] = R2F_D(yl);
832 bb->lower[BB_Z] = R2F_D(zl);
833 bb->upper[BB_X] = R2F_U(xh);
834 bb->upper[BB_Y] = R2F_U(yh);
835 bb->upper[BB_Z] = R2F_U(zh);
838 /* Packed coordinates, bb order xyz0 */
839 static void calc_bounding_box_x_x4_halves(int na, const real *x,
840 nbnxn_bb_t *bb, nbnxn_bb_t *bbj)
842 calc_bounding_box_x_x4(min(na, 2), x, bbj);
846 calc_bounding_box_x_x4(min(na-2, 2), x+(PACK_X4>>1), bbj+1);
850 /* Set the "empty" bounding box to the same as the first one,
851 * so we don't need to treat special cases in the rest of the code.
853 #ifdef NBNXN_SEARCH_BB_SIMD4
854 gmx_simd4_store_f(&bbj[1].lower[0], gmx_simd4_load_f(&bbj[0].lower[0]));
855 gmx_simd4_store_f(&bbj[1].upper[0], gmx_simd4_load_f(&bbj[0].upper[0]));
861 #ifdef NBNXN_SEARCH_BB_SIMD4
862 gmx_simd4_store_f(&bb->lower[0],
863 gmx_simd4_min_f(gmx_simd4_load_f(&bbj[0].lower[0]),
864 gmx_simd4_load_f(&bbj[1].lower[0])));
865 gmx_simd4_store_f(&bb->upper[0],
866 gmx_simd4_max_f(gmx_simd4_load_f(&bbj[0].upper[0]),
867 gmx_simd4_load_f(&bbj[1].upper[0])));
872 for (i = 0; i < NNBSBB_C; i++)
874 bb->lower[i] = min(bbj[0].lower[i], bbj[1].lower[i]);
875 bb->upper[i] = max(bbj[0].upper[i], bbj[1].upper[i]);
881 #ifdef NBNXN_SEARCH_BB_SIMD4
883 /* Coordinate order xyz, bb order xxxxyyyyzzzz */
884 static void calc_bounding_box_xxxx(int na, int stride, const real *x, float *bb)
887 real xl, xh, yl, yh, zl, zh;
897 for (j = 1; j < na; j++)
899 xl = min(xl, x[i+XX]);
900 xh = max(xh, x[i+XX]);
901 yl = min(yl, x[i+YY]);
902 yh = max(yh, x[i+YY]);
903 zl = min(zl, x[i+ZZ]);
904 zh = max(zh, x[i+ZZ]);
907 /* Note: possible double to float conversion here */
908 bb[0*STRIDE_PBB] = R2F_D(xl);
909 bb[1*STRIDE_PBB] = R2F_D(yl);
910 bb[2*STRIDE_PBB] = R2F_D(zl);
911 bb[3*STRIDE_PBB] = R2F_U(xh);
912 bb[4*STRIDE_PBB] = R2F_U(yh);
913 bb[5*STRIDE_PBB] = R2F_U(zh);
916 #endif /* NBNXN_SEARCH_BB_SIMD4 */
918 #ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
920 /* Coordinate order xyz?, bb order xyz0 */
921 static void calc_bounding_box_simd4(int na, const float *x, nbnxn_bb_t *bb)
923 gmx_simd4_float_t bb_0_S, bb_1_S;
924 gmx_simd4_float_t x_S;
928 bb_0_S = gmx_simd4_load_f(x);
931 for (i = 1; i < na; i++)
933 x_S = gmx_simd4_load_f(x+i*NNBSBB_C);
934 bb_0_S = gmx_simd4_min_f(bb_0_S, x_S);
935 bb_1_S = gmx_simd4_max_f(bb_1_S, x_S);
938 gmx_simd4_store_f(&bb->lower[0], bb_0_S);
939 gmx_simd4_store_f(&bb->upper[0], bb_1_S);
942 /* Coordinate order xyz?, bb order xxxxyyyyzzzz */
943 static void calc_bounding_box_xxxx_simd4(int na, const float *x,
944 nbnxn_bb_t *bb_work_aligned,
947 calc_bounding_box_simd4(na, x, bb_work_aligned);
949 bb[0*STRIDE_PBB] = bb_work_aligned->lower[BB_X];
950 bb[1*STRIDE_PBB] = bb_work_aligned->lower[BB_Y];
951 bb[2*STRIDE_PBB] = bb_work_aligned->lower[BB_Z];
952 bb[3*STRIDE_PBB] = bb_work_aligned->upper[BB_X];
953 bb[4*STRIDE_PBB] = bb_work_aligned->upper[BB_Y];
954 bb[5*STRIDE_PBB] = bb_work_aligned->upper[BB_Z];
957 #endif /* NBNXN_SEARCH_SIMD4_FLOAT_X_BB */
960 /* Combines pairs of consecutive bounding boxes */
961 static void combine_bounding_box_pairs(nbnxn_grid_t *grid, const nbnxn_bb_t *bb)
963 int i, j, sc2, nc2, c2;
965 for (i = 0; i < grid->ncx*grid->ncy; i++)
967 /* Starting bb in a column is expected to be 2-aligned */
968 sc2 = grid->cxy_ind[i]>>1;
969 /* For odd numbers skip the last bb here */
970 nc2 = (grid->cxy_na[i]+3)>>(2+1);
971 for (c2 = sc2; c2 < sc2+nc2; c2++)
973 #ifdef NBNXN_SEARCH_BB_SIMD4
974 gmx_simd4_float_t min_S, max_S;
976 min_S = gmx_simd4_min_f(gmx_simd4_load_f(&bb[c2*2+0].lower[0]),
977 gmx_simd4_load_f(&bb[c2*2+1].lower[0]));
978 max_S = gmx_simd4_max_f(gmx_simd4_load_f(&bb[c2*2+0].upper[0]),
979 gmx_simd4_load_f(&bb[c2*2+1].upper[0]));
980 gmx_simd4_store_f(&grid->bbj[c2].lower[0], min_S);
981 gmx_simd4_store_f(&grid->bbj[c2].upper[0], max_S);
983 for (j = 0; j < NNBSBB_C; j++)
985 grid->bbj[c2].lower[j] = min(bb[c2*2+0].lower[j],
986 bb[c2*2+1].lower[j]);
987 grid->bbj[c2].upper[j] = max(bb[c2*2+0].upper[j],
988 bb[c2*2+1].upper[j]);
992 if (((grid->cxy_na[i]+3)>>2) & 1)
994 /* The bb count in this column is odd: duplicate the last bb */
995 for (j = 0; j < NNBSBB_C; j++)
997 grid->bbj[c2].lower[j] = bb[c2*2].lower[j];
998 grid->bbj[c2].upper[j] = bb[c2*2].upper[j];
1005 /* Prints the average bb size, used for debug output */
1006 static void print_bbsizes_simple(FILE *fp,
1007 const nbnxn_search_t nbs,
1008 const nbnxn_grid_t *grid)
1014 for (c = 0; c < grid->nc; c++)
1016 for (d = 0; d < DIM; d++)
1018 ba[d] += grid->bb[c].upper[d] - grid->bb[c].lower[d];
1021 dsvmul(1.0/grid->nc, ba, ba);
1023 fprintf(fp, "ns bb: %4.2f %4.2f %4.2f %4.2f %4.2f %4.2f rel %4.2f %4.2f %4.2f\n",
1024 nbs->box[XX][XX]/grid->ncx,
1025 nbs->box[YY][YY]/grid->ncy,
1026 nbs->box[ZZ][ZZ]*grid->ncx*grid->ncy/grid->nc,
1027 ba[XX], ba[YY], ba[ZZ],
1028 ba[XX]*grid->ncx/nbs->box[XX][XX],
1029 ba[YY]*grid->ncy/nbs->box[YY][YY],
1030 ba[ZZ]*grid->nc/(grid->ncx*grid->ncy*nbs->box[ZZ][ZZ]));
1033 /* Prints the average bb size, used for debug output */
1034 static void print_bbsizes_supersub(FILE *fp,
1035 const nbnxn_search_t nbs,
1036 const nbnxn_grid_t *grid)
1043 for (c = 0; c < grid->nc; c++)
1046 for (s = 0; s < grid->nsubc[c]; s += STRIDE_PBB)
1050 cs_w = (c*GPU_NSUBCELL + s)/STRIDE_PBB;
1051 for (i = 0; i < STRIDE_PBB; i++)
1053 for (d = 0; d < DIM; d++)
1056 grid->pbb[cs_w*NNBSBB_XXXX+(DIM+d)*STRIDE_PBB+i] -
1057 grid->pbb[cs_w*NNBSBB_XXXX+ d *STRIDE_PBB+i];
1062 for (s = 0; s < grid->nsubc[c]; s++)
1066 cs = c*GPU_NSUBCELL + s;
1067 for (d = 0; d < DIM; d++)
1069 ba[d] += grid->bb[cs].upper[d] - grid->bb[cs].lower[d];
1073 ns += grid->nsubc[c];
1075 dsvmul(1.0/ns, ba, ba);
1077 fprintf(fp, "ns bb: %4.2f %4.2f %4.2f %4.2f %4.2f %4.2f rel %4.2f %4.2f %4.2f\n",
1078 nbs->box[XX][XX]/(grid->ncx*GPU_NSUBCELL_X),
1079 nbs->box[YY][YY]/(grid->ncy*GPU_NSUBCELL_Y),
1080 nbs->box[ZZ][ZZ]*grid->ncx*grid->ncy/(grid->nc*GPU_NSUBCELL_Z),
1081 ba[XX], ba[YY], ba[ZZ],
1082 ba[XX]*grid->ncx*GPU_NSUBCELL_X/nbs->box[XX][XX],
1083 ba[YY]*grid->ncy*GPU_NSUBCELL_Y/nbs->box[YY][YY],
1084 ba[ZZ]*grid->nc*GPU_NSUBCELL_Z/(grid->ncx*grid->ncy*nbs->box[ZZ][ZZ]));
1087 /* Potentially sorts atoms on LJ coefficients !=0 and ==0.
1088 * Also sets interaction flags.
1090 void sort_on_lj(int na_c,
1091 int a0, int a1, const int *atinfo,
1095 int subc, s, a, n1, n2, a_lj_max, i, j;
1096 int sort1[NBNXN_NA_SC_MAX/GPU_NSUBCELL];
1097 int sort2[NBNXN_NA_SC_MAX/GPU_NSUBCELL];
1098 gmx_bool haveQ, bFEP;
1103 for (s = a0; s < a1; s += na_c)
1105 /* Make lists for this (sub-)cell on atoms with and without LJ */
1110 for (a = s; a < min(s+na_c, a1); a++)
1112 haveQ = haveQ || GET_CGINFO_HAS_Q(atinfo[order[a]]);
1114 if (GET_CGINFO_HAS_VDW(atinfo[order[a]]))
1116 sort1[n1++] = order[a];
1121 sort2[n2++] = order[a];
1125 /* If we don't have atoms with LJ, there's nothing to sort */
1128 *flags |= NBNXN_CI_DO_LJ(subc);
1132 /* Only sort when strictly necessary. Ordering particles
1133 * Ordering particles can lead to less accurate summation
1134 * due to rounding, both for LJ and Coulomb interactions.
1136 if (2*(a_lj_max - s) >= na_c)
1138 for (i = 0; i < n1; i++)
1140 order[a0+i] = sort1[i];
1142 for (j = 0; j < n2; j++)
1144 order[a0+n1+j] = sort2[j];
1148 *flags |= NBNXN_CI_HALF_LJ(subc);
1153 *flags |= NBNXN_CI_DO_COUL(subc);
1159 /* Fill a pair search cell with atoms.
1160 * Potentially sorts atoms and sets the interaction flags.
1162 void fill_cell(const nbnxn_search_t nbs,
1164 nbnxn_atomdata_t *nbat,
1168 int sx, int sy, int sz,
1169 nbnxn_bb_t gmx_unused *bb_work_aligned)
1182 sort_on_lj(grid->na_c, a0, a1, atinfo, nbs->a,
1183 grid->flags+(a0>>grid->na_c_2log)-grid->cell0);
1188 /* Set the fep flag for perturbed atoms in this (sub-)cell */
1191 /* The grid-local cluster/(sub-)cell index */
1192 c = (a0 >> grid->na_c_2log) - grid->cell0*(grid->bSimple ? 1 : GPU_NSUBCELL);
1194 for (at = a0; at < a1; at++)
1196 if (nbs->a[at] >= 0 && GET_CGINFO_FEP(atinfo[nbs->a[at]]))
1198 grid->fep[c] |= (1 << (at - a0));
1203 /* Now we have sorted the atoms, set the cell indices */
1204 for (a = a0; a < a1; a++)
1206 nbs->cell[nbs->a[a]] = a;
1209 copy_rvec_to_nbat_real(nbs->a+a0, a1-a0, grid->na_c, x,
1210 nbat->XFormat, nbat->x, a0,
1213 if (nbat->XFormat == nbatX4)
1215 /* Store the bounding boxes as xyz.xyz. */
1216 offset = (a0 - grid->cell0*grid->na_sc) >> grid->na_c_2log;
1217 bb_ptr = grid->bb + offset;
1219 #if defined GMX_NBNXN_SIMD && GMX_SIMD_REAL_WIDTH == 2
1220 if (2*grid->na_cj == grid->na_c)
1222 calc_bounding_box_x_x4_halves(na, nbat->x+X4_IND_A(a0), bb_ptr,
1223 grid->bbj+offset*2);
1228 calc_bounding_box_x_x4(na, nbat->x+X4_IND_A(a0), bb_ptr);
1231 else if (nbat->XFormat == nbatX8)
1233 /* Store the bounding boxes as xyz.xyz. */
1234 offset = (a0 - grid->cell0*grid->na_sc) >> grid->na_c_2log;
1235 bb_ptr = grid->bb + offset;
1237 calc_bounding_box_x_x8(na, nbat->x+X8_IND_A(a0), bb_ptr);
1240 else if (!grid->bSimple)
1242 /* Store the bounding boxes in a format convenient
1243 * for SIMD4 calculations: xxxxyyyyzzzz...
1247 ((a0-grid->cell0*grid->na_sc)>>(grid->na_c_2log+STRIDE_PBB_2LOG))*NNBSBB_XXXX +
1248 (((a0-grid->cell0*grid->na_sc)>>grid->na_c_2log) & (STRIDE_PBB-1));
1250 #ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
1251 if (nbat->XFormat == nbatXYZQ)
1253 calc_bounding_box_xxxx_simd4(na, nbat->x+a0*nbat->xstride,
1254 bb_work_aligned, pbb_ptr);
1259 calc_bounding_box_xxxx(na, nbat->xstride, nbat->x+a0*nbat->xstride,
1264 fprintf(debug, "%2d %2d %2d bb %5.2f %5.2f %5.2f %5.2f %5.2f %5.2f\n",
1266 pbb_ptr[0*STRIDE_PBB], pbb_ptr[3*STRIDE_PBB],
1267 pbb_ptr[1*STRIDE_PBB], pbb_ptr[4*STRIDE_PBB],
1268 pbb_ptr[2*STRIDE_PBB], pbb_ptr[5*STRIDE_PBB]);
1274 /* Store the bounding boxes as xyz.xyz. */
1275 bb_ptr = grid->bb+((a0-grid->cell0*grid->na_sc)>>grid->na_c_2log);
1277 calc_bounding_box(na, nbat->xstride, nbat->x+a0*nbat->xstride,
1283 bbo = (a0 - grid->cell0*grid->na_sc)/grid->na_c;
1284 fprintf(debug, "%2d %2d %2d bb %5.2f %5.2f %5.2f %5.2f %5.2f %5.2f\n",
1286 grid->bb[bbo].lower[BB_X],
1287 grid->bb[bbo].lower[BB_Y],
1288 grid->bb[bbo].lower[BB_Z],
1289 grid->bb[bbo].upper[BB_X],
1290 grid->bb[bbo].upper[BB_Y],
1291 grid->bb[bbo].upper[BB_Z]);
1296 /* Spatially sort the atoms within one grid column */
1297 static void sort_columns_simple(const nbnxn_search_t nbs,
1303 nbnxn_atomdata_t *nbat,
1304 int cxy_start, int cxy_end,
1308 int cx, cy, cz, ncz, cfilled, c;
1309 int na, ash, ind, a;
1314 fprintf(debug, "cell0 %d sorting columns %d - %d, atoms %d - %d\n",
1315 grid->cell0, cxy_start, cxy_end, a0, a1);
1318 /* Sort the atoms within each x,y column in 3 dimensions */
1319 for (cxy = cxy_start; cxy < cxy_end; cxy++)
1322 cy = cxy - cx*grid->ncy;
1324 na = grid->cxy_na[cxy];
1325 ncz = grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy];
1326 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1328 /* Sort the atoms within each x,y column on z coordinate */
1329 sort_atoms(ZZ, FALSE, dd_zone,
1332 1.0/nbs->box[ZZ][ZZ], ncz*grid->na_sc,
1335 /* Fill the ncz cells in this column */
1336 cfilled = grid->cxy_ind[cxy];
1337 for (cz = 0; cz < ncz; cz++)
1339 c = grid->cxy_ind[cxy] + cz;
1341 ash_c = ash + cz*grid->na_sc;
1342 na_c = min(grid->na_sc, na-(ash_c-ash));
1344 fill_cell(nbs, grid, nbat,
1345 ash_c, ash_c+na_c, atinfo, x,
1346 grid->na_sc*cx + (dd_zone >> 2),
1347 grid->na_sc*cy + (dd_zone & 3),
1351 /* This copy to bbcz is not really necessary.
1352 * But it allows to use the same grid search code
1353 * for the simple and supersub cell setups.
1359 grid->bbcz[c*NNBSBB_D ] = grid->bb[cfilled].lower[BB_Z];
1360 grid->bbcz[c*NNBSBB_D+1] = grid->bb[cfilled].upper[BB_Z];
1363 /* Set the unused atom indices to -1 */
1364 for (ind = na; ind < ncz*grid->na_sc; ind++)
1366 nbs->a[ash+ind] = -1;
1371 /* Spatially sort the atoms within one grid column */
1372 static void sort_columns_supersub(const nbnxn_search_t nbs,
1378 nbnxn_atomdata_t *nbat,
1379 int cxy_start, int cxy_end,
1383 int cx, cy, cz = -1, c = -1, ncz;
1384 int na, ash, na_c, ind, a;
1385 int subdiv_z, sub_z, na_z, ash_z;
1386 int subdiv_y, sub_y, na_y, ash_y;
1387 int subdiv_x, sub_x, na_x, ash_x;
1389 nbnxn_bb_t bb_work_array[2], *bb_work_aligned;
1391 bb_work_aligned = (nbnxn_bb_t *)(((size_t)(bb_work_array+1)) & (~((size_t)15)));
1395 fprintf(debug, "cell0 %d sorting columns %d - %d, atoms %d - %d\n",
1396 grid->cell0, cxy_start, cxy_end, a0, a1);
1399 subdiv_x = grid->na_c;
1400 subdiv_y = GPU_NSUBCELL_X*subdiv_x;
1401 subdiv_z = GPU_NSUBCELL_Y*subdiv_y;
1403 /* Sort the atoms within each x,y column in 3 dimensions */
1404 for (cxy = cxy_start; cxy < cxy_end; cxy++)
1407 cy = cxy - cx*grid->ncy;
1409 na = grid->cxy_na[cxy];
1410 ncz = grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy];
1411 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1413 /* Sort the atoms within each x,y column on z coordinate */
1414 sort_atoms(ZZ, FALSE, dd_zone,
1417 1.0/nbs->box[ZZ][ZZ], ncz*grid->na_sc,
1420 /* This loop goes over the supercells and subcells along z at once */
1421 for (sub_z = 0; sub_z < ncz*GPU_NSUBCELL_Z; sub_z++)
1423 ash_z = ash + sub_z*subdiv_z;
1424 na_z = min(subdiv_z, na-(ash_z-ash));
1426 /* We have already sorted on z */
1428 if (sub_z % GPU_NSUBCELL_Z == 0)
1430 cz = sub_z/GPU_NSUBCELL_Z;
1431 c = grid->cxy_ind[cxy] + cz;
1433 /* The number of atoms in this supercell */
1434 na_c = min(grid->na_sc, na-(ash_z-ash));
1436 grid->nsubc[c] = min(GPU_NSUBCELL, (na_c+grid->na_c-1)/grid->na_c);
1438 /* Store the z-boundaries of the super cell */
1439 grid->bbcz[c*NNBSBB_D ] = x[nbs->a[ash_z]][ZZ];
1440 grid->bbcz[c*NNBSBB_D+1] = x[nbs->a[ash_z+na_c-1]][ZZ];
1443 #if GPU_NSUBCELL_Y > 1
1444 /* Sort the atoms along y */
1445 sort_atoms(YY, (sub_z & 1), dd_zone,
1446 nbs->a+ash_z, na_z, x,
1447 grid->c0[YY]+cy*grid->sy,
1448 grid->inv_sy, subdiv_z,
1452 for (sub_y = 0; sub_y < GPU_NSUBCELL_Y; sub_y++)
1454 ash_y = ash_z + sub_y*subdiv_y;
1455 na_y = min(subdiv_y, na-(ash_y-ash));
1457 #if GPU_NSUBCELL_X > 1
1458 /* Sort the atoms along x */
1459 sort_atoms(XX, ((cz*GPU_NSUBCELL_Y + sub_y) & 1), dd_zone,
1460 nbs->a+ash_y, na_y, x,
1461 grid->c0[XX]+cx*grid->sx,
1462 grid->inv_sx, subdiv_y,
1466 for (sub_x = 0; sub_x < GPU_NSUBCELL_X; sub_x++)
1468 ash_x = ash_y + sub_x*subdiv_x;
1469 na_x = min(subdiv_x, na-(ash_x-ash));
1471 fill_cell(nbs, grid, nbat,
1472 ash_x, ash_x+na_x, atinfo, x,
1473 grid->na_c*(cx*GPU_NSUBCELL_X+sub_x) + (dd_zone >> 2),
1474 grid->na_c*(cy*GPU_NSUBCELL_Y+sub_y) + (dd_zone & 3),
1481 /* Set the unused atom indices to -1 */
1482 for (ind = na; ind < ncz*grid->na_sc; ind++)
1484 nbs->a[ash+ind] = -1;
1489 /* Determine in which grid column atoms should go */
1490 static void calc_column_indices(nbnxn_grid_t *grid,
1493 int dd_zone, const int *move,
1494 int thread, int nthread,
1501 /* We add one extra cell for particles which moved during DD */
1502 for (i = 0; i < grid->ncx*grid->ncy+1; i++)
1507 n0 = a0 + (int)((thread+0)*(a1 - a0))/nthread;
1508 n1 = a0 + (int)((thread+1)*(a1 - a0))/nthread;
1512 for (i = n0; i < n1; i++)
1514 if (move == NULL || move[i] >= 0)
1516 /* We need to be careful with rounding,
1517 * particles might be a few bits outside the local zone.
1518 * The int cast takes care of the lower bound,
1519 * we will explicitly take care of the upper bound.
1521 cx = (int)((x[i][XX] - grid->c0[XX])*grid->inv_sx);
1522 cy = (int)((x[i][YY] - grid->c0[YY])*grid->inv_sy);
1525 if (cx < 0 || cx > grid->ncx ||
1526 cy < 0 || cy > grid->ncy)
1529 "grid cell cx %d cy %d out of range (max %d %d)\n"
1530 "atom %f %f %f, grid->c0 %f %f",
1531 cx, cy, grid->ncx, grid->ncy,
1532 x[i][XX], x[i][YY], x[i][ZZ], grid->c0[XX], grid->c0[YY]);
1535 /* Take care of potential rouding issues */
1536 cx = min(cx, grid->ncx - 1);
1537 cy = min(cy, grid->ncy - 1);
1539 /* For the moment cell will contain only the, grid local,
1540 * x and y indices, not z.
1542 cell[i] = cx*grid->ncy + cy;
1546 /* Put this moved particle after the end of the grid,
1547 * so we can process it later without using conditionals.
1549 cell[i] = grid->ncx*grid->ncy;
1558 for (i = n0; i < n1; i++)
1560 cx = (int)((x[i][XX] - grid->c0[XX])*grid->inv_sx);
1561 cy = (int)((x[i][YY] - grid->c0[YY])*grid->inv_sy);
1563 /* For non-home zones there could be particles outside
1564 * the non-bonded cut-off range, which have been communicated
1565 * for bonded interactions only. For the result it doesn't
1566 * matter where these end up on the grid. For performance
1567 * we put them in an extra row at the border.
1570 cx = min(cx, grid->ncx - 1);
1572 cy = min(cy, grid->ncy - 1);
1574 /* For the moment cell will contain only the, grid local,
1575 * x and y indices, not z.
1577 cell[i] = cx*grid->ncy + cy;
1584 /* Determine in which grid cells the atoms should go */
1585 static void calc_cell_indices(const nbnxn_search_t nbs,
1592 nbnxn_atomdata_t *nbat)
1595 int cx, cy, cxy, ncz_max, ncz;
1596 int nthread, thread;
1597 int *cxy_na, cxy_na_i;
1599 nthread = gmx_omp_nthreads_get(emntPairsearch);
1601 #pragma omp parallel for num_threads(nthread) schedule(static)
1602 for (thread = 0; thread < nthread; thread++)
1604 calc_column_indices(grid, a0, a1, x, dd_zone, move, thread, nthread,
1605 nbs->cell, nbs->work[thread].cxy_na);
1608 /* Make the cell index as a function of x and y */
1611 grid->cxy_ind[0] = 0;
1612 for (i = 0; i < grid->ncx*grid->ncy+1; i++)
1614 /* We set ncz_max at the beginning of the loop iso at the end
1615 * to skip i=grid->ncx*grid->ncy which are moved particles
1616 * that do not need to be ordered on the grid.
1622 cxy_na_i = nbs->work[0].cxy_na[i];
1623 for (thread = 1; thread < nthread; thread++)
1625 cxy_na_i += nbs->work[thread].cxy_na[i];
1627 ncz = (cxy_na_i + grid->na_sc - 1)/grid->na_sc;
1628 if (nbat->XFormat == nbatX8)
1630 /* Make the number of cell a multiple of 2 */
1631 ncz = (ncz + 1) & ~1;
1633 grid->cxy_ind[i+1] = grid->cxy_ind[i] + ncz;
1634 /* Clear cxy_na, so we can reuse the array below */
1635 grid->cxy_na[i] = 0;
1637 grid->nc = grid->cxy_ind[grid->ncx*grid->ncy] - grid->cxy_ind[0];
1639 nbat->natoms = (grid->cell0 + grid->nc)*grid->na_sc;
1643 fprintf(debug, "ns na_sc %d na_c %d super-cells: %d x %d y %d z %.1f maxz %d\n",
1644 grid->na_sc, grid->na_c, grid->nc,
1645 grid->ncx, grid->ncy, grid->nc/((double)(grid->ncx*grid->ncy)),
1650 for (cy = 0; cy < grid->ncy; cy++)
1652 for (cx = 0; cx < grid->ncx; cx++)
1654 fprintf(debug, " %2d", grid->cxy_ind[i+1]-grid->cxy_ind[i]);
1657 fprintf(debug, "\n");
1662 /* Make sure the work array for sorting is large enough */
1663 if (ncz_max*grid->na_sc*SGSF > nbs->work[0].sort_work_nalloc)
1665 for (thread = 0; thread < nbs->nthread_max; thread++)
1667 nbs->work[thread].sort_work_nalloc =
1668 over_alloc_large(ncz_max*grid->na_sc*SGSF);
1669 srenew(nbs->work[thread].sort_work,
1670 nbs->work[thread].sort_work_nalloc);
1671 /* When not in use, all elements should be -1 */
1672 for (i = 0; i < nbs->work[thread].sort_work_nalloc; i++)
1674 nbs->work[thread].sort_work[i] = -1;
1679 /* Now we know the dimensions we can fill the grid.
1680 * This is the first, unsorted fill. We sort the columns after this.
1682 for (i = a0; i < a1; i++)
1684 /* At this point nbs->cell contains the local grid x,y indices */
1686 nbs->a[(grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc + grid->cxy_na[cxy]++] = i;
1691 /* Set the cell indices for the moved particles */
1692 n0 = grid->nc*grid->na_sc;
1693 n1 = grid->nc*grid->na_sc+grid->cxy_na[grid->ncx*grid->ncy];
1696 for (i = n0; i < n1; i++)
1698 nbs->cell[nbs->a[i]] = i;
1703 /* Sort the super-cell columns along z into the sub-cells. */
1704 #pragma omp parallel for num_threads(nthread) schedule(static)
1705 for (thread = 0; thread < nthread; thread++)
1709 sort_columns_simple(nbs, dd_zone, grid, a0, a1, atinfo, x, nbat,
1710 ((thread+0)*grid->ncx*grid->ncy)/nthread,
1711 ((thread+1)*grid->ncx*grid->ncy)/nthread,
1712 nbs->work[thread].sort_work);
1716 sort_columns_supersub(nbs, dd_zone, grid, a0, a1, atinfo, x, nbat,
1717 ((thread+0)*grid->ncx*grid->ncy)/nthread,
1718 ((thread+1)*grid->ncx*grid->ncy)/nthread,
1719 nbs->work[thread].sort_work);
1723 if (grid->bSimple && nbat->XFormat == nbatX8)
1725 combine_bounding_box_pairs(grid, grid->bb);
1730 grid->nsubc_tot = 0;
1731 for (i = 0; i < grid->nc; i++)
1733 grid->nsubc_tot += grid->nsubc[i];
1741 print_bbsizes_simple(debug, nbs, grid);
1745 fprintf(debug, "ns non-zero sub-cells: %d average atoms %.2f\n",
1746 grid->nsubc_tot, (a1-a0)/(double)grid->nsubc_tot);
1748 print_bbsizes_supersub(debug, nbs, grid);
1753 static void init_buffer_flags(nbnxn_buffer_flags_t *flags,
1758 flags->nflag = (natoms + NBNXN_BUFFERFLAG_SIZE - 1)/NBNXN_BUFFERFLAG_SIZE;
1759 if (flags->nflag > flags->flag_nalloc)
1761 flags->flag_nalloc = over_alloc_large(flags->nflag);
1762 srenew(flags->flag, flags->flag_nalloc);
1764 for (b = 0; b < flags->nflag; b++)
1770 /* Sets up a grid and puts the atoms on the grid.
1771 * This function only operates on one domain of the domain decompostion.
1772 * Note that without domain decomposition there is only one domain.
1774 void nbnxn_put_on_grid(nbnxn_search_t nbs,
1775 int ePBC, matrix box,
1777 rvec corner0, rvec corner1,
1782 int nmoved, int *move,
1784 nbnxn_atomdata_t *nbat)
1788 int nc_max_grid, nc_max;
1790 grid = &nbs->grid[dd_zone];
1792 nbs_cycle_start(&nbs->cc[enbsCCgrid]);
1794 grid->bSimple = nbnxn_kernel_pairlist_simple(nb_kernel_type);
1796 grid->na_c = nbnxn_kernel_to_ci_size(nb_kernel_type);
1797 grid->na_cj = nbnxn_kernel_to_cj_size(nb_kernel_type);
1798 grid->na_sc = (grid->bSimple ? 1 : GPU_NSUBCELL)*grid->na_c;
1799 grid->na_c_2log = get_2log(grid->na_c);
1801 nbat->na_c = grid->na_c;
1810 (nbs->grid[dd_zone-1].cell0 + nbs->grid[dd_zone-1].nc)*
1811 nbs->grid[dd_zone-1].na_sc/grid->na_sc;
1819 copy_mat(box, nbs->box);
1821 /* Avoid zero density */
1822 if (atom_density > 0)
1824 grid->atom_density = atom_density;
1828 grid->atom_density = grid_atom_density(n-nmoved, corner0, corner1);
1833 nbs->natoms_local = a1 - nmoved;
1834 /* We assume that nbnxn_put_on_grid is called first
1835 * for the local atoms (dd_zone=0).
1837 nbs->natoms_nonlocal = a1 - nmoved;
1841 fprintf(debug, "natoms_local = %5d atom_density = %5.1f\n",
1842 nbs->natoms_local, grid->atom_density);
1847 nbs->natoms_nonlocal = max(nbs->natoms_nonlocal, a1);
1850 /* We always use the home zone (grid[0]) for setting the cell size,
1851 * since determining densities for non-local zones is difficult.
1853 nc_max_grid = set_grid_size_xy(nbs, grid,
1854 dd_zone, n-nmoved, corner0, corner1,
1855 nbs->grid[0].atom_density);
1857 nc_max = grid->cell0 + nc_max_grid;
1859 if (a1 > nbs->cell_nalloc)
1861 nbs->cell_nalloc = over_alloc_large(a1);
1862 srenew(nbs->cell, nbs->cell_nalloc);
1865 /* To avoid conditionals we store the moved particles at the end of a,
1866 * make sure we have enough space.
1868 if (nc_max*grid->na_sc + nmoved > nbs->a_nalloc)
1870 nbs->a_nalloc = over_alloc_large(nc_max*grid->na_sc + nmoved);
1871 srenew(nbs->a, nbs->a_nalloc);
1874 /* We need padding up to a multiple of the buffer flag size: simply add */
1875 if (nc_max*grid->na_sc + NBNXN_BUFFERFLAG_SIZE > nbat->nalloc)
1877 nbnxn_atomdata_realloc(nbat, nc_max*grid->na_sc+NBNXN_BUFFERFLAG_SIZE);
1880 calc_cell_indices(nbs, dd_zone, grid, a0, a1, atinfo, x, move, nbat);
1884 nbat->natoms_local = nbat->natoms;
1887 nbs_cycle_stop(&nbs->cc[enbsCCgrid]);
1890 /* Calls nbnxn_put_on_grid for all non-local domains */
1891 void nbnxn_put_on_grid_nonlocal(nbnxn_search_t nbs,
1892 const gmx_domdec_zones_t *zones,
1896 nbnxn_atomdata_t *nbat)
1901 for (zone = 1; zone < zones->n; zone++)
1903 for (d = 0; d < DIM; d++)
1905 c0[d] = zones->size[zone].bb_x0[d];
1906 c1[d] = zones->size[zone].bb_x1[d];
1909 nbnxn_put_on_grid(nbs, nbs->ePBC, NULL,
1911 zones->cg_range[zone],
1912 zones->cg_range[zone+1],
1922 /* Add simple grid type information to the local super/sub grid */
1923 void nbnxn_grid_add_simple(nbnxn_search_t nbs,
1924 nbnxn_atomdata_t *nbat)
1931 grid = &nbs->grid[0];
1935 gmx_incons("nbnxn_grid_simple called with a simple grid");
1938 ncd = grid->na_sc/NBNXN_CPU_CLUSTER_I_SIZE;
1940 if (grid->nc*ncd > grid->nc_nalloc_simple)
1942 grid->nc_nalloc_simple = over_alloc_large(grid->nc*ncd);
1943 srenew(grid->bbcz_simple, grid->nc_nalloc_simple*NNBSBB_D);
1944 srenew(grid->bb_simple, grid->nc_nalloc_simple);
1945 srenew(grid->flags_simple, grid->nc_nalloc_simple);
1948 sfree_aligned(grid->bbj);
1949 snew_aligned(grid->bbj, grid->nc_nalloc_simple/2, 16);
1953 bbcz = grid->bbcz_simple;
1954 bb = grid->bb_simple;
1956 #pragma omp parallel for num_threads(gmx_omp_nthreads_get(emntPairsearch)) schedule(static)
1957 for (sc = 0; sc < grid->nc; sc++)
1961 for (c = 0; c < ncd; c++)
1965 na = NBNXN_CPU_CLUSTER_I_SIZE;
1967 nbat->type[tx*NBNXN_CPU_CLUSTER_I_SIZE+na-1] == nbat->ntype-1)
1974 switch (nbat->XFormat)
1977 /* PACK_X4==NBNXN_CPU_CLUSTER_I_SIZE, so this is simple */
1978 calc_bounding_box_x_x4(na, nbat->x+tx*STRIDE_P4,
1982 /* PACK_X8>NBNXN_CPU_CLUSTER_I_SIZE, more complicated */
1983 calc_bounding_box_x_x8(na, nbat->x+X8_IND_A(tx*NBNXN_CPU_CLUSTER_I_SIZE),
1987 calc_bounding_box(na, nbat->xstride,
1988 nbat->x+tx*NBNXN_CPU_CLUSTER_I_SIZE*nbat->xstride,
1992 bbcz[tx*NNBSBB_D+0] = bb[tx].lower[BB_Z];
1993 bbcz[tx*NNBSBB_D+1] = bb[tx].upper[BB_Z];
1995 /* No interaction optimization yet here */
1996 grid->flags_simple[tx] = NBNXN_CI_DO_LJ(0) | NBNXN_CI_DO_COUL(0);
2000 grid->flags_simple[tx] = 0;
2005 if (grid->bSimple && nbat->XFormat == nbatX8)
2007 combine_bounding_box_pairs(grid, grid->bb_simple);
2011 void nbnxn_get_ncells(nbnxn_search_t nbs, int *ncx, int *ncy)
2013 *ncx = nbs->grid[0].ncx;
2014 *ncy = nbs->grid[0].ncy;
2017 void nbnxn_get_atomorder(nbnxn_search_t nbs, int **a, int *n)
2019 const nbnxn_grid_t *grid;
2021 grid = &nbs->grid[0];
2023 /* Return the atom order for the home cell (index 0) */
2026 *n = grid->cxy_ind[grid->ncx*grid->ncy]*grid->na_sc;
2029 void nbnxn_set_atomorder(nbnxn_search_t nbs)
2032 int ao, cx, cy, cxy, cz, j;
2034 /* Set the atom order for the home cell (index 0) */
2035 grid = &nbs->grid[0];
2038 for (cx = 0; cx < grid->ncx; cx++)
2040 for (cy = 0; cy < grid->ncy; cy++)
2042 cxy = cx*grid->ncy + cy;
2043 j = grid->cxy_ind[cxy]*grid->na_sc;
2044 for (cz = 0; cz < grid->cxy_na[cxy]; cz++)
2055 /* Determines the cell range along one dimension that
2056 * the bounding box b0 - b1 sees.
2058 static void get_cell_range(real b0, real b1,
2059 int nc, real c0, real s, real invs,
2060 real d2, real r2, int *cf, int *cl)
2062 *cf = max((int)((b0 - c0)*invs), 0);
2064 while (*cf > 0 && d2 + sqr((b0 - c0) - (*cf-1+1)*s) < r2)
2069 *cl = min((int)((b1 - c0)*invs), nc-1);
2070 while (*cl < nc-1 && d2 + sqr((*cl+1)*s - (b1 - c0)) < r2)
2076 /* Reference code calculating the distance^2 between two bounding boxes */
2077 static float box_dist2(float bx0, float bx1, float by0,
2078 float by1, float bz0, float bz1,
2079 const nbnxn_bb_t *bb)
2082 float dl, dh, dm, dm0;
2086 dl = bx0 - bb->upper[BB_X];
2087 dh = bb->lower[BB_X] - bx1;
2092 dl = by0 - bb->upper[BB_Y];
2093 dh = bb->lower[BB_Y] - by1;
2098 dl = bz0 - bb->upper[BB_Z];
2099 dh = bb->lower[BB_Z] - bz1;
2107 /* Plain C code calculating the distance^2 between two bounding boxes */
2108 static float subc_bb_dist2(int si, const nbnxn_bb_t *bb_i_ci,
2109 int csj, const nbnxn_bb_t *bb_j_all)
2111 const nbnxn_bb_t *bb_i, *bb_j;
2113 float dl, dh, dm, dm0;
2115 bb_i = bb_i_ci + si;
2116 bb_j = bb_j_all + csj;
2120 dl = bb_i->lower[BB_X] - bb_j->upper[BB_X];
2121 dh = bb_j->lower[BB_X] - bb_i->upper[BB_X];
2126 dl = bb_i->lower[BB_Y] - bb_j->upper[BB_Y];
2127 dh = bb_j->lower[BB_Y] - bb_i->upper[BB_Y];
2132 dl = bb_i->lower[BB_Z] - bb_j->upper[BB_Z];
2133 dh = bb_j->lower[BB_Z] - bb_i->upper[BB_Z];
2141 #ifdef NBNXN_SEARCH_BB_SIMD4
2143 /* 4-wide SIMD code for bb distance for bb format xyz0 */
2144 static float subc_bb_dist2_simd4(int si, const nbnxn_bb_t *bb_i_ci,
2145 int csj, const nbnxn_bb_t *bb_j_all)
2147 gmx_simd4_float_t bb_i_S0, bb_i_S1;
2148 gmx_simd4_float_t bb_j_S0, bb_j_S1;
2149 gmx_simd4_float_t dl_S;
2150 gmx_simd4_float_t dh_S;
2151 gmx_simd4_float_t dm_S;
2152 gmx_simd4_float_t dm0_S;
2154 bb_i_S0 = gmx_simd4_load_f(&bb_i_ci[si].lower[0]);
2155 bb_i_S1 = gmx_simd4_load_f(&bb_i_ci[si].upper[0]);
2156 bb_j_S0 = gmx_simd4_load_f(&bb_j_all[csj].lower[0]);
2157 bb_j_S1 = gmx_simd4_load_f(&bb_j_all[csj].upper[0]);
2159 dl_S = gmx_simd4_sub_f(bb_i_S0, bb_j_S1);
2160 dh_S = gmx_simd4_sub_f(bb_j_S0, bb_i_S1);
2162 dm_S = gmx_simd4_max_f(dl_S, dh_S);
2163 dm0_S = gmx_simd4_max_f(dm_S, gmx_simd4_setzero_f());
2165 return gmx_simd4_dotproduct3_f(dm0_S, dm0_S);
2168 /* Calculate bb bounding distances of bb_i[si,...,si+3] and store them in d2 */
2169 #define SUBC_BB_DIST2_SIMD4_XXXX_INNER(si, bb_i, d2) \
2173 gmx_simd4_float_t dx_0, dy_0, dz_0; \
2174 gmx_simd4_float_t dx_1, dy_1, dz_1; \
2176 gmx_simd4_float_t mx, my, mz; \
2177 gmx_simd4_float_t m0x, m0y, m0z; \
2179 gmx_simd4_float_t d2x, d2y, d2z; \
2180 gmx_simd4_float_t d2s, d2t; \
2182 shi = si*NNBSBB_D*DIM; \
2184 xi_l = gmx_simd4_load_f(bb_i+shi+0*STRIDE_PBB); \
2185 yi_l = gmx_simd4_load_f(bb_i+shi+1*STRIDE_PBB); \
2186 zi_l = gmx_simd4_load_f(bb_i+shi+2*STRIDE_PBB); \
2187 xi_h = gmx_simd4_load_f(bb_i+shi+3*STRIDE_PBB); \
2188 yi_h = gmx_simd4_load_f(bb_i+shi+4*STRIDE_PBB); \
2189 zi_h = gmx_simd4_load_f(bb_i+shi+5*STRIDE_PBB); \
2191 dx_0 = gmx_simd4_sub_f(xi_l, xj_h); \
2192 dy_0 = gmx_simd4_sub_f(yi_l, yj_h); \
2193 dz_0 = gmx_simd4_sub_f(zi_l, zj_h); \
2195 dx_1 = gmx_simd4_sub_f(xj_l, xi_h); \
2196 dy_1 = gmx_simd4_sub_f(yj_l, yi_h); \
2197 dz_1 = gmx_simd4_sub_f(zj_l, zi_h); \
2199 mx = gmx_simd4_max_f(dx_0, dx_1); \
2200 my = gmx_simd4_max_f(dy_0, dy_1); \
2201 mz = gmx_simd4_max_f(dz_0, dz_1); \
2203 m0x = gmx_simd4_max_f(mx, zero); \
2204 m0y = gmx_simd4_max_f(my, zero); \
2205 m0z = gmx_simd4_max_f(mz, zero); \
2207 d2x = gmx_simd4_mul_f(m0x, m0x); \
2208 d2y = gmx_simd4_mul_f(m0y, m0y); \
2209 d2z = gmx_simd4_mul_f(m0z, m0z); \
2211 d2s = gmx_simd4_add_f(d2x, d2y); \
2212 d2t = gmx_simd4_add_f(d2s, d2z); \
2214 gmx_simd4_store_f(d2+si, d2t); \
2217 /* 4-wide SIMD code for nsi bb distances for bb format xxxxyyyyzzzz */
2218 static void subc_bb_dist2_simd4_xxxx(const float *bb_j,
2219 int nsi, const float *bb_i,
2222 gmx_simd4_float_t xj_l, yj_l, zj_l;
2223 gmx_simd4_float_t xj_h, yj_h, zj_h;
2224 gmx_simd4_float_t xi_l, yi_l, zi_l;
2225 gmx_simd4_float_t xi_h, yi_h, zi_h;
2227 gmx_simd4_float_t zero;
2229 zero = gmx_simd4_setzero_f();
2231 xj_l = gmx_simd4_set1_f(bb_j[0*STRIDE_PBB]);
2232 yj_l = gmx_simd4_set1_f(bb_j[1*STRIDE_PBB]);
2233 zj_l = gmx_simd4_set1_f(bb_j[2*STRIDE_PBB]);
2234 xj_h = gmx_simd4_set1_f(bb_j[3*STRIDE_PBB]);
2235 yj_h = gmx_simd4_set1_f(bb_j[4*STRIDE_PBB]);
2236 zj_h = gmx_simd4_set1_f(bb_j[5*STRIDE_PBB]);
2238 /* Here we "loop" over si (0,STRIDE_PBB) from 0 to nsi with step STRIDE_PBB.
2239 * But as we know the number of iterations is 1 or 2, we unroll manually.
2241 SUBC_BB_DIST2_SIMD4_XXXX_INNER(0, bb_i, d2);
2242 if (STRIDE_PBB < nsi)
2244 SUBC_BB_DIST2_SIMD4_XXXX_INNER(STRIDE_PBB, bb_i, d2);
2248 #endif /* NBNXN_SEARCH_BB_SIMD4 */
2250 /* Plain C function which determines if any atom pair between two cells
2251 * is within distance sqrt(rl2).
2253 static gmx_bool subc_in_range_x(int na_c,
2254 int si, const real *x_i,
2255 int csj, int stride, const real *x_j,
2261 for (i = 0; i < na_c; i++)
2263 i0 = (si*na_c + i)*DIM;
2264 for (j = 0; j < na_c; j++)
2266 j0 = (csj*na_c + j)*stride;
2268 d2 = sqr(x_i[i0 ] - x_j[j0 ]) +
2269 sqr(x_i[i0+1] - x_j[j0+1]) +
2270 sqr(x_i[i0+2] - x_j[j0+2]);
2282 #ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
2284 /* 4-wide SIMD function which determines if any atom pair between two cells,
2285 * both with 8 atoms, is within distance sqrt(rl2).
2286 * Using 8-wide AVX is not faster on Intel Sandy Bridge.
2288 static gmx_bool subc_in_range_simd4(int na_c,
2289 int si, const real *x_i,
2290 int csj, int stride, const real *x_j,
2293 gmx_simd4_real_t ix_S0, iy_S0, iz_S0;
2294 gmx_simd4_real_t ix_S1, iy_S1, iz_S1;
2296 gmx_simd4_real_t rc2_S;
2301 rc2_S = gmx_simd4_set1_r(rl2);
2303 dim_stride = NBNXN_GPU_CLUSTER_SIZE/STRIDE_PBB*DIM;
2304 ix_S0 = gmx_simd4_load_r(x_i+(si*dim_stride+0)*STRIDE_PBB);
2305 iy_S0 = gmx_simd4_load_r(x_i+(si*dim_stride+1)*STRIDE_PBB);
2306 iz_S0 = gmx_simd4_load_r(x_i+(si*dim_stride+2)*STRIDE_PBB);
2307 ix_S1 = gmx_simd4_load_r(x_i+(si*dim_stride+3)*STRIDE_PBB);
2308 iy_S1 = gmx_simd4_load_r(x_i+(si*dim_stride+4)*STRIDE_PBB);
2309 iz_S1 = gmx_simd4_load_r(x_i+(si*dim_stride+5)*STRIDE_PBB);
2311 /* We loop from the outer to the inner particles to maximize
2312 * the chance that we find a pair in range quickly and return.
2318 gmx_simd4_real_t jx0_S, jy0_S, jz0_S;
2319 gmx_simd4_real_t jx1_S, jy1_S, jz1_S;
2321 gmx_simd4_real_t dx_S0, dy_S0, dz_S0;
2322 gmx_simd4_real_t dx_S1, dy_S1, dz_S1;
2323 gmx_simd4_real_t dx_S2, dy_S2, dz_S2;
2324 gmx_simd4_real_t dx_S3, dy_S3, dz_S3;
2326 gmx_simd4_real_t rsq_S0;
2327 gmx_simd4_real_t rsq_S1;
2328 gmx_simd4_real_t rsq_S2;
2329 gmx_simd4_real_t rsq_S3;
2331 gmx_simd4_bool_t wco_S0;
2332 gmx_simd4_bool_t wco_S1;
2333 gmx_simd4_bool_t wco_S2;
2334 gmx_simd4_bool_t wco_S3;
2335 gmx_simd4_bool_t wco_any_S01, wco_any_S23, wco_any_S;
2337 jx0_S = gmx_simd4_set1_r(x_j[j0*stride+0]);
2338 jy0_S = gmx_simd4_set1_r(x_j[j0*stride+1]);
2339 jz0_S = gmx_simd4_set1_r(x_j[j0*stride+2]);
2341 jx1_S = gmx_simd4_set1_r(x_j[j1*stride+0]);
2342 jy1_S = gmx_simd4_set1_r(x_j[j1*stride+1]);
2343 jz1_S = gmx_simd4_set1_r(x_j[j1*stride+2]);
2345 /* Calculate distance */
2346 dx_S0 = gmx_simd4_sub_r(ix_S0, jx0_S);
2347 dy_S0 = gmx_simd4_sub_r(iy_S0, jy0_S);
2348 dz_S0 = gmx_simd4_sub_r(iz_S0, jz0_S);
2349 dx_S1 = gmx_simd4_sub_r(ix_S1, jx0_S);
2350 dy_S1 = gmx_simd4_sub_r(iy_S1, jy0_S);
2351 dz_S1 = gmx_simd4_sub_r(iz_S1, jz0_S);
2352 dx_S2 = gmx_simd4_sub_r(ix_S0, jx1_S);
2353 dy_S2 = gmx_simd4_sub_r(iy_S0, jy1_S);
2354 dz_S2 = gmx_simd4_sub_r(iz_S0, jz1_S);
2355 dx_S3 = gmx_simd4_sub_r(ix_S1, jx1_S);
2356 dy_S3 = gmx_simd4_sub_r(iy_S1, jy1_S);
2357 dz_S3 = gmx_simd4_sub_r(iz_S1, jz1_S);
2359 /* rsq = dx*dx+dy*dy+dz*dz */
2360 rsq_S0 = gmx_simd4_calc_rsq_r(dx_S0, dy_S0, dz_S0);
2361 rsq_S1 = gmx_simd4_calc_rsq_r(dx_S1, dy_S1, dz_S1);
2362 rsq_S2 = gmx_simd4_calc_rsq_r(dx_S2, dy_S2, dz_S2);
2363 rsq_S3 = gmx_simd4_calc_rsq_r(dx_S3, dy_S3, dz_S3);
2365 wco_S0 = gmx_simd4_cmplt_r(rsq_S0, rc2_S);
2366 wco_S1 = gmx_simd4_cmplt_r(rsq_S1, rc2_S);
2367 wco_S2 = gmx_simd4_cmplt_r(rsq_S2, rc2_S);
2368 wco_S3 = gmx_simd4_cmplt_r(rsq_S3, rc2_S);
2370 wco_any_S01 = gmx_simd4_or_b(wco_S0, wco_S1);
2371 wco_any_S23 = gmx_simd4_or_b(wco_S2, wco_S3);
2372 wco_any_S = gmx_simd4_or_b(wco_any_S01, wco_any_S23);
2374 if (gmx_simd4_anytrue_b(wco_any_S))
2388 /* Returns the j sub-cell for index cj_ind */
2389 static int nbl_cj(const nbnxn_pairlist_t *nbl, int cj_ind)
2391 return nbl->cj4[cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG].cj[cj_ind & (NBNXN_GPU_JGROUP_SIZE - 1)];
2394 /* Returns the i-interaction mask of the j sub-cell for index cj_ind */
2395 static unsigned int nbl_imask0(const nbnxn_pairlist_t *nbl, int cj_ind)
2397 return nbl->cj4[cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG].imei[0].imask;
2400 /* Ensures there is enough space for extra extra exclusion masks */
2401 static void check_excl_space(nbnxn_pairlist_t *nbl, int extra)
2403 if (nbl->nexcl+extra > nbl->excl_nalloc)
2405 nbl->excl_nalloc = over_alloc_small(nbl->nexcl+extra);
2406 nbnxn_realloc_void((void **)&nbl->excl,
2407 nbl->nexcl*sizeof(*nbl->excl),
2408 nbl->excl_nalloc*sizeof(*nbl->excl),
2409 nbl->alloc, nbl->free);
2413 /* Ensures there is enough space for ncell extra j-cells in the list */
2414 static void check_subcell_list_space_simple(nbnxn_pairlist_t *nbl,
2419 cj_max = nbl->ncj + ncell;
2421 if (cj_max > nbl->cj_nalloc)
2423 nbl->cj_nalloc = over_alloc_small(cj_max);
2424 nbnxn_realloc_void((void **)&nbl->cj,
2425 nbl->ncj*sizeof(*nbl->cj),
2426 nbl->cj_nalloc*sizeof(*nbl->cj),
2427 nbl->alloc, nbl->free);
2431 /* Ensures there is enough space for ncell extra j-subcells in the list */
2432 static void check_subcell_list_space_supersub(nbnxn_pairlist_t *nbl,
2435 int ncj4_max, j4, j, w, t;
2438 #define WARP_SIZE 32
2440 /* We can have maximally nsupercell*GPU_NSUBCELL sj lists */
2441 /* We can store 4 j-subcell - i-supercell pairs in one struct.
2442 * since we round down, we need one extra entry.
2444 ncj4_max = ((nbl->work->cj_ind + nsupercell*GPU_NSUBCELL + NBNXN_GPU_JGROUP_SIZE - 1) >> NBNXN_GPU_JGROUP_SIZE_2LOG);
2446 if (ncj4_max > nbl->cj4_nalloc)
2448 nbl->cj4_nalloc = over_alloc_small(ncj4_max);
2449 nbnxn_realloc_void((void **)&nbl->cj4,
2450 nbl->work->cj4_init*sizeof(*nbl->cj4),
2451 nbl->cj4_nalloc*sizeof(*nbl->cj4),
2452 nbl->alloc, nbl->free);
2455 if (ncj4_max > nbl->work->cj4_init)
2457 for (j4 = nbl->work->cj4_init; j4 < ncj4_max; j4++)
2459 /* No i-subcells and no excl's in the list initially */
2460 for (w = 0; w < NWARP; w++)
2462 nbl->cj4[j4].imei[w].imask = 0U;
2463 nbl->cj4[j4].imei[w].excl_ind = 0;
2467 nbl->work->cj4_init = ncj4_max;
2471 /* Set all excl masks for one GPU warp no exclusions */
2472 static void set_no_excls(nbnxn_excl_t *excl)
2476 for (t = 0; t < WARP_SIZE; t++)
2478 /* Turn all interaction bits on */
2479 excl->pair[t] = NBNXN_INTERACTION_MASK_ALL;
2483 /* Initializes a single nbnxn_pairlist_t data structure */
2484 static void nbnxn_init_pairlist(nbnxn_pairlist_t *nbl,
2486 nbnxn_alloc_t *alloc,
2491 nbl->alloc = nbnxn_alloc_aligned;
2499 nbl->free = nbnxn_free_aligned;
2506 nbl->bSimple = bSimple;
2517 /* We need one element extra in sj, so alloc initially with 1 */
2518 nbl->cj4_nalloc = 0;
2525 nbl->excl_nalloc = 0;
2527 check_excl_space(nbl, 1);
2529 set_no_excls(&nbl->excl[0]);
2535 snew_aligned(nbl->work->bb_ci, 1, NBNXN_SEARCH_BB_MEM_ALIGN);
2540 snew_aligned(nbl->work->pbb_ci, GPU_NSUBCELL/STRIDE_PBB*NNBSBB_XXXX, NBNXN_SEARCH_BB_MEM_ALIGN);
2542 snew_aligned(nbl->work->bb_ci, GPU_NSUBCELL, NBNXN_SEARCH_BB_MEM_ALIGN);
2545 snew_aligned(nbl->work->x_ci, NBNXN_NA_SC_MAX*DIM, NBNXN_SEARCH_BB_MEM_ALIGN);
2546 #ifdef GMX_NBNXN_SIMD
2547 snew_aligned(nbl->work->x_ci_simd_4xn, 1, NBNXN_MEM_ALIGN);
2548 snew_aligned(nbl->work->x_ci_simd_2xnn, 1, NBNXN_MEM_ALIGN);
2550 snew_aligned(nbl->work->d2, GPU_NSUBCELL, NBNXN_SEARCH_BB_MEM_ALIGN);
2552 nbl->work->sort = NULL;
2553 nbl->work->sort_nalloc = 0;
2554 nbl->work->sci_sort = NULL;
2555 nbl->work->sci_sort_nalloc = 0;
2558 void nbnxn_init_pairlist_set(nbnxn_pairlist_set_t *nbl_list,
2559 gmx_bool bSimple, gmx_bool bCombined,
2560 nbnxn_alloc_t *alloc,
2565 nbl_list->bSimple = bSimple;
2566 nbl_list->bCombined = bCombined;
2568 nbl_list->nnbl = gmx_omp_nthreads_get(emntNonbonded);
2570 if (!nbl_list->bCombined &&
2571 nbl_list->nnbl > NBNXN_BUFFERFLAG_MAX_THREADS)
2573 gmx_fatal(FARGS, "%d OpenMP threads were requested. Since the non-bonded force buffer reduction is prohibitively slow with more than %d threads, we do not allow this. Use %d or less OpenMP threads.",
2574 nbl_list->nnbl, NBNXN_BUFFERFLAG_MAX_THREADS, NBNXN_BUFFERFLAG_MAX_THREADS);
2577 snew(nbl_list->nbl, nbl_list->nnbl);
2578 snew(nbl_list->nbl_fep, nbl_list->nnbl);
2579 /* Execute in order to avoid memory interleaving between threads */
2580 #pragma omp parallel for num_threads(nbl_list->nnbl) schedule(static)
2581 for (i = 0; i < nbl_list->nnbl; i++)
2583 /* Allocate the nblist data structure locally on each thread
2584 * to optimize memory access for NUMA architectures.
2586 snew(nbl_list->nbl[i], 1);
2588 /* Only list 0 is used on the GPU, use normal allocation for i>0 */
2591 nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, alloc, free);
2595 nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, NULL, NULL);
2598 snew(nbl_list->nbl_fep[i], 1);
2599 nbnxn_init_pairlist_fep(nbl_list->nbl_fep[i]);
2603 /* Print statistics of a pair list, used for debug output */
2604 static void print_nblist_statistics_simple(FILE *fp, const nbnxn_pairlist_t *nbl,
2605 const nbnxn_search_t nbs, real rl)
2607 const nbnxn_grid_t *grid;
2612 /* This code only produces correct statistics with domain decomposition */
2613 grid = &nbs->grid[0];
2615 fprintf(fp, "nbl nci %d ncj %d\n",
2616 nbl->nci, nbl->ncj);
2617 fprintf(fp, "nbl na_sc %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
2618 nbl->na_sc, rl, nbl->ncj, nbl->ncj/(double)grid->nc,
2619 nbl->ncj/(double)grid->nc*grid->na_sc,
2620 nbl->ncj/(double)grid->nc*grid->na_sc/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid->nc*grid->na_sc/det(nbs->box)));
2622 fprintf(fp, "nbl average j cell list length %.1f\n",
2623 0.25*nbl->ncj/(double)nbl->nci);
2625 for (s = 0; s < SHIFTS; s++)
2630 for (i = 0; i < nbl->nci; i++)
2632 cs[nbl->ci[i].shift & NBNXN_CI_SHIFT] +=
2633 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start;
2635 j = nbl->ci[i].cj_ind_start;
2636 while (j < nbl->ci[i].cj_ind_end &&
2637 nbl->cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
2643 fprintf(fp, "nbl cell pairs, total: %d excl: %d %.1f%%\n",
2644 nbl->ncj, npexcl, 100*npexcl/(double)nbl->ncj);
2645 for (s = 0; s < SHIFTS; s++)
2649 fprintf(fp, "nbl shift %2d ncj %3d\n", s, cs[s]);
2654 /* Print statistics of a pair lists, used for debug output */
2655 static void print_nblist_statistics_supersub(FILE *fp, const nbnxn_pairlist_t *nbl,
2656 const nbnxn_search_t nbs, real rl)
2658 const nbnxn_grid_t *grid;
2659 int i, j4, j, si, b;
2660 int c[GPU_NSUBCELL+1];
2662 /* This code only produces correct statistics with domain decomposition */
2663 grid = &nbs->grid[0];
2665 fprintf(fp, "nbl nsci %d ncj4 %d nsi %d excl4 %d\n",
2666 nbl->nsci, nbl->ncj4, nbl->nci_tot, nbl->nexcl);
2667 fprintf(fp, "nbl na_c %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
2668 nbl->na_ci, rl, nbl->nci_tot, nbl->nci_tot/(double)grid->nsubc_tot,
2669 nbl->nci_tot/(double)grid->nsubc_tot*grid->na_c,
2670 nbl->nci_tot/(double)grid->nsubc_tot*grid->na_c/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid->nsubc_tot*grid->na_c/det(nbs->box)));
2672 fprintf(fp, "nbl average j super cell list length %.1f\n",
2673 0.25*nbl->ncj4/(double)nbl->nsci);
2674 fprintf(fp, "nbl average i sub cell list length %.1f\n",
2675 nbl->nci_tot/((double)nbl->ncj4));
2677 for (si = 0; si <= GPU_NSUBCELL; si++)
2681 for (i = 0; i < nbl->nsci; i++)
2683 for (j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
2685 for (j = 0; j < NBNXN_GPU_JGROUP_SIZE; j++)
2688 for (si = 0; si < GPU_NSUBCELL; si++)
2690 if (nbl->cj4[j4].imei[0].imask & (1U << (j*GPU_NSUBCELL + si)))
2699 for (b = 0; b <= GPU_NSUBCELL; b++)
2701 fprintf(fp, "nbl j-list #i-subcell %d %7d %4.1f\n",
2702 b, c[b], 100.0*c[b]/(double)(nbl->ncj4*NBNXN_GPU_JGROUP_SIZE));
2706 /* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp */
2707 static void low_get_nbl_exclusions(nbnxn_pairlist_t *nbl, int cj4,
2708 int warp, nbnxn_excl_t **excl)
2710 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
2712 /* No exclusions set, make a new list entry */
2713 nbl->cj4[cj4].imei[warp].excl_ind = nbl->nexcl;
2715 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
2716 set_no_excls(*excl);
2720 /* We already have some exclusions, new ones can be added to the list */
2721 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
2725 /* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp,
2726 * generates a new element and allocates extra memory, if necessary.
2728 static void get_nbl_exclusions_1(nbnxn_pairlist_t *nbl, int cj4,
2729 int warp, nbnxn_excl_t **excl)
2731 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
2733 /* We need to make a new list entry, check if we have space */
2734 check_excl_space(nbl, 1);
2736 low_get_nbl_exclusions(nbl, cj4, warp, excl);
2739 /* Returns pointers to the exclusion mask for cj4-unit cj4 for both warps,
2740 * generates a new element and allocates extra memory, if necessary.
2742 static void get_nbl_exclusions_2(nbnxn_pairlist_t *nbl, int cj4,
2743 nbnxn_excl_t **excl_w0,
2744 nbnxn_excl_t **excl_w1)
2746 /* Check for space we might need */
2747 check_excl_space(nbl, 2);
2749 low_get_nbl_exclusions(nbl, cj4, 0, excl_w0);
2750 low_get_nbl_exclusions(nbl, cj4, 1, excl_w1);
2753 /* Sets the self exclusions i=j and pair exclusions i>j */
2754 static void set_self_and_newton_excls_supersub(nbnxn_pairlist_t *nbl,
2755 int cj4_ind, int sj_offset,
2758 nbnxn_excl_t *excl[2];
2761 /* Here we only set the set self and double pair exclusions */
2763 get_nbl_exclusions_2(nbl, cj4_ind, &excl[0], &excl[1]);
2765 /* Only minor < major bits set */
2766 for (ej = 0; ej < nbl->na_ci; ej++)
2769 for (ei = ej; ei < nbl->na_ci; ei++)
2771 excl[w]->pair[(ej & (NBNXN_GPU_JGROUP_SIZE-1))*nbl->na_ci + ei] &=
2772 ~(1U << (sj_offset*GPU_NSUBCELL + si));
2777 /* Returns a diagonal or off-diagonal interaction mask for plain C lists */
2778 static unsigned int get_imask(gmx_bool rdiag, int ci, int cj)
2780 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
2783 /* Returns a diagonal or off-diagonal interaction mask for cj-size=2 */
2784 static unsigned int get_imask_simd_j2(gmx_bool rdiag, int ci, int cj)
2786 return (rdiag && ci*2 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_0 :
2787 (rdiag && ci*2+1 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_1 :
2788 NBNXN_INTERACTION_MASK_ALL));
2791 /* Returns a diagonal or off-diagonal interaction mask for cj-size=4 */
2792 static unsigned int get_imask_simd_j4(gmx_bool rdiag, int ci, int cj)
2794 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
2797 /* Returns a diagonal or off-diagonal interaction mask for cj-size=8 */
2798 static unsigned int get_imask_simd_j8(gmx_bool rdiag, int ci, int cj)
2800 return (rdiag && ci == cj*2 ? NBNXN_INTERACTION_MASK_DIAG_J8_0 :
2801 (rdiag && ci == cj*2+1 ? NBNXN_INTERACTION_MASK_DIAG_J8_1 :
2802 NBNXN_INTERACTION_MASK_ALL));
2805 #ifdef GMX_NBNXN_SIMD
2806 #if GMX_SIMD_REAL_WIDTH == 2
2807 #define get_imask_simd_4xn get_imask_simd_j2
2809 #if GMX_SIMD_REAL_WIDTH == 4
2810 #define get_imask_simd_4xn get_imask_simd_j4
2812 #if GMX_SIMD_REAL_WIDTH == 8
2813 #define get_imask_simd_4xn get_imask_simd_j8
2814 #define get_imask_simd_2xnn get_imask_simd_j4
2816 #if GMX_SIMD_REAL_WIDTH == 16
2817 #define get_imask_simd_2xnn get_imask_simd_j8
2821 /* Plain C code for making a pair list of cell ci vs cell cjf-cjl.
2822 * Checks bounding box distances and possibly atom pair distances.
2824 static void make_cluster_list_simple(const nbnxn_grid_t *gridj,
2825 nbnxn_pairlist_t *nbl,
2826 int ci, int cjf, int cjl,
2827 gmx_bool remove_sub_diag,
2829 real rl2, float rbb2,
2832 const nbnxn_list_work_t *work;
2834 const nbnxn_bb_t *bb_ci;
2839 int cjf_gl, cjl_gl, cj;
2843 bb_ci = nbl->work->bb_ci;
2844 x_ci = nbl->work->x_ci;
2847 while (!InRange && cjf <= cjl)
2849 d2 = subc_bb_dist2(0, bb_ci, cjf, gridj->bb);
2852 /* Check if the distance is within the distance where
2853 * we use only the bounding box distance rbb,
2854 * or within the cut-off and there is at least one atom pair
2855 * within the cut-off.
2865 cjf_gl = gridj->cell0 + cjf;
2866 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
2868 for (j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE; j++)
2870 InRange = InRange ||
2871 (sqr(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
2872 sqr(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
2873 sqr(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rl2);
2876 *ndistc += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
2889 while (!InRange && cjl > cjf)
2891 d2 = subc_bb_dist2(0, bb_ci, cjl, gridj->bb);
2894 /* Check if the distance is within the distance where
2895 * we use only the bounding box distance rbb,
2896 * or within the cut-off and there is at least one atom pair
2897 * within the cut-off.
2907 cjl_gl = gridj->cell0 + cjl;
2908 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
2910 for (j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE; j++)
2912 InRange = InRange ||
2913 (sqr(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
2914 sqr(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
2915 sqr(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rl2);
2918 *ndistc += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
2928 for (cj = cjf; cj <= cjl; cj++)
2930 /* Store cj and the interaction mask */
2931 nbl->cj[nbl->ncj].cj = gridj->cell0 + cj;
2932 nbl->cj[nbl->ncj].excl = get_imask(remove_sub_diag, ci, cj);
2935 /* Increase the closing index in i super-cell list */
2936 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
2940 #ifdef GMX_NBNXN_SIMD_4XN
2941 #include "nbnxn_search_simd_4xn.h"
2943 #ifdef GMX_NBNXN_SIMD_2XNN
2944 #include "nbnxn_search_simd_2xnn.h"
2947 /* Plain C or SIMD4 code for making a pair list of super-cell sci vs scj.
2948 * Checks bounding box distances and possibly atom pair distances.
2950 static void make_cluster_list_supersub(const nbnxn_grid_t *gridi,
2951 const nbnxn_grid_t *gridj,
2952 nbnxn_pairlist_t *nbl,
2954 gmx_bool sci_equals_scj,
2955 int stride, const real *x,
2956 real rl2, float rbb2,
2961 int cjo, ci1, ci, cj, cj_gl;
2962 int cj4_ind, cj_offset;
2966 const float *pbb_ci;
2968 const nbnxn_bb_t *bb_ci;
2973 #define PRUNE_LIST_CPU_ONE
2974 #ifdef PRUNE_LIST_CPU_ONE
2978 d2l = nbl->work->d2;
2981 pbb_ci = nbl->work->pbb_ci;
2983 bb_ci = nbl->work->bb_ci;
2985 x_ci = nbl->work->x_ci;
2989 for (cjo = 0; cjo < gridj->nsubc[scj]; cjo++)
2991 cj4_ind = (nbl->work->cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG);
2992 cj_offset = nbl->work->cj_ind - cj4_ind*NBNXN_GPU_JGROUP_SIZE;
2993 cj4 = &nbl->cj4[cj4_ind];
2995 cj = scj*GPU_NSUBCELL + cjo;
2997 cj_gl = gridj->cell0*GPU_NSUBCELL + cj;
2999 /* Initialize this j-subcell i-subcell list */
3000 cj4->cj[cj_offset] = cj_gl;
3009 ci1 = gridi->nsubc[sci];
3013 /* Determine all ci1 bb distances in one call with SIMD4 */
3014 subc_bb_dist2_simd4_xxxx(gridj->pbb+(cj>>STRIDE_PBB_2LOG)*NNBSBB_XXXX+(cj & (STRIDE_PBB-1)),
3020 /* We use a fixed upper-bound instead of ci1 to help optimization */
3021 for (ci = 0; ci < GPU_NSUBCELL; ci++)
3028 #ifndef NBNXN_BBXXXX
3029 /* Determine the bb distance between ci and cj */
3030 d2l[ci] = subc_bb_dist2(ci, bb_ci, cj, gridj->bb);
3035 #ifdef PRUNE_LIST_CPU_ALL
3036 /* Check if the distance is within the distance where
3037 * we use only the bounding box distance rbb,
3038 * or within the cut-off and there is at least one atom pair
3039 * within the cut-off. This check is very costly.
3041 *ndistc += na_c*na_c;
3044 #ifdef NBNXN_PBB_SIMD4
3049 (na_c, ci, x_ci, cj_gl, stride, x, rl2)))
3051 /* Check if the distance between the two bounding boxes
3052 * in within the pair-list cut-off.
3057 /* Flag this i-subcell to be taken into account */
3058 imask |= (1U << (cj_offset*GPU_NSUBCELL+ci));
3060 #ifdef PRUNE_LIST_CPU_ONE
3068 #ifdef PRUNE_LIST_CPU_ONE
3069 /* If we only found 1 pair, check if any atoms are actually
3070 * within the cut-off, so we could get rid of it.
3072 if (npair == 1 && d2l[ci_last] >= rbb2)
3074 /* Avoid using function pointers here, as it's slower */
3076 #ifdef NBNXN_PBB_SIMD4
3077 !subc_in_range_simd4
3081 (na_c, ci_last, x_ci, cj_gl, stride, x, rl2))
3083 imask &= ~(1U << (cj_offset*GPU_NSUBCELL+ci_last));
3091 /* We have a useful sj entry, close it now */
3093 /* Set the exclucions for the ci== sj entry.
3094 * Here we don't bother to check if this entry is actually flagged,
3095 * as it will nearly always be in the list.
3099 set_self_and_newton_excls_supersub(nbl, cj4_ind, cj_offset, cjo);
3102 /* Copy the cluster interaction mask to the list */
3103 for (w = 0; w < NWARP; w++)
3105 cj4->imei[w].imask |= imask;
3108 nbl->work->cj_ind++;
3110 /* Keep the count */
3111 nbl->nci_tot += npair;
3113 /* Increase the closing index in i super-cell list */
3114 nbl->sci[nbl->nsci].cj4_ind_end =
3115 ((nbl->work->cj_ind+NBNXN_GPU_JGROUP_SIZE-1) >> NBNXN_GPU_JGROUP_SIZE_2LOG);
3120 /* Set all atom-pair exclusions from the topology stored in excl
3121 * as masks in the pair-list for simple list i-entry nbl_ci
3123 static void set_ci_top_excls(const nbnxn_search_t nbs,
3124 nbnxn_pairlist_t *nbl,
3125 gmx_bool diagRemoved,
3128 const nbnxn_ci_t *nbl_ci,
3129 const t_blocka *excl)
3133 int cj_ind_first, cj_ind_last;
3134 int cj_first, cj_last;
3136 int i, ai, aj, si, eind, ge, se;
3137 int found, cj_ind_0, cj_ind_1, cj_ind_m;
3141 nbnxn_excl_t *nbl_excl;
3142 int inner_i, inner_e;
3146 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
3154 cj_ind_first = nbl_ci->cj_ind_start;
3155 cj_ind_last = nbl->ncj - 1;
3157 cj_first = nbl->cj[cj_ind_first].cj;
3158 cj_last = nbl->cj[cj_ind_last].cj;
3160 /* Determine how many contiguous j-cells we have starting
3161 * from the first i-cell. This number can be used to directly
3162 * calculate j-cell indices for excluded atoms.
3165 if (na_ci_2log == na_cj_2log)
3167 while (cj_ind_first + ndirect <= cj_ind_last &&
3168 nbl->cj[cj_ind_first+ndirect].cj == ci + ndirect)
3173 #ifdef NBNXN_SEARCH_BB_SIMD4
3176 while (cj_ind_first + ndirect <= cj_ind_last &&
3177 nbl->cj[cj_ind_first+ndirect].cj == ci_to_cj(na_cj_2log, ci) + ndirect)
3184 /* Loop over the atoms in the i super-cell */
3185 for (i = 0; i < nbl->na_sc; i++)
3187 ai = nbs->a[ci*nbl->na_sc+i];
3190 si = (i>>na_ci_2log);
3192 /* Loop over the topology-based exclusions for this i-atom */
3193 for (eind = excl->index[ai]; eind < excl->index[ai+1]; eind++)
3199 /* The self exclusion are already set, save some time */
3205 /* Without shifts we only calculate interactions j>i
3206 * for one-way pair-lists.
3208 if (diagRemoved && ge <= ci*nbl->na_sc + i)
3213 se = (ge >> na_cj_2log);
3215 /* Could the cluster se be in our list? */
3216 if (se >= cj_first && se <= cj_last)
3218 if (se < cj_first + ndirect)
3220 /* We can calculate cj_ind directly from se */
3221 found = cj_ind_first + se - cj_first;
3225 /* Search for se using bisection */
3227 cj_ind_0 = cj_ind_first + ndirect;
3228 cj_ind_1 = cj_ind_last + 1;
3229 while (found == -1 && cj_ind_0 < cj_ind_1)
3231 cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
3233 cj_m = nbl->cj[cj_ind_m].cj;
3241 cj_ind_1 = cj_ind_m;
3245 cj_ind_0 = cj_ind_m + 1;
3252 inner_i = i - (si << na_ci_2log);
3253 inner_e = ge - (se << na_cj_2log);
3255 nbl->cj[found].excl &= ~(1U<<((inner_i<<na_cj_2log) + inner_e));
3256 /* The next code line is usually not needed. We do not want to version
3257 * away the above line, because there is logic that relies on being
3258 * able to detect easily whether any exclusions exist. */
3259 #if (defined GMX_SIMD_IBM_QPX)
3260 nbl->cj[found].interaction_mask_indices[inner_i] &= ~(1U << inner_e);
3269 /* Add a new i-entry to the FEP list and copy the i-properties */
3270 static gmx_inline void fep_list_new_nri_copy(t_nblist *nlist)
3272 /* Add a new i-entry */
3275 assert(nlist->nri < nlist->maxnri);
3277 /* Duplicate the last i-entry, except for jindex, which continues */
3278 nlist->iinr[nlist->nri] = nlist->iinr[nlist->nri-1];
3279 nlist->shift[nlist->nri] = nlist->shift[nlist->nri-1];
3280 nlist->gid[nlist->nri] = nlist->gid[nlist->nri-1];
3281 nlist->jindex[nlist->nri] = nlist->nrj;
3284 /* For load balancing of the free-energy lists over threads, we set
3285 * the maximum nrj size of an i-entry to 40. This leads to good
3286 * load balancing in the worst case scenario of a single perturbed
3287 * particle on 16 threads, while not introducing significant overhead.
3288 * Note that half of the perturbed pairs will anyhow end up in very small lists,
3289 * since non perturbed i-particles will see few perturbed j-particles).
3291 const int max_nrj_fep = 40;
3293 /* Exclude the perturbed pairs from the Verlet list. This is only done to avoid
3294 * singularities for overlapping particles (0/0), since the charges and
3295 * LJ parameters have been zeroed in the nbnxn data structure.
3296 * Simultaneously make a group pair list for the perturbed pairs.
3298 static void make_fep_list(const nbnxn_search_t nbs,
3299 const nbnxn_atomdata_t *nbat,
3300 nbnxn_pairlist_t *nbl,
3301 gmx_bool bDiagRemoved,
3303 const nbnxn_grid_t *gridi,
3304 const nbnxn_grid_t *gridj,
3307 int ci, cj_ind_start, cj_ind_end, cj_ind, cja, cjr;
3309 int ngid, gid_i = 0, gid_j, gid;
3310 int egp_shift, egp_mask;
3312 int i, j, ind_i, ind_j, ai, aj;
3314 gmx_bool bFEP_i, bFEP_i_all;
3316 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
3324 cj_ind_start = nbl_ci->cj_ind_start;
3325 cj_ind_end = nbl_ci->cj_ind_end;
3327 /* In worst case we have alternating energy groups
3328 * and create #atom-pair lists, which means we need the size
3329 * of a cluster pair (na_ci*na_cj) times the number of cj's.
3331 nri_max = nbl->na_ci*nbl->na_cj*(cj_ind_end - cj_ind_start);
3332 if (nlist->nri + nri_max > nlist->maxnri)
3334 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
3335 reallocate_nblist(nlist);
3338 ngid = nbat->nenergrp;
3340 if (ngid*gridj->na_cj > sizeof(gid_cj)*8)
3342 gmx_fatal(FARGS, "The Verlet scheme with %dx%d kernels and free-energy only supports up to %d energy groups",
3343 gridi->na_c, gridj->na_cj, (sizeof(gid_cj)*8)/gridj->na_cj);
3346 egp_shift = nbat->neg_2log;
3347 egp_mask = (1<<nbat->neg_2log) - 1;
3349 /* Loop over the atoms in the i sub-cell */
3351 for (i = 0; i < nbl->na_ci; i++)
3353 ind_i = ci*nbl->na_ci + i;
3358 nlist->jindex[nri+1] = nlist->jindex[nri];
3359 nlist->iinr[nri] = ai;
3360 /* The actual energy group pair index is set later */
3361 nlist->gid[nri] = 0;
3362 nlist->shift[nri] = nbl_ci->shift & NBNXN_CI_SHIFT;
3364 bFEP_i = gridi->fep[ci - gridi->cell0] & (1 << i);
3366 bFEP_i_all = bFEP_i_all && bFEP_i;
3368 if ((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj > nlist->maxnrj)
3370 nlist->maxnrj = over_alloc_small((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj);
3371 srenew(nlist->jjnr, nlist->maxnrj);
3372 srenew(nlist->excl_fep, nlist->maxnrj);
3377 gid_i = (nbat->energrp[ci] >> (egp_shift*i)) & egp_mask;
3380 for (cj_ind = cj_ind_start; cj_ind < cj_ind_end; cj_ind++)
3382 unsigned int fep_cj;
3384 cja = nbl->cj[cj_ind].cj;
3386 if (gridj->na_cj == gridj->na_c)
3388 cjr = cja - gridj->cell0;
3389 fep_cj = gridj->fep[cjr];
3392 gid_cj = nbat->energrp[cja];
3395 else if (2*gridj->na_cj == gridj->na_c)
3397 cjr = cja - gridj->cell0*2;
3398 /* Extract half of the ci fep/energrp mask */
3399 fep_cj = (gridj->fep[cjr>>1] >> ((cjr&1)*gridj->na_cj)) & ((1<<gridj->na_cj) - 1);
3402 gid_cj = nbat->energrp[cja>>1] >> ((cja&1)*gridj->na_cj*egp_shift) & ((1<<(gridj->na_cj*egp_shift)) - 1);
3407 cjr = cja - (gridj->cell0>>1);
3408 /* Combine two ci fep masks/energrp */
3409 fep_cj = gridj->fep[cjr*2] + (gridj->fep[cjr*2+1] << gridj->na_c);
3412 gid_cj = nbat->energrp[cja*2] + (nbat->energrp[cja*2+1] << (gridj->na_c*egp_shift));
3416 if (bFEP_i || fep_cj != 0)
3418 for (j = 0; j < nbl->na_cj; j++)
3420 /* Is this interaction perturbed and not excluded? */
3421 ind_j = cja*nbl->na_cj + j;
3424 (bFEP_i || (fep_cj & (1 << j))) &&
3425 (!bDiagRemoved || ind_j >= ind_i))
3429 gid_j = (gid_cj >> (j*egp_shift)) & egp_mask;
3430 gid = GID(gid_i, gid_j, ngid);
3432 if (nlist->nrj > nlist->jindex[nri] &&
3433 nlist->gid[nri] != gid)
3435 /* Energy group pair changed: new list */
3436 fep_list_new_nri_copy(nlist);
3439 nlist->gid[nri] = gid;
3442 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
3444 fep_list_new_nri_copy(nlist);
3448 /* Add it to the FEP list */
3449 nlist->jjnr[nlist->nrj] = aj;
3450 nlist->excl_fep[nlist->nrj] = (nbl->cj[cj_ind].excl >> (i*nbl->na_cj + j)) & 1;
3453 /* Exclude it from the normal list.
3454 * Note that the charge has been set to zero,
3455 * but we need to avoid 0/0, as perturbed atoms
3456 * can be on top of each other.
3458 nbl->cj[cj_ind].excl &= ~(1U << (i*nbl->na_cj + j));
3464 if (nlist->nrj > nlist->jindex[nri])
3466 /* Actually add this new, non-empty, list */
3468 nlist->jindex[nlist->nri] = nlist->nrj;
3475 /* All interactions are perturbed, we can skip this entry */
3476 nbl_ci->cj_ind_end = cj_ind_start;
3480 /* Return the index of atom a within a cluster */
3481 static gmx_inline int cj_mod_cj4(int cj)
3483 return cj & (NBNXN_GPU_JGROUP_SIZE - 1);
3486 /* Convert a j-cluster to a cj4 group */
3487 static gmx_inline int cj_to_cj4(int cj)
3489 return cj >> NBNXN_GPU_JGROUP_SIZE_2LOG;
3492 /* Return the index of an j-atom within a warp */
3493 static gmx_inline int a_mod_wj(int a)
3495 return a & (NBNXN_GPU_CLUSTER_SIZE/2 - 1);
3498 /* As make_fep_list above, but for super/sub lists. */
3499 static void make_fep_list_supersub(const nbnxn_search_t nbs,
3500 const nbnxn_atomdata_t *nbat,
3501 nbnxn_pairlist_t *nbl,
3502 gmx_bool bDiagRemoved,
3503 const nbnxn_sci_t *nbl_sci,
3508 const nbnxn_grid_t *gridi,
3509 const nbnxn_grid_t *gridj,
3512 int sci, cj4_ind_start, cj4_ind_end, cj4_ind, gcj, cjr;
3515 int i, j, ind_i, ind_j, ai, aj;
3519 const nbnxn_cj4_t *cj4;
3521 if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
3529 cj4_ind_start = nbl_sci->cj4_ind_start;
3530 cj4_ind_end = nbl_sci->cj4_ind_end;
3532 /* Here we process one super-cell, max #atoms na_sc, versus a list
3533 * cj4 entries, each with max NBNXN_GPU_JGROUP_SIZE cj's, each
3534 * of size na_cj atoms.
3535 * On the GPU we don't support energy groups (yet).
3536 * So for each of the na_sc i-atoms, we need max one FEP list
3537 * for each max_nrj_fep j-atoms.
3539 nri_max = nbl->na_sc*nbl->na_cj*(1 + ((cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE)/max_nrj_fep);
3540 if (nlist->nri + nri_max > nlist->maxnri)
3542 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
3543 reallocate_nblist(nlist);
3546 /* Loop over the atoms in the i super-cluster */
3547 for (c = 0; c < GPU_NSUBCELL; c++)
3549 c_abs = sci*GPU_NSUBCELL + c;
3551 for (i = 0; i < nbl->na_ci; i++)
3553 ind_i = c_abs*nbl->na_ci + i;
3558 nlist->jindex[nri+1] = nlist->jindex[nri];
3559 nlist->iinr[nri] = ai;
3560 /* With GPUs, energy groups are not supported */
3561 nlist->gid[nri] = 0;
3562 nlist->shift[nri] = nbl_sci->shift & NBNXN_CI_SHIFT;
3564 bFEP_i = (gridi->fep[c_abs - gridi->cell0] & (1 << i));
3566 xi = nbat->x[ind_i*nbat->xstride+XX] + shx;
3567 yi = nbat->x[ind_i*nbat->xstride+YY] + shy;
3568 zi = nbat->x[ind_i*nbat->xstride+ZZ] + shz;
3570 if ((nlist->nrj + cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE*nbl->na_cj > nlist->maxnrj)
3572 nlist->maxnrj = over_alloc_small((nlist->nrj + cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE*nbl->na_cj);
3573 srenew(nlist->jjnr, nlist->maxnrj);
3574 srenew(nlist->excl_fep, nlist->maxnrj);
3577 for (cj4_ind = cj4_ind_start; cj4_ind < cj4_ind_end; cj4_ind++)
3579 cj4 = &nbl->cj4[cj4_ind];
3581 for (gcj = 0; gcj < NBNXN_GPU_JGROUP_SIZE; gcj++)
3583 unsigned int fep_cj;
3585 if ((cj4->imei[0].imask & (1U << (gcj*GPU_NSUBCELL + c))) == 0)
3587 /* Skip this ci for this cj */
3591 cjr = cj4->cj[gcj] - gridj->cell0*GPU_NSUBCELL;
3593 fep_cj = gridj->fep[cjr];
3595 if (bFEP_i || fep_cj != 0)
3597 for (j = 0; j < nbl->na_cj; j++)
3599 /* Is this interaction perturbed and not excluded? */
3600 ind_j = (gridj->cell0*GPU_NSUBCELL + cjr)*nbl->na_cj + j;
3603 (bFEP_i || (fep_cj & (1 << j))) &&
3604 (!bDiagRemoved || ind_j >= ind_i))
3608 unsigned int excl_bit;
3611 get_nbl_exclusions_1(nbl, cj4_ind, j>>2, &excl);
3613 excl_pair = a_mod_wj(j)*nbl->na_ci + i;
3614 excl_bit = (1U << (gcj*GPU_NSUBCELL + c));
3616 dx = nbat->x[ind_j*nbat->xstride+XX] - xi;
3617 dy = nbat->x[ind_j*nbat->xstride+YY] - yi;
3618 dz = nbat->x[ind_j*nbat->xstride+ZZ] - zi;
3620 /* The unpruned GPU list has more than 2/3
3621 * of the atom pairs beyond rlist. Using
3622 * this list will cause a lot of overhead
3623 * in the CPU FEP kernels, especially
3624 * relative to the fast GPU kernels.
3625 * So we prune the FEP list here.
3627 if (dx*dx + dy*dy + dz*dz < rlist_fep2)
3629 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
3631 fep_list_new_nri_copy(nlist);
3635 /* Add it to the FEP list */
3636 nlist->jjnr[nlist->nrj] = aj;
3637 nlist->excl_fep[nlist->nrj] = (excl->pair[excl_pair] & excl_bit) ? 1 : 0;
3641 /* Exclude it from the normal list.
3642 * Note that the charge and LJ parameters have
3643 * been set to zero, but we need to avoid 0/0,
3644 * as perturbed atoms can be on top of each other.
3646 excl->pair[excl_pair] &= ~excl_bit;
3650 /* Note that we could mask out this pair in imask
3651 * if all i- and/or all j-particles are perturbed.
3652 * But since the perturbed pairs on the CPU will
3653 * take an order of magnitude more time, the GPU
3654 * will finish before the CPU and there is no gain.
3660 if (nlist->nrj > nlist->jindex[nri])
3662 /* Actually add this new, non-empty, list */
3664 nlist->jindex[nlist->nri] = nlist->nrj;
3671 /* Set all atom-pair exclusions from the topology stored in excl
3672 * as masks in the pair-list for i-super-cell entry nbl_sci
3674 static void set_sci_top_excls(const nbnxn_search_t nbs,
3675 nbnxn_pairlist_t *nbl,
3676 gmx_bool diagRemoved,
3678 const nbnxn_sci_t *nbl_sci,
3679 const t_blocka *excl)
3684 int cj_ind_first, cj_ind_last;
3685 int cj_first, cj_last;
3687 int i, ai, aj, si, eind, ge, se;
3688 int found, cj_ind_0, cj_ind_1, cj_ind_m;
3692 nbnxn_excl_t *nbl_excl;
3693 int inner_i, inner_e, w;
3699 if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
3707 cj_ind_first = nbl_sci->cj4_ind_start*NBNXN_GPU_JGROUP_SIZE;
3708 cj_ind_last = nbl->work->cj_ind - 1;
3710 cj_first = nbl->cj4[nbl_sci->cj4_ind_start].cj[0];
3711 cj_last = nbl_cj(nbl, cj_ind_last);
3713 /* Determine how many contiguous j-clusters we have starting
3714 * from the first i-cluster. This number can be used to directly
3715 * calculate j-cluster indices for excluded atoms.
3718 while (cj_ind_first + ndirect <= cj_ind_last &&
3719 nbl_cj(nbl, cj_ind_first+ndirect) == sci*GPU_NSUBCELL + ndirect)
3724 /* Loop over the atoms in the i super-cell */
3725 for (i = 0; i < nbl->na_sc; i++)
3727 ai = nbs->a[sci*nbl->na_sc+i];
3730 si = (i>>na_c_2log);
3732 /* Loop over the topology-based exclusions for this i-atom */
3733 for (eind = excl->index[ai]; eind < excl->index[ai+1]; eind++)
3739 /* The self exclusion are already set, save some time */
3745 /* Without shifts we only calculate interactions j>i
3746 * for one-way pair-lists.
3748 if (diagRemoved && ge <= sci*nbl->na_sc + i)
3754 /* Could the cluster se be in our list? */
3755 if (se >= cj_first && se <= cj_last)
3757 if (se < cj_first + ndirect)
3759 /* We can calculate cj_ind directly from se */
3760 found = cj_ind_first + se - cj_first;
3764 /* Search for se using bisection */
3766 cj_ind_0 = cj_ind_first + ndirect;
3767 cj_ind_1 = cj_ind_last + 1;
3768 while (found == -1 && cj_ind_0 < cj_ind_1)
3770 cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
3772 cj_m = nbl_cj(nbl, cj_ind_m);
3780 cj_ind_1 = cj_ind_m;
3784 cj_ind_0 = cj_ind_m + 1;
3791 inner_i = i - si*na_c;
3792 inner_e = ge - se*na_c;
3794 if (nbl_imask0(nbl, found) & (1U << (cj_mod_cj4(found)*GPU_NSUBCELL + si)))
3798 get_nbl_exclusions_1(nbl, cj_to_cj4(found), w, &nbl_excl);
3800 nbl_excl->pair[a_mod_wj(inner_e)*nbl->na_ci+inner_i] &=
3801 ~(1U << (cj_mod_cj4(found)*GPU_NSUBCELL + si));
3810 /* Reallocate the simple ci list for at least n entries */
3811 static void nb_realloc_ci(nbnxn_pairlist_t *nbl, int n)
3813 nbl->ci_nalloc = over_alloc_small(n);
3814 nbnxn_realloc_void((void **)&nbl->ci,
3815 nbl->nci*sizeof(*nbl->ci),
3816 nbl->ci_nalloc*sizeof(*nbl->ci),
3817 nbl->alloc, nbl->free);
3820 /* Reallocate the super-cell sci list for at least n entries */
3821 static void nb_realloc_sci(nbnxn_pairlist_t *nbl, int n)
3823 nbl->sci_nalloc = over_alloc_small(n);
3824 nbnxn_realloc_void((void **)&nbl->sci,
3825 nbl->nsci*sizeof(*nbl->sci),
3826 nbl->sci_nalloc*sizeof(*nbl->sci),
3827 nbl->alloc, nbl->free);
3830 /* Make a new ci entry at index nbl->nci */
3831 static void new_ci_entry(nbnxn_pairlist_t *nbl, int ci, int shift, int flags)
3833 if (nbl->nci + 1 > nbl->ci_nalloc)
3835 nb_realloc_ci(nbl, nbl->nci+1);
3837 nbl->ci[nbl->nci].ci = ci;
3838 nbl->ci[nbl->nci].shift = shift;
3839 /* Store the interaction flags along with the shift */
3840 nbl->ci[nbl->nci].shift |= flags;
3841 nbl->ci[nbl->nci].cj_ind_start = nbl->ncj;
3842 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
3845 /* Make a new sci entry at index nbl->nsci */
3846 static void new_sci_entry(nbnxn_pairlist_t *nbl, int sci, int shift)
3848 if (nbl->nsci + 1 > nbl->sci_nalloc)
3850 nb_realloc_sci(nbl, nbl->nsci+1);
3852 nbl->sci[nbl->nsci].sci = sci;
3853 nbl->sci[nbl->nsci].shift = shift;
3854 nbl->sci[nbl->nsci].cj4_ind_start = nbl->ncj4;
3855 nbl->sci[nbl->nsci].cj4_ind_end = nbl->ncj4;
3858 /* Sort the simple j-list cj on exclusions.
3859 * Entries with exclusions will all be sorted to the beginning of the list.
3861 static void sort_cj_excl(nbnxn_cj_t *cj, int ncj,
3862 nbnxn_list_work_t *work)
3866 if (ncj > work->cj_nalloc)
3868 work->cj_nalloc = over_alloc_large(ncj);
3869 srenew(work->cj, work->cj_nalloc);
3872 /* Make a list of the j-cells involving exclusions */
3874 for (j = 0; j < ncj; j++)
3876 if (cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
3878 work->cj[jnew++] = cj[j];
3881 /* Check if there are exclusions at all or not just the first entry */
3882 if (!((jnew == 0) ||
3883 (jnew == 1 && cj[0].excl != NBNXN_INTERACTION_MASK_ALL)))
3885 for (j = 0; j < ncj; j++)
3887 if (cj[j].excl == NBNXN_INTERACTION_MASK_ALL)
3889 work->cj[jnew++] = cj[j];
3892 for (j = 0; j < ncj; j++)
3894 cj[j] = work->cj[j];
3899 /* Close this simple list i entry */
3900 static void close_ci_entry_simple(nbnxn_pairlist_t *nbl)
3904 /* All content of the new ci entry have already been filled correctly,
3905 * we only need to increase the count here (for non empty lists).
3907 jlen = nbl->ci[nbl->nci].cj_ind_end - nbl->ci[nbl->nci].cj_ind_start;
3910 sort_cj_excl(nbl->cj+nbl->ci[nbl->nci].cj_ind_start, jlen, nbl->work);
3912 /* The counts below are used for non-bonded pair/flop counts
3913 * and should therefore match the available kernel setups.
3915 if (!(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_COUL(0)))
3917 nbl->work->ncj_noq += jlen;
3919 else if ((nbl->ci[nbl->nci].shift & NBNXN_CI_HALF_LJ(0)) ||
3920 !(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_LJ(0)))
3922 nbl->work->ncj_hlj += jlen;
3929 /* Split sci entry for load balancing on the GPU.
3930 * Splitting ensures we have enough lists to fully utilize the whole GPU.
3931 * With progBal we generate progressively smaller lists, which improves
3932 * load balancing. As we only know the current count on our own thread,
3933 * we will need to estimate the current total amount of i-entries.
3934 * As the lists get concatenated later, this estimate depends
3935 * both on nthread and our own thread index.
3937 static void split_sci_entry(nbnxn_pairlist_t *nbl,
3938 int nsp_max_av, gmx_bool progBal, int nc_bal,
3939 int thread, int nthread)
3943 int cj4_start, cj4_end, j4len, cj4;
3945 int nsp, nsp_sci, nsp_cj4, nsp_cj4_e, nsp_cj4_p;
3950 /* Estimate the total numbers of ci's of the nblist combined
3951 * over all threads using the target number of ci's.
3953 nsci_est = nc_bal*thread/nthread + nbl->nsci;
3955 /* The first ci blocks should be larger, to avoid overhead.
3956 * The last ci blocks should be smaller, to improve load balancing.
3959 nsp_max_av*nc_bal*3/(2*(nsci_est - 1 + nc_bal)));
3963 nsp_max = nsp_max_av;
3966 cj4_start = nbl->sci[nbl->nsci-1].cj4_ind_start;
3967 cj4_end = nbl->sci[nbl->nsci-1].cj4_ind_end;
3968 j4len = cj4_end - cj4_start;
3970 if (j4len > 1 && j4len*GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE > nsp_max)
3972 /* Remove the last ci entry and process the cj4's again */
3980 for (cj4 = cj4_start; cj4 < cj4_end; cj4++)
3982 nsp_cj4_p = nsp_cj4;
3983 /* Count the number of cluster pairs in this cj4 group */
3985 for (p = 0; p < GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE; p++)
3987 nsp_cj4 += (nbl->cj4[cj4].imei[0].imask >> p) & 1;
3990 if (nsp_cj4 > 0 && nsp + nsp_cj4 > nsp_max)
3992 /* Split the list at cj4 */
3993 nbl->sci[sci].cj4_ind_end = cj4;
3994 /* Create a new sci entry */
3997 if (nbl->nsci+1 > nbl->sci_nalloc)
3999 nb_realloc_sci(nbl, nbl->nsci+1);
4001 nbl->sci[sci].sci = nbl->sci[nbl->nsci-1].sci;
4002 nbl->sci[sci].shift = nbl->sci[nbl->nsci-1].shift;
4003 nbl->sci[sci].cj4_ind_start = cj4;
4005 nsp_cj4_e = nsp_cj4_p;
4011 /* Put the remaining cj4's in the last sci entry */
4012 nbl->sci[sci].cj4_ind_end = cj4_end;
4014 /* Possibly balance out the last two sci's
4015 * by moving the last cj4 of the second last sci.
4017 if (nsp_sci - nsp_cj4_e >= nsp + nsp_cj4_e)
4019 nbl->sci[sci-1].cj4_ind_end--;
4020 nbl->sci[sci].cj4_ind_start--;
4027 /* Clost this super/sub list i entry */
4028 static void close_ci_entry_supersub(nbnxn_pairlist_t *nbl,
4030 gmx_bool progBal, int nc_bal,
4031 int thread, int nthread)
4036 /* All content of the new ci entry have already been filled correctly,
4037 * we only need to increase the count here (for non empty lists).
4039 j4len = nbl->sci[nbl->nsci].cj4_ind_end - nbl->sci[nbl->nsci].cj4_ind_start;
4042 /* We can only have complete blocks of 4 j-entries in a list,
4043 * so round the count up before closing.
4045 nbl->ncj4 = ((nbl->work->cj_ind + NBNXN_GPU_JGROUP_SIZE - 1) >> NBNXN_GPU_JGROUP_SIZE_2LOG);
4046 nbl->work->cj_ind = nbl->ncj4*NBNXN_GPU_JGROUP_SIZE;
4052 /* Measure the size of the new entry and potentially split it */
4053 split_sci_entry(nbl, nsp_max_av, progBal, nc_bal, thread, nthread);
4058 /* Syncs the working array before adding another grid pair to the list */
4059 static void sync_work(nbnxn_pairlist_t *nbl)
4063 nbl->work->cj_ind = nbl->ncj4*NBNXN_GPU_JGROUP_SIZE;
4064 nbl->work->cj4_init = nbl->ncj4;
4068 /* Clears an nbnxn_pairlist_t data structure */
4069 static void clear_pairlist(nbnxn_pairlist_t *nbl)
4078 nbl->work->ncj_noq = 0;
4079 nbl->work->ncj_hlj = 0;
4082 /* Clears a group scheme pair list */
4083 static void clear_pairlist_fep(t_nblist *nl)
4087 if (nl->jindex == NULL)
4089 snew(nl->jindex, 1);
4094 /* Sets a simple list i-cell bounding box, including PBC shift */
4095 static gmx_inline void set_icell_bb_simple(const nbnxn_bb_t *bb, int ci,
4096 real shx, real shy, real shz,
4099 bb_ci->lower[BB_X] = bb[ci].lower[BB_X] + shx;
4100 bb_ci->lower[BB_Y] = bb[ci].lower[BB_Y] + shy;
4101 bb_ci->lower[BB_Z] = bb[ci].lower[BB_Z] + shz;
4102 bb_ci->upper[BB_X] = bb[ci].upper[BB_X] + shx;
4103 bb_ci->upper[BB_Y] = bb[ci].upper[BB_Y] + shy;
4104 bb_ci->upper[BB_Z] = bb[ci].upper[BB_Z] + shz;
4108 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
4109 static void set_icell_bbxxxx_supersub(const float *bb, int ci,
4110 real shx, real shy, real shz,
4115 ia = ci*(GPU_NSUBCELL>>STRIDE_PBB_2LOG)*NNBSBB_XXXX;
4116 for (m = 0; m < (GPU_NSUBCELL>>STRIDE_PBB_2LOG)*NNBSBB_XXXX; m += NNBSBB_XXXX)
4118 for (i = 0; i < STRIDE_PBB; i++)
4120 bb_ci[m+0*STRIDE_PBB+i] = bb[ia+m+0*STRIDE_PBB+i] + shx;
4121 bb_ci[m+1*STRIDE_PBB+i] = bb[ia+m+1*STRIDE_PBB+i] + shy;
4122 bb_ci[m+2*STRIDE_PBB+i] = bb[ia+m+2*STRIDE_PBB+i] + shz;
4123 bb_ci[m+3*STRIDE_PBB+i] = bb[ia+m+3*STRIDE_PBB+i] + shx;
4124 bb_ci[m+4*STRIDE_PBB+i] = bb[ia+m+4*STRIDE_PBB+i] + shy;
4125 bb_ci[m+5*STRIDE_PBB+i] = bb[ia+m+5*STRIDE_PBB+i] + shz;
4131 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
4132 static void set_icell_bb_supersub(const nbnxn_bb_t *bb, int ci,
4133 real shx, real shy, real shz,
4138 for (i = 0; i < GPU_NSUBCELL; i++)
4140 set_icell_bb_simple(bb, ci*GPU_NSUBCELL+i,
4146 /* Copies PBC shifted i-cell atom coordinates x,y,z to working array */
4147 static void icell_set_x_simple(int ci,
4148 real shx, real shy, real shz,
4149 int gmx_unused na_c,
4150 int stride, const real *x,
4151 nbnxn_list_work_t *work)
4155 ia = ci*NBNXN_CPU_CLUSTER_I_SIZE;
4157 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE; i++)
4159 work->x_ci[i*STRIDE_XYZ+XX] = x[(ia+i)*stride+XX] + shx;
4160 work->x_ci[i*STRIDE_XYZ+YY] = x[(ia+i)*stride+YY] + shy;
4161 work->x_ci[i*STRIDE_XYZ+ZZ] = x[(ia+i)*stride+ZZ] + shz;
4165 /* Copies PBC shifted super-cell atom coordinates x,y,z to working array */
4166 static void icell_set_x_supersub(int ci,
4167 real shx, real shy, real shz,
4169 int stride, const real *x,
4170 nbnxn_list_work_t *work)
4177 ia = ci*GPU_NSUBCELL*na_c;
4178 for (i = 0; i < GPU_NSUBCELL*na_c; i++)
4180 x_ci[i*DIM + XX] = x[(ia+i)*stride + XX] + shx;
4181 x_ci[i*DIM + YY] = x[(ia+i)*stride + YY] + shy;
4182 x_ci[i*DIM + ZZ] = x[(ia+i)*stride + ZZ] + shz;
4186 #ifdef NBNXN_SEARCH_BB_SIMD4
4187 /* Copies PBC shifted super-cell packed atom coordinates to working array */
4188 static void icell_set_x_supersub_simd4(int ci,
4189 real shx, real shy, real shz,
4191 int stride, const real *x,
4192 nbnxn_list_work_t *work)
4194 int si, io, ia, i, j;
4199 for (si = 0; si < GPU_NSUBCELL; si++)
4201 for (i = 0; i < na_c; i += STRIDE_PBB)
4204 ia = ci*GPU_NSUBCELL*na_c + io;
4205 for (j = 0; j < STRIDE_PBB; j++)
4207 x_ci[io*DIM + j + XX*STRIDE_PBB] = x[(ia+j)*stride+XX] + shx;
4208 x_ci[io*DIM + j + YY*STRIDE_PBB] = x[(ia+j)*stride+YY] + shy;
4209 x_ci[io*DIM + j + ZZ*STRIDE_PBB] = x[(ia+j)*stride+ZZ] + shz;
4216 static real minimum_subgrid_size_xy(const nbnxn_grid_t *grid)
4220 return min(grid->sx, grid->sy);
4224 return min(grid->sx/GPU_NSUBCELL_X, grid->sy/GPU_NSUBCELL_Y);
4228 static real effective_buffer_1x1_vs_MxN(const nbnxn_grid_t *gridi,
4229 const nbnxn_grid_t *gridj)
4231 const real eff_1x1_buffer_fac_overest = 0.1;
4233 /* Determine an atom-pair list cut-off buffer size for atom pairs,
4234 * to be added to rlist (including buffer) used for MxN.
4235 * This is for converting an MxN list to a 1x1 list. This means we can't
4236 * use the normal buffer estimate, as we have an MxN list in which
4237 * some atom pairs beyond rlist are missing. We want to capture
4238 * the beneficial effect of buffering by extra pairs just outside rlist,
4239 * while removing the useless pairs that are further away from rlist.
4240 * (Also the buffer could have been set manually not using the estimate.)
4241 * This buffer size is an overestimate.
4242 * We add 10% of the smallest grid sub-cell dimensions.
4243 * Note that the z-size differs per cell and we don't use this,
4244 * so we overestimate.
4245 * With PME, the 10% value gives a buffer that is somewhat larger
4246 * than the effective buffer with a tolerance of 0.005 kJ/mol/ps.
4247 * Smaller tolerances or using RF lead to a smaller effective buffer,
4248 * so 10% gives a safe overestimate.
4250 return eff_1x1_buffer_fac_overest*(minimum_subgrid_size_xy(gridi) +
4251 minimum_subgrid_size_xy(gridj));
4254 /* Clusters at the cut-off only increase rlist by 60% of their size */
4255 static real nbnxn_rlist_inc_outside_fac = 0.6;
4257 /* Due to the cluster size the effective pair-list is longer than
4258 * that of a simple atom pair-list. This function gives the extra distance.
4260 real nbnxn_get_rlist_effective_inc(int cluster_size_j, real atom_density)
4263 real vol_inc_i, vol_inc_j;
4265 /* We should get this from the setup, but currently it's the same for
4266 * all setups, including GPUs.
4268 cluster_size_i = NBNXN_CPU_CLUSTER_I_SIZE;
4270 vol_inc_i = (cluster_size_i - 1)/atom_density;
4271 vol_inc_j = (cluster_size_j - 1)/atom_density;
4273 return nbnxn_rlist_inc_outside_fac*pow(vol_inc_i + vol_inc_j, 1.0/3.0);
4276 /* Estimates the interaction volume^2 for non-local interactions */
4277 static real nonlocal_vol2(const gmx_domdec_zones_t *zones, rvec ls, real r)
4286 /* Here we simply add up the volumes of 1, 2 or 3 1D decomposition
4287 * not home interaction volume^2. As these volumes are not additive,
4288 * this is an overestimate, but it would only be significant in the limit
4289 * of small cells, where we anyhow need to split the lists into
4290 * as small parts as possible.
4293 for (z = 0; z < zones->n; z++)
4295 if (zones->shift[z][XX] + zones->shift[z][YY] + zones->shift[z][ZZ] == 1)
4300 for (d = 0; d < DIM; d++)
4302 if (zones->shift[z][d] == 0)
4306 za *= zones->size[z].x1[d] - zones->size[z].x0[d];
4310 /* 4 octants of a sphere */
4311 vold_est = 0.25*M_PI*r*r*r*r;
4312 /* 4 quarter pie slices on the edges */
4313 vold_est += 4*cl*M_PI/6.0*r*r*r;
4314 /* One rectangular volume on a face */
4315 vold_est += ca*0.5*r*r;
4317 vol2_est_tot += vold_est*za;
4321 return vol2_est_tot;
4324 /* Estimates the average size of a full j-list for super/sub setup */
4325 static int get_nsubpair_max(const nbnxn_search_t nbs,
4328 int min_ci_balanced)
4330 const nbnxn_grid_t *grid;
4332 real xy_diag2, r_eff_sup, vol_est, nsp_est, nsp_est_nl;
4335 grid = &nbs->grid[0];
4337 ls[XX] = (grid->c1[XX] - grid->c0[XX])/(grid->ncx*GPU_NSUBCELL_X);
4338 ls[YY] = (grid->c1[YY] - grid->c0[YY])/(grid->ncy*GPU_NSUBCELL_Y);
4339 ls[ZZ] = (grid->c1[ZZ] - grid->c0[ZZ])*grid->ncx*grid->ncy/(grid->nc*GPU_NSUBCELL_Z);
4341 /* The average squared length of the diagonal of a sub cell */
4342 xy_diag2 = ls[XX]*ls[XX] + ls[YY]*ls[YY] + ls[ZZ]*ls[ZZ];
4344 /* The formulas below are a heuristic estimate of the average nsj per si*/
4345 r_eff_sup = rlist + nbnxn_rlist_inc_outside_fac*sqr((grid->na_c - 1.0)/grid->na_c)*sqrt(xy_diag2/3);
4347 if (!nbs->DomDec || nbs->zones->n == 1)
4354 sqr(grid->atom_density/grid->na_c)*
4355 nonlocal_vol2(nbs->zones, ls, r_eff_sup);
4360 /* Sub-cell interacts with itself */
4361 vol_est = ls[XX]*ls[YY]*ls[ZZ];
4362 /* 6/2 rectangular volume on the faces */
4363 vol_est += (ls[XX]*ls[YY] + ls[XX]*ls[ZZ] + ls[YY]*ls[ZZ])*r_eff_sup;
4364 /* 12/2 quarter pie slices on the edges */
4365 vol_est += 2*(ls[XX] + ls[YY] + ls[ZZ])*0.25*M_PI*sqr(r_eff_sup);
4366 /* 4 octants of a sphere */
4367 vol_est += 0.5*4.0/3.0*M_PI*pow(r_eff_sup, 3);
4369 nsp_est = grid->nsubc_tot*vol_est*grid->atom_density/grid->na_c;
4371 /* Subtract the non-local pair count */
4372 nsp_est -= nsp_est_nl;
4376 fprintf(debug, "nsp_est local %5.1f non-local %5.1f\n",
4377 nsp_est, nsp_est_nl);
4382 nsp_est = nsp_est_nl;
4385 if (min_ci_balanced <= 0 || grid->nc >= min_ci_balanced || grid->nc == 0)
4387 /* We don't need to worry */
4392 /* Thus the (average) maximum j-list size should be as follows */
4393 nsubpair_max = max(1, (int)(nsp_est/min_ci_balanced+0.5));
4395 /* Since the target value is a maximum (this avoids high outliers,
4396 * which lead to load imbalance), not average, we add half the
4397 * number of pairs in a cj4 block to get the average about right.
4399 nsubpair_max += GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE/2;
4404 fprintf(debug, "nbl nsp estimate %.1f, nsubpair_max %d\n",
4405 nsp_est, nsubpair_max);
4408 return nsubpair_max;
4411 /* Debug list print function */
4412 static void print_nblist_ci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
4416 for (i = 0; i < nbl->nci; i++)
4418 fprintf(fp, "ci %4d shift %2d ncj %3d\n",
4419 nbl->ci[i].ci, nbl->ci[i].shift,
4420 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start);
4422 for (j = nbl->ci[i].cj_ind_start; j < nbl->ci[i].cj_ind_end; j++)
4424 fprintf(fp, " cj %5d imask %x\n",
4431 /* Debug list print function */
4432 static void print_nblist_sci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
4434 int i, j4, j, ncp, si;
4436 for (i = 0; i < nbl->nsci; i++)
4438 fprintf(fp, "ci %4d shift %2d ncj4 %2d\n",
4439 nbl->sci[i].sci, nbl->sci[i].shift,
4440 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start);
4443 for (j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
4445 for (j = 0; j < NBNXN_GPU_JGROUP_SIZE; j++)
4447 fprintf(fp, " sj %5d imask %x\n",
4449 nbl->cj4[j4].imei[0].imask);
4450 for (si = 0; si < GPU_NSUBCELL; si++)
4452 if (nbl->cj4[j4].imei[0].imask & (1U << (j*GPU_NSUBCELL + si)))
4459 fprintf(fp, "ci %4d shift %2d ncj4 %2d ncp %3d\n",
4460 nbl->sci[i].sci, nbl->sci[i].shift,
4461 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start,
4466 /* Combine pair lists *nbl generated on multiple threads nblc */
4467 static void combine_nblists(int nnbl, nbnxn_pairlist_t **nbl,
4468 nbnxn_pairlist_t *nblc)
4470 int nsci, ncj4, nexcl;
4475 gmx_incons("combine_nblists does not support simple lists");
4480 nexcl = nblc->nexcl;
4481 for (i = 0; i < nnbl; i++)
4483 nsci += nbl[i]->nsci;
4484 ncj4 += nbl[i]->ncj4;
4485 nexcl += nbl[i]->nexcl;
4488 if (nsci > nblc->sci_nalloc)
4490 nb_realloc_sci(nblc, nsci);
4492 if (ncj4 > nblc->cj4_nalloc)
4494 nblc->cj4_nalloc = over_alloc_small(ncj4);
4495 nbnxn_realloc_void((void **)&nblc->cj4,
4496 nblc->ncj4*sizeof(*nblc->cj4),
4497 nblc->cj4_nalloc*sizeof(*nblc->cj4),
4498 nblc->alloc, nblc->free);
4500 if (nexcl > nblc->excl_nalloc)
4502 nblc->excl_nalloc = over_alloc_small(nexcl);
4503 nbnxn_realloc_void((void **)&nblc->excl,
4504 nblc->nexcl*sizeof(*nblc->excl),
4505 nblc->excl_nalloc*sizeof(*nblc->excl),
4506 nblc->alloc, nblc->free);
4509 /* Each thread should copy its own data to the combined arrays,
4510 * as otherwise data will go back and forth between different caches.
4512 #pragma omp parallel for num_threads(gmx_omp_nthreads_get(emntPairsearch)) schedule(static)
4513 for (n = 0; n < nnbl; n++)
4520 const nbnxn_pairlist_t *nbli;
4522 /* Determine the offset in the combined data for our thread */
4523 sci_offset = nblc->nsci;
4524 cj4_offset = nblc->ncj4;
4525 ci_offset = nblc->nci_tot;
4526 excl_offset = nblc->nexcl;
4528 for (i = 0; i < n; i++)
4530 sci_offset += nbl[i]->nsci;
4531 cj4_offset += nbl[i]->ncj4;
4532 ci_offset += nbl[i]->nci_tot;
4533 excl_offset += nbl[i]->nexcl;
4538 for (i = 0; i < nbli->nsci; i++)
4540 nblc->sci[sci_offset+i] = nbli->sci[i];
4541 nblc->sci[sci_offset+i].cj4_ind_start += cj4_offset;
4542 nblc->sci[sci_offset+i].cj4_ind_end += cj4_offset;
4545 for (j4 = 0; j4 < nbli->ncj4; j4++)
4547 nblc->cj4[cj4_offset+j4] = nbli->cj4[j4];
4548 nblc->cj4[cj4_offset+j4].imei[0].excl_ind += excl_offset;
4549 nblc->cj4[cj4_offset+j4].imei[1].excl_ind += excl_offset;
4552 for (j4 = 0; j4 < nbli->nexcl; j4++)
4554 nblc->excl[excl_offset+j4] = nbli->excl[j4];
4558 for (n = 0; n < nnbl; n++)
4560 nblc->nsci += nbl[n]->nsci;
4561 nblc->ncj4 += nbl[n]->ncj4;
4562 nblc->nci_tot += nbl[n]->nci_tot;
4563 nblc->nexcl += nbl[n]->nexcl;
4567 static void balance_fep_lists(const nbnxn_search_t nbs,
4568 nbnxn_pairlist_set_t *nbl_lists)
4571 int nri_tot, nrj_tot, nrj_target;
4575 nnbl = nbl_lists->nnbl;
4579 /* Nothing to balance */
4583 /* Count the total i-lists and pairs */
4586 for (th = 0; th < nnbl; th++)
4588 nri_tot += nbl_lists->nbl_fep[th]->nri;
4589 nrj_tot += nbl_lists->nbl_fep[th]->nrj;
4592 nrj_target = (nrj_tot + nnbl - 1)/nnbl;
4594 assert(gmx_omp_nthreads_get(emntNonbonded) == nnbl);
4596 #pragma omp parallel for schedule(static) num_threads(nnbl)
4597 for (th = 0; th < nnbl; th++)
4601 nbl = nbs->work[th].nbl_fep;
4603 /* Note that here we allocate for the total size, instead of
4604 * a per-thread esimate (which is hard to obtain).
4606 if (nri_tot > nbl->maxnri)
4608 nbl->maxnri = over_alloc_large(nri_tot);
4609 reallocate_nblist(nbl);
4611 if (nri_tot > nbl->maxnri || nrj_tot > nbl->maxnrj)
4613 nbl->maxnrj = over_alloc_small(nrj_tot);
4614 srenew(nbl->jjnr, nbl->maxnrj);
4615 srenew(nbl->excl_fep, nbl->maxnrj);
4618 clear_pairlist_fep(nbl);
4621 /* Loop over the source lists and assign and copy i-entries */
4623 nbld = nbs->work[th_dest].nbl_fep;
4624 for (th = 0; th < nnbl; th++)
4629 nbls = nbl_lists->nbl_fep[th];
4631 for (i = 0; i < nbls->nri; i++)
4635 /* The number of pairs in this i-entry */
4636 nrj = nbls->jindex[i+1] - nbls->jindex[i];
4638 /* Decide if list th_dest is too large and we should procede
4639 * to the next destination list.
4641 if (th_dest+1 < nnbl && nbld->nrj > 0 &&
4642 nbld->nrj + nrj - nrj_target > nrj_target - nbld->nrj)
4645 nbld = nbs->work[th_dest].nbl_fep;
4648 nbld->iinr[nbld->nri] = nbls->iinr[i];
4649 nbld->gid[nbld->nri] = nbls->gid[i];
4650 nbld->shift[nbld->nri] = nbls->shift[i];
4652 for (j = nbls->jindex[i]; j < nbls->jindex[i+1]; j++)
4654 nbld->jjnr[nbld->nrj] = nbls->jjnr[j];
4655 nbld->excl_fep[nbld->nrj] = nbls->excl_fep[j];
4659 nbld->jindex[nbld->nri] = nbld->nrj;
4663 /* Swap the list pointers */
4664 for (th = 0; th < nnbl; th++)
4668 nbl_tmp = nbl_lists->nbl_fep[th];
4669 nbl_lists->nbl_fep[th] = nbs->work[th].nbl_fep;
4670 nbs->work[th].nbl_fep = nbl_tmp;
4674 fprintf(debug, "nbl_fep[%d] nri %4d nrj %4d\n",
4676 nbl_lists->nbl_fep[th]->nri,
4677 nbl_lists->nbl_fep[th]->nrj);
4682 /* Returns the next ci to be processes by our thread */
4683 static gmx_bool next_ci(const nbnxn_grid_t *grid,
4685 int nth, int ci_block,
4686 int *ci_x, int *ci_y,
4692 if (*ci_b == ci_block)
4694 /* Jump to the next block assigned to this task */
4695 *ci += (nth - 1)*ci_block;
4699 if (*ci >= grid->nc*conv)
4704 while (*ci >= grid->cxy_ind[*ci_x*grid->ncy + *ci_y + 1]*conv)
4707 if (*ci_y == grid->ncy)
4717 /* Returns the distance^2 for which we put cell pairs in the list
4718 * without checking atom pair distances. This is usually < rlist^2.
4720 static float boundingbox_only_distance2(const nbnxn_grid_t *gridi,
4721 const nbnxn_grid_t *gridj,
4725 /* If the distance between two sub-cell bounding boxes is less
4726 * than this distance, do not check the distance between
4727 * all particle pairs in the sub-cell, since then it is likely
4728 * that the box pair has atom pairs within the cut-off.
4729 * We use the nblist cut-off minus 0.5 times the average x/y diagonal
4730 * spacing of the sub-cells. Around 40% of the checked pairs are pruned.
4731 * Using more than 0.5 gains at most 0.5%.
4732 * If forces are calculated more than twice, the performance gain
4733 * in the force calculation outweighs the cost of checking.
4734 * Note that with subcell lists, the atom-pair distance check
4735 * is only performed when only 1 out of 8 sub-cells in within range,
4736 * this is because the GPU is much faster than the cpu.
4741 bbx = 0.5*(gridi->sx + gridj->sx);
4742 bby = 0.5*(gridi->sy + gridj->sy);
4745 bbx /= GPU_NSUBCELL_X;
4746 bby /= GPU_NSUBCELL_Y;
4749 rbb2 = sqr(max(0, rlist - 0.5*sqrt(bbx*bbx + bby*bby)));
4754 return (float)((1+GMX_FLOAT_EPS)*rbb2);
4758 static int get_ci_block_size(const nbnxn_grid_t *gridi,
4759 gmx_bool bDomDec, int nth)
4761 const int ci_block_enum = 5;
4762 const int ci_block_denom = 11;
4763 const int ci_block_min_atoms = 16;
4766 /* Here we decide how to distribute the blocks over the threads.
4767 * We use prime numbers to try to avoid that the grid size becomes
4768 * a multiple of the number of threads, which would lead to some
4769 * threads getting "inner" pairs and others getting boundary pairs,
4770 * which in turns will lead to load imbalance between threads.
4771 * Set the block size as 5/11/ntask times the average number of cells
4772 * in a y,z slab. This should ensure a quite uniform distribution
4773 * of the grid parts of the different thread along all three grid
4774 * zone boundaries with 3D domain decomposition. At the same time
4775 * the blocks will not become too small.
4777 ci_block = (gridi->nc*ci_block_enum)/(ci_block_denom*gridi->ncx*nth);
4779 /* Ensure the blocks are not too small: avoids cache invalidation */
4780 if (ci_block*gridi->na_sc < ci_block_min_atoms)
4782 ci_block = (ci_block_min_atoms + gridi->na_sc - 1)/gridi->na_sc;
4785 /* Without domain decomposition
4786 * or with less than 3 blocks per task, divide in nth blocks.
4788 if (!bDomDec || ci_block*3*nth > gridi->nc)
4790 ci_block = (gridi->nc + nth - 1)/nth;
4796 /* Generates the part of pair-list nbl assigned to our thread */
4797 static void nbnxn_make_pairlist_part(const nbnxn_search_t nbs,
4798 const nbnxn_grid_t *gridi,
4799 const nbnxn_grid_t *gridj,
4800 nbnxn_search_work_t *work,
4801 const nbnxn_atomdata_t *nbat,
4802 const t_blocka *excl,
4806 gmx_bool bFBufferFlag,
4809 int min_ci_balanced,
4811 nbnxn_pairlist_t *nbl,
4816 real rl2, rl_fep2 = 0;
4819 int ci_b, ci, ci_x, ci_y, ci_xy, cj;
4825 int conv_i, cell0_i;
4826 const nbnxn_bb_t *bb_i = NULL;
4828 const float *pbb_i = NULL;
4830 const float *bbcz_i, *bbcz_j;
4832 real bx0, bx1, by0, by1, bz0, bz1;
4834 real d2cx, d2z, d2z_cx, d2z_cy, d2zx, d2zxy, d2xy;
4835 int cxf, cxl, cyf, cyf_x, cyl;
4837 int c0, c1, cs, cf, cl;
4840 int gridi_flag_shift = 0, gridj_flag_shift = 0;
4841 unsigned int *gridj_flag = NULL;
4842 int ncj_old_i, ncj_old_j;
4844 nbs_cycle_start(&work->cc[enbsCCsearch]);
4846 if (gridj->bSimple != nbl->bSimple)
4848 gmx_incons("Grid incompatible with pair-list");
4852 nbl->na_sc = gridj->na_sc;
4853 nbl->na_ci = gridj->na_c;
4854 nbl->na_cj = nbnxn_kernel_to_cj_size(nb_kernel_type);
4855 na_cj_2log = get_2log(nbl->na_cj);
4861 /* Determine conversion of clusters to flag blocks */
4862 gridi_flag_shift = 0;
4863 while ((nbl->na_ci<<gridi_flag_shift) < NBNXN_BUFFERFLAG_SIZE)
4867 gridj_flag_shift = 0;
4868 while ((nbl->na_cj<<gridj_flag_shift) < NBNXN_BUFFERFLAG_SIZE)
4873 gridj_flag = work->buffer_flags.flag;
4876 copy_mat(nbs->box, box);
4878 rl2 = nbl->rlist*nbl->rlist;
4880 if (nbs->bFEP && !nbl->bSimple)
4882 /* Determine an atom-pair list cut-off distance for FEP atom pairs.
4883 * We should not simply use rlist, since then we would not have
4884 * the small, effective buffering of the NxN lists.
4885 * The buffer is on overestimate, but the resulting cost for pairs
4886 * beyond rlist is neglible compared to the FEP pairs within rlist.
4888 rl_fep2 = nbl->rlist + effective_buffer_1x1_vs_MxN(gridi, gridj);
4892 fprintf(debug, "nbl_fep atom-pair rlist %f\n", rl_fep2);
4894 rl_fep2 = rl_fep2*rl_fep2;
4897 rbb2 = boundingbox_only_distance2(gridi, gridj, nbl->rlist, nbl->bSimple);
4901 fprintf(debug, "nbl bounding box only distance %f\n", sqrt(rbb2));
4904 /* Set the shift range */
4905 for (d = 0; d < DIM; d++)
4907 /* Check if we need periodicity shifts.
4908 * Without PBC or with domain decomposition we don't need them.
4910 if (d >= ePBC2npbcdim(nbs->ePBC) || nbs->dd_dim[d])
4917 box[XX][XX] - fabs(box[YY][XX]) - fabs(box[ZZ][XX]) < sqrt(rl2))
4928 if (nbl->bSimple && !gridi->bSimple)
4930 conv_i = gridi->na_sc/gridj->na_sc;
4931 bb_i = gridi->bb_simple;
4932 bbcz_i = gridi->bbcz_simple;
4933 flags_i = gridi->flags_simple;
4948 /* We use the normal bounding box format for both grid types */
4951 bbcz_i = gridi->bbcz;
4952 flags_i = gridi->flags;
4954 cell0_i = gridi->cell0*conv_i;
4956 bbcz_j = gridj->bbcz;
4960 /* Blocks of the conversion factor - 1 give a large repeat count
4961 * combined with a small block size. This should result in good
4962 * load balancing for both small and large domains.
4964 ci_block = conv_i - 1;
4968 fprintf(debug, "nbl nc_i %d col.av. %.1f ci_block %d\n",
4969 gridi->nc, gridi->nc/(double)(gridi->ncx*gridi->ncy), ci_block);
4975 /* Initially ci_b and ci to 1 before where we want them to start,
4976 * as they will both be incremented in next_ci.
4979 ci = th*ci_block - 1;
4982 while (next_ci(gridi, conv_i, nth, ci_block, &ci_x, &ci_y, &ci_b, &ci))
4984 if (nbl->bSimple && flags_i[ci] == 0)
4989 ncj_old_i = nbl->ncj;
4992 if (gridj != gridi && shp[XX] == 0)
4996 bx1 = bb_i[ci].upper[BB_X];
5000 bx1 = gridi->c0[XX] + (ci_x+1)*gridi->sx;
5002 if (bx1 < gridj->c0[XX])
5004 d2cx = sqr(gridj->c0[XX] - bx1);
5013 ci_xy = ci_x*gridi->ncy + ci_y;
5015 /* Loop over shift vectors in three dimensions */
5016 for (tz = -shp[ZZ]; tz <= shp[ZZ]; tz++)
5018 shz = tz*box[ZZ][ZZ];
5020 bz0 = bbcz_i[ci*NNBSBB_D ] + shz;
5021 bz1 = bbcz_i[ci*NNBSBB_D+1] + shz;
5033 d2z = sqr(bz0 - box[ZZ][ZZ]);
5036 d2z_cx = d2z + d2cx;
5044 bz1/((real)(gridi->cxy_ind[ci_xy+1] - gridi->cxy_ind[ci_xy]));
5049 /* The check with bz1_frac close to or larger than 1 comes later */
5051 for (ty = -shp[YY]; ty <= shp[YY]; ty++)
5053 shy = ty*box[YY][YY] + tz*box[ZZ][YY];
5057 by0 = bb_i[ci].lower[BB_Y] + shy;
5058 by1 = bb_i[ci].upper[BB_Y] + shy;
5062 by0 = gridi->c0[YY] + (ci_y )*gridi->sy + shy;
5063 by1 = gridi->c0[YY] + (ci_y+1)*gridi->sy + shy;
5066 get_cell_range(by0, by1,
5067 gridj->ncy, gridj->c0[YY], gridj->sy, gridj->inv_sy,
5077 if (by1 < gridj->c0[YY])
5079 d2z_cy += sqr(gridj->c0[YY] - by1);
5081 else if (by0 > gridj->c1[YY])
5083 d2z_cy += sqr(by0 - gridj->c1[YY]);
5086 for (tx = -shp[XX]; tx <= shp[XX]; tx++)
5088 shift = XYZ2IS(tx, ty, tz);
5090 #ifdef NBNXN_SHIFT_BACKWARD
5091 if (gridi == gridj && shift > CENTRAL)
5097 shx = tx*box[XX][XX] + ty*box[YY][XX] + tz*box[ZZ][XX];
5101 bx0 = bb_i[ci].lower[BB_X] + shx;
5102 bx1 = bb_i[ci].upper[BB_X] + shx;
5106 bx0 = gridi->c0[XX] + (ci_x )*gridi->sx + shx;
5107 bx1 = gridi->c0[XX] + (ci_x+1)*gridi->sx + shx;
5110 get_cell_range(bx0, bx1,
5111 gridj->ncx, gridj->c0[XX], gridj->sx, gridj->inv_sx,
5122 new_ci_entry(nbl, cell0_i+ci, shift, flags_i[ci]);
5126 new_sci_entry(nbl, cell0_i+ci, shift);
5129 #ifndef NBNXN_SHIFT_BACKWARD
5132 if (shift == CENTRAL && gridi == gridj &&
5136 /* Leave the pairs with i > j.
5137 * x is the major index, so skip half of it.
5144 set_icell_bb_simple(bb_i, ci, shx, shy, shz,
5150 set_icell_bbxxxx_supersub(pbb_i, ci, shx, shy, shz,
5153 set_icell_bb_supersub(bb_i, ci, shx, shy, shz,
5158 nbs->icell_set_x(cell0_i+ci, shx, shy, shz,
5159 gridi->na_c, nbat->xstride, nbat->x,
5162 for (cx = cxf; cx <= cxl; cx++)
5165 if (gridj->c0[XX] + cx*gridj->sx > bx1)
5167 d2zx += sqr(gridj->c0[XX] + cx*gridj->sx - bx1);
5169 else if (gridj->c0[XX] + (cx+1)*gridj->sx < bx0)
5171 d2zx += sqr(gridj->c0[XX] + (cx+1)*gridj->sx - bx0);
5174 #ifndef NBNXN_SHIFT_BACKWARD
5175 if (gridi == gridj &&
5176 cx == 0 && cyf < ci_y)
5178 if (gridi == gridj &&
5179 cx == 0 && shift == CENTRAL && cyf < ci_y)
5182 /* Leave the pairs with i > j.
5183 * Skip half of y when i and j have the same x.
5192 for (cy = cyf_x; cy <= cyl; cy++)
5194 c0 = gridj->cxy_ind[cx*gridj->ncy+cy];
5195 c1 = gridj->cxy_ind[cx*gridj->ncy+cy+1];
5196 #ifdef NBNXN_SHIFT_BACKWARD
5197 if (gridi == gridj &&
5198 shift == CENTRAL && c0 < ci)
5205 if (gridj->c0[YY] + cy*gridj->sy > by1)
5207 d2zxy += sqr(gridj->c0[YY] + cy*gridj->sy - by1);
5209 else if (gridj->c0[YY] + (cy+1)*gridj->sy < by0)
5211 d2zxy += sqr(gridj->c0[YY] + (cy+1)*gridj->sy - by0);
5213 if (c1 > c0 && d2zxy < rl2)
5215 cs = c0 + (int)(bz1_frac*(c1 - c0));
5223 /* Find the lowest cell that can possibly
5228 (bbcz_j[cf*NNBSBB_D+1] >= bz0 ||
5229 d2xy + sqr(bbcz_j[cf*NNBSBB_D+1] - bz0) < rl2))
5234 /* Find the highest cell that can possibly
5239 (bbcz_j[cl*NNBSBB_D] <= bz1 ||
5240 d2xy + sqr(bbcz_j[cl*NNBSBB_D] - bz1) < rl2))
5245 #ifdef NBNXN_REFCODE
5247 /* Simple reference code, for debugging,
5248 * overrides the more complex code above.
5253 for (k = c0; k < c1; k++)
5255 if (box_dist2(bx0, bx1, by0, by1, bz0, bz1, bb+k) < rl2 &&
5260 if (box_dist2(bx0, bx1, by0, by1, bz0, bz1, bb+k) < rl2 &&
5271 /* We want each atom/cell pair only once,
5272 * only use cj >= ci.
5274 #ifndef NBNXN_SHIFT_BACKWARD
5277 if (shift == CENTRAL)
5286 /* For f buffer flags with simple lists */
5287 ncj_old_j = nbl->ncj;
5289 switch (nb_kernel_type)
5291 case nbnxnk4x4_PlainC:
5292 check_subcell_list_space_simple(nbl, cl-cf+1);
5294 make_cluster_list_simple(gridj,
5296 (gridi == gridj && shift == CENTRAL),
5301 #ifdef GMX_NBNXN_SIMD_4XN
5302 case nbnxnk4xN_SIMD_4xN:
5303 check_subcell_list_space_simple(nbl, ci_to_cj(na_cj_2log, cl-cf)+2);
5304 make_cluster_list_simd_4xn(gridj,
5306 (gridi == gridj && shift == CENTRAL),
5312 #ifdef GMX_NBNXN_SIMD_2XNN
5313 case nbnxnk4xN_SIMD_2xNN:
5314 check_subcell_list_space_simple(nbl, ci_to_cj(na_cj_2log, cl-cf)+2);
5315 make_cluster_list_simd_2xnn(gridj,
5317 (gridi == gridj && shift == CENTRAL),
5323 case nbnxnk8x8x8_PlainC:
5324 case nbnxnk8x8x8_CUDA:
5325 check_subcell_list_space_supersub(nbl, cl-cf+1);
5326 for (cj = cf; cj <= cl; cj++)
5328 make_cluster_list_supersub(gridi, gridj,
5330 (gridi == gridj && shift == CENTRAL && ci == cj),
5331 nbat->xstride, nbat->x,
5337 ncpcheck += cl - cf + 1;
5339 if (bFBufferFlag && nbl->ncj > ncj_old_j)
5343 cbf = nbl->cj[ncj_old_j].cj >> gridj_flag_shift;
5344 cbl = nbl->cj[nbl->ncj-1].cj >> gridj_flag_shift;
5345 for (cb = cbf; cb <= cbl; cb++)
5347 gridj_flag[cb] = 1U<<th;
5355 /* Set the exclusions for this ci list */
5358 set_ci_top_excls(nbs,
5360 shift == CENTRAL && gridi == gridj,
5363 &(nbl->ci[nbl->nci]),
5368 make_fep_list(nbs, nbat, nbl,
5369 shift == CENTRAL && gridi == gridj,
5370 &(nbl->ci[nbl->nci]),
5371 gridi, gridj, nbl_fep);
5376 set_sci_top_excls(nbs,
5378 shift == CENTRAL && gridi == gridj,
5380 &(nbl->sci[nbl->nsci]),
5385 make_fep_list_supersub(nbs, nbat, nbl,
5386 shift == CENTRAL && gridi == gridj,
5387 &(nbl->sci[nbl->nsci]),
5390 gridi, gridj, nbl_fep);
5394 /* Close this ci list */
5397 close_ci_entry_simple(nbl);
5401 close_ci_entry_supersub(nbl,
5403 progBal, min_ci_balanced,
5410 if (bFBufferFlag && nbl->ncj > ncj_old_i)
5412 work->buffer_flags.flag[(gridi->cell0+ci)>>gridi_flag_shift] = 1U<<th;
5416 work->ndistc = ndistc;
5418 nbs_cycle_stop(&work->cc[enbsCCsearch]);
5422 fprintf(debug, "number of distance checks %d\n", ndistc);
5423 fprintf(debug, "ncpcheck %s %d\n", gridi == gridj ? "local" : "non-local",
5428 print_nblist_statistics_simple(debug, nbl, nbs, rlist);
5432 print_nblist_statistics_supersub(debug, nbl, nbs, rlist);
5437 fprintf(debug, "nbl FEP list pairs: %d\n", nbl_fep->nrj);
5442 static void reduce_buffer_flags(const nbnxn_search_t nbs,
5444 const nbnxn_buffer_flags_t *dest)
5447 const unsigned int *flag;
5449 for (s = 0; s < nsrc; s++)
5451 flag = nbs->work[s].buffer_flags.flag;
5453 for (b = 0; b < dest->nflag; b++)
5455 dest->flag[b] |= flag[b];
5460 static void print_reduction_cost(const nbnxn_buffer_flags_t *flags, int nout)
5462 int nelem, nkeep, ncopy, nred, b, c, out;
5468 for (b = 0; b < flags->nflag; b++)
5470 if (flags->flag[b] == 1)
5472 /* Only flag 0 is set, no copy of reduction required */
5476 else if (flags->flag[b] > 0)
5479 for (out = 0; out < nout; out++)
5481 if (flags->flag[b] & (1U<<out))
5498 fprintf(debug, "nbnxn reduction: #flag %d #list %d elem %4.2f, keep %4.2f copy %4.2f red %4.2f\n",
5500 nelem/(double)(flags->nflag),
5501 nkeep/(double)(flags->nflag),
5502 ncopy/(double)(flags->nflag),
5503 nred/(double)(flags->nflag));
5506 /* Perform a count (linear) sort to sort the smaller lists to the end.
5507 * This avoids load imbalance on the GPU, as large lists will be
5508 * scheduled and executed first and the smaller lists later.
5509 * Load balancing between multi-processors only happens at the end
5510 * and there smaller lists lead to more effective load balancing.
5511 * The sorting is done on the cj4 count, not on the actual pair counts.
5512 * Not only does this make the sort faster, but it also results in
5513 * better load balancing than using a list sorted on exact load.
5514 * This function swaps the pointer in the pair list to avoid a copy operation.
5516 static void sort_sci(nbnxn_pairlist_t *nbl)
5518 nbnxn_list_work_t *work;
5519 int m, i, s, s0, s1;
5520 nbnxn_sci_t *sci_sort;
5522 if (nbl->ncj4 <= nbl->nsci)
5524 /* nsci = 0 or all sci have size 1, sorting won't change the order */
5530 /* We will distinguish differences up to double the average */
5531 m = (2*nbl->ncj4)/nbl->nsci;
5533 if (m + 1 > work->sort_nalloc)
5535 work->sort_nalloc = over_alloc_large(m + 1);
5536 srenew(work->sort, work->sort_nalloc);
5539 if (work->sci_sort_nalloc != nbl->sci_nalloc)
5541 work->sci_sort_nalloc = nbl->sci_nalloc;
5542 nbnxn_realloc_void((void **)&work->sci_sort,
5544 work->sci_sort_nalloc*sizeof(*work->sci_sort),
5545 nbl->alloc, nbl->free);
5548 /* Count the entries of each size */
5549 for (i = 0; i <= m; i++)
5553 for (s = 0; s < nbl->nsci; s++)
5555 i = min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start);
5558 /* Calculate the offset for each count */
5561 for (i = m - 1; i >= 0; i--)
5564 work->sort[i] = work->sort[i + 1] + s0;
5568 /* Sort entries directly into place */
5569 sci_sort = work->sci_sort;
5570 for (s = 0; s < nbl->nsci; s++)
5572 i = min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start);
5573 sci_sort[work->sort[i]++] = nbl->sci[s];
5576 /* Swap the sci pointers so we use the new, sorted list */
5577 work->sci_sort = nbl->sci;
5578 nbl->sci = sci_sort;
5581 /* Make a local or non-local pair-list, depending on iloc */
5582 void nbnxn_make_pairlist(const nbnxn_search_t nbs,
5583 nbnxn_atomdata_t *nbat,
5584 const t_blocka *excl,
5586 int min_ci_balanced,
5587 nbnxn_pairlist_set_t *nbl_list,
5592 nbnxn_grid_t *gridi, *gridj;
5594 int nzi, zi, zj0, zj1, zj;
5598 nbnxn_pairlist_t **nbl;
5600 gmx_bool CombineNBLists;
5602 int np_tot, np_noq, np_hlj, nap;
5604 /* Check if we are running hybrid GPU + CPU nbnxn mode */
5605 bGPUCPU = (!nbs->grid[0].bSimple && nbl_list->bSimple);
5607 nnbl = nbl_list->nnbl;
5608 nbl = nbl_list->nbl;
5609 CombineNBLists = nbl_list->bCombined;
5613 fprintf(debug, "ns making %d nblists\n", nnbl);
5616 nbat->bUseBufferFlags = (nbat->nout > 1);
5617 /* We should re-init the flags before making the first list */
5618 if (nbat->bUseBufferFlags && (LOCAL_I(iloc) || bGPUCPU))
5620 init_buffer_flags(&nbat->buffer_flags, nbat->natoms);
5623 if (nbl_list->bSimple)
5625 switch (nb_kernel_type)
5627 #ifdef GMX_NBNXN_SIMD_4XN
5628 case nbnxnk4xN_SIMD_4xN:
5629 nbs->icell_set_x = icell_set_x_simd_4xn;
5632 #ifdef GMX_NBNXN_SIMD_2XNN
5633 case nbnxnk4xN_SIMD_2xNN:
5634 nbs->icell_set_x = icell_set_x_simd_2xnn;
5638 nbs->icell_set_x = icell_set_x_simple;
5644 #ifdef NBNXN_SEARCH_BB_SIMD4
5645 nbs->icell_set_x = icell_set_x_supersub_simd4;
5647 nbs->icell_set_x = icell_set_x_supersub;
5653 /* Only zone (grid) 0 vs 0 */
5660 nzi = nbs->zones->nizone;
5663 if (!nbl_list->bSimple && min_ci_balanced > 0)
5665 nsubpair_max = get_nsubpair_max(nbs, iloc, rlist, min_ci_balanced);
5672 /* Clear all pair-lists */
5673 for (th = 0; th < nnbl; th++)
5675 clear_pairlist(nbl[th]);
5679 clear_pairlist_fep(nbl_list->nbl_fep[th]);
5683 for (zi = 0; zi < nzi; zi++)
5685 gridi = &nbs->grid[zi];
5687 if (NONLOCAL_I(iloc))
5689 zj0 = nbs->zones->izone[zi].j0;
5690 zj1 = nbs->zones->izone[zi].j1;
5696 for (zj = zj0; zj < zj1; zj++)
5698 gridj = &nbs->grid[zj];
5702 fprintf(debug, "ns search grid %d vs %d\n", zi, zj);
5705 nbs_cycle_start(&nbs->cc[enbsCCsearch]);
5707 if (nbl[0]->bSimple && !gridi->bSimple)
5709 /* Hybrid list, determine blocking later */
5714 ci_block = get_ci_block_size(gridi, nbs->DomDec, nnbl);
5717 /* With GPU: generate progressively smaller lists for
5718 * load balancing for local only or non-local with 2 zones.
5720 progBal = (LOCAL_I(iloc) || nbs->zones->n <= 2);
5722 #pragma omp parallel for num_threads(nnbl) schedule(static)
5723 for (th = 0; th < nnbl; th++)
5725 /* Re-init the thread-local work flag data before making
5726 * the first list (not an elegant conditional).
5728 if (nbat->bUseBufferFlags && ((zi == 0 && zj == 0) ||
5729 (bGPUCPU && zi == 0 && zj == 1)))
5731 init_buffer_flags(&nbs->work[th].buffer_flags, nbat->natoms);
5734 if (CombineNBLists && th > 0)
5736 clear_pairlist(nbl[th]);
5739 /* Divide the i super cell equally over the nblists */
5740 nbnxn_make_pairlist_part(nbs, gridi, gridj,
5741 &nbs->work[th], nbat, excl,
5745 nbat->bUseBufferFlags,
5747 progBal, min_ci_balanced,
5750 nbl_list->nbl_fep[th]);
5752 nbs_cycle_stop(&nbs->cc[enbsCCsearch]);
5757 for (th = 0; th < nnbl; th++)
5759 inc_nrnb(nrnb, eNR_NBNXN_DIST2, nbs->work[th].ndistc);
5761 if (nbl_list->bSimple)
5763 np_tot += nbl[th]->ncj;
5764 np_noq += nbl[th]->work->ncj_noq;
5765 np_hlj += nbl[th]->work->ncj_hlj;
5769 /* This count ignores potential subsequent pair pruning */
5770 np_tot += nbl[th]->nci_tot;
5773 nap = nbl[0]->na_ci*nbl[0]->na_cj;
5774 nbl_list->natpair_ljq = (np_tot - np_noq)*nap - np_hlj*nap/2;
5775 nbl_list->natpair_lj = np_noq*nap;
5776 nbl_list->natpair_q = np_hlj*nap/2;
5778 if (CombineNBLists && nnbl > 1)
5780 nbs_cycle_start(&nbs->cc[enbsCCcombine]);
5782 combine_nblists(nnbl-1, nbl+1, nbl[0]);
5784 nbs_cycle_stop(&nbs->cc[enbsCCcombine]);
5789 if (!nbl_list->bSimple)
5791 /* Sort the entries on size, large ones first */
5792 if (CombineNBLists || nnbl == 1)
5798 #pragma omp parallel for num_threads(nnbl) schedule(static)
5799 for (th = 0; th < nnbl; th++)
5806 if (nbat->bUseBufferFlags)
5808 reduce_buffer_flags(nbs, nnbl, &nbat->buffer_flags);
5813 /* Balance the free-energy lists over all the threads */
5814 balance_fep_lists(nbs, nbl_list);
5817 /* Special performance logging stuff (env.var. GMX_NBNXN_CYCLE) */
5820 nbs->search_count++;
5822 if (nbs->print_cycles &&
5823 (!nbs->DomDec || (nbs->DomDec && !LOCAL_I(iloc))) &&
5824 nbs->search_count % 100 == 0)
5826 nbs_cycle_print(stderr, nbs);
5829 if (debug && (CombineNBLists && nnbl > 1))
5831 if (nbl[0]->bSimple)
5833 print_nblist_statistics_simple(debug, nbl[0], nbs, rlist);
5837 print_nblist_statistics_supersub(debug, nbl[0], nbs, rlist);
5845 if (nbl[0]->bSimple)
5847 print_nblist_ci_cj(debug, nbl[0]);
5851 print_nblist_sci_cj(debug, nbl[0]);
5855 if (nbat->bUseBufferFlags)
5857 print_reduction_cost(&nbat->buffer_flags, nnbl);