2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015, 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.
38 #include "nbnxn_search.h"
44 #include "gromacs/legacyheaders/gmx_omp_nthreads.h"
45 #include "gromacs/legacyheaders/macros.h"
46 #include "gromacs/legacyheaders/nrnb.h"
47 #include "gromacs/legacyheaders/ns.h"
48 #include "gromacs/legacyheaders/types/commrec.h"
49 #include "gromacs/math/utilities.h"
50 #include "gromacs/math/vec.h"
51 #include "gromacs/mdlib/nb_verlet.h"
52 #include "gromacs/mdlib/nbnxn_atomdata.h"
53 #include "gromacs/mdlib/nbnxn_consts.h"
54 #include "gromacs/mdlib/nbnxn_internal.h"
55 #include "gromacs/pbcutil/ishift.h"
56 #include "gromacs/pbcutil/pbc.h"
57 #include "gromacs/simd/simd.h"
58 #include "gromacs/simd/vector_operations.h"
59 #include "gromacs/utility/smalloc.h"
61 #ifdef NBNXN_SEARCH_BB_SIMD4
62 /* Always use 4-wide SIMD for bounding box calculations */
65 /* Single precision BBs + coordinates, we can also load coordinates with SIMD */
66 # define NBNXN_SEARCH_SIMD4_FLOAT_X_BB
69 # if defined NBNXN_SEARCH_SIMD4_FLOAT_X_BB && (GPU_NSUBCELL == 4 || GPU_NSUBCELL == 8)
70 /* Store bounding boxes with x, y and z coordinates in packs of 4 */
71 # define NBNXN_PBB_SIMD4
74 /* The packed bounding box coordinate stride is always set to 4.
75 * With AVX we could use 8, but that turns out not to be faster.
78 # define STRIDE_PBB_2LOG 2
80 #endif /* NBNXN_SEARCH_BB_SIMD4 */
84 /* The functions below are macros as they are performance sensitive */
86 /* 4x4 list, pack=4: no complex conversion required */
87 /* i-cluster to j-cluster conversion */
88 #define CI_TO_CJ_J4(ci) (ci)
89 /* cluster index to coordinate array index conversion */
90 #define X_IND_CI_J4(ci) ((ci)*STRIDE_P4)
91 #define X_IND_CJ_J4(cj) ((cj)*STRIDE_P4)
93 /* 4x2 list, pack=4: j-cluster size is half the packing width */
94 /* i-cluster to j-cluster conversion */
95 #define CI_TO_CJ_J2(ci) ((ci)<<1)
96 /* cluster index to coordinate array index conversion */
97 #define X_IND_CI_J2(ci) ((ci)*STRIDE_P4)
98 #define X_IND_CJ_J2(cj) (((cj)>>1)*STRIDE_P4 + ((cj) & 1)*(PACK_X4>>1))
100 /* 4x8 list, pack=8: i-cluster size is half the packing width */
101 /* i-cluster to j-cluster conversion */
102 #define CI_TO_CJ_J8(ci) ((ci)>>1)
103 /* cluster index to coordinate array index conversion */
104 #define X_IND_CI_J8(ci) (((ci)>>1)*STRIDE_P8 + ((ci) & 1)*(PACK_X8>>1))
105 #define X_IND_CJ_J8(cj) ((cj)*STRIDE_P8)
107 /* The j-cluster size is matched to the SIMD width */
108 #if GMX_SIMD_REAL_WIDTH == 2
109 #define CI_TO_CJ_SIMD_4XN(ci) CI_TO_CJ_J2(ci)
110 #define X_IND_CI_SIMD_4XN(ci) X_IND_CI_J2(ci)
111 #define X_IND_CJ_SIMD_4XN(cj) X_IND_CJ_J2(cj)
113 #if GMX_SIMD_REAL_WIDTH == 4
114 #define CI_TO_CJ_SIMD_4XN(ci) CI_TO_CJ_J4(ci)
115 #define X_IND_CI_SIMD_4XN(ci) X_IND_CI_J4(ci)
116 #define X_IND_CJ_SIMD_4XN(cj) X_IND_CJ_J4(cj)
118 #if GMX_SIMD_REAL_WIDTH == 8
119 #define CI_TO_CJ_SIMD_4XN(ci) CI_TO_CJ_J8(ci)
120 #define X_IND_CI_SIMD_4XN(ci) X_IND_CI_J8(ci)
121 #define X_IND_CJ_SIMD_4XN(cj) X_IND_CJ_J8(cj)
122 /* Half SIMD with j-cluster size */
123 #define CI_TO_CJ_SIMD_2XNN(ci) CI_TO_CJ_J4(ci)
124 #define X_IND_CI_SIMD_2XNN(ci) X_IND_CI_J4(ci)
125 #define X_IND_CJ_SIMD_2XNN(cj) X_IND_CJ_J4(cj)
127 #if GMX_SIMD_REAL_WIDTH == 16
128 #define CI_TO_CJ_SIMD_2XNN(ci) CI_TO_CJ_J8(ci)
129 #define X_IND_CI_SIMD_2XNN(ci) X_IND_CI_J8(ci)
130 #define X_IND_CJ_SIMD_2XNN(cj) X_IND_CJ_J8(cj)
132 #error "unsupported GMX_SIMD_REAL_WIDTH"
138 #endif /* GMX_NBNXN_SIMD */
141 #ifdef NBNXN_SEARCH_BB_SIMD4
142 /* Store bounding boxes corners as quadruplets: xxxxyyyyzzzz */
144 /* Size of bounding box corners quadruplet */
145 #define NNBSBB_XXXX (NNBSBB_D*DIM*STRIDE_PBB)
148 /* We shift the i-particles backward for PBC.
149 * This leads to more conditionals than shifting forward.
150 * We do this to get more balanced pair lists.
152 #define NBNXN_SHIFT_BACKWARD
155 /* This define is a lazy way to avoid interdependence of the grid
156 * and searching data structures.
158 #define NBNXN_NA_SC_MAX (GPU_NSUBCELL*NBNXN_GPU_CLUSTER_SIZE)
161 static void nbs_cycle_clear(nbnxn_cycle_t *cc)
165 for (i = 0; i < enbsCCnr; i++)
172 static double Mcyc_av(const nbnxn_cycle_t *cc)
174 return (double)cc->c*1e-6/cc->count;
177 static void nbs_cycle_print(FILE *fp, const nbnxn_search_t nbs)
183 fprintf(fp, "ns %4d grid %4.1f search %4.1f red.f %5.3f",
184 nbs->cc[enbsCCgrid].count,
185 Mcyc_av(&nbs->cc[enbsCCgrid]),
186 Mcyc_av(&nbs->cc[enbsCCsearch]),
187 Mcyc_av(&nbs->cc[enbsCCreducef]));
189 if (nbs->nthread_max > 1)
191 if (nbs->cc[enbsCCcombine].count > 0)
193 fprintf(fp, " comb %5.2f",
194 Mcyc_av(&nbs->cc[enbsCCcombine]));
196 fprintf(fp, " s. th");
197 for (t = 0; t < nbs->nthread_max; t++)
199 fprintf(fp, " %4.1f",
200 Mcyc_av(&nbs->work[t].cc[enbsCCsearch]));
206 static void nbnxn_grid_init(nbnxn_grid_t * grid)
209 grid->cxy_ind = NULL;
210 grid->cxy_nalloc = 0;
216 static int get_2log(int n)
221 while ((1<<log2) < n)
227 gmx_fatal(FARGS, "nbnxn na_c (%d) is not a power of 2", n);
233 int nbnxn_kernel_to_ci_size(int nb_kernel_type)
235 switch (nb_kernel_type)
237 case nbnxnk4x4_PlainC:
238 case nbnxnk4xN_SIMD_4xN:
239 case nbnxnk4xN_SIMD_2xNN:
240 return NBNXN_CPU_CLUSTER_I_SIZE;
241 case nbnxnk8x8x8_GPU:
242 case nbnxnk8x8x8_PlainC:
243 /* The cluster size for super/sub lists is only set here.
244 * Any value should work for the pair-search and atomdata code.
245 * The kernels, of course, might require a particular value.
247 return NBNXN_GPU_CLUSTER_SIZE;
249 gmx_incons("unknown kernel type");
255 int nbnxn_kernel_to_cj_size(int nb_kernel_type)
257 int nbnxn_simd_width = 0;
260 #ifdef GMX_NBNXN_SIMD
261 nbnxn_simd_width = GMX_SIMD_REAL_WIDTH;
264 switch (nb_kernel_type)
266 case nbnxnk4x4_PlainC:
267 cj_size = NBNXN_CPU_CLUSTER_I_SIZE;
269 case nbnxnk4xN_SIMD_4xN:
270 cj_size = nbnxn_simd_width;
272 case nbnxnk4xN_SIMD_2xNN:
273 cj_size = nbnxn_simd_width/2;
275 case nbnxnk8x8x8_GPU:
276 case nbnxnk8x8x8_PlainC:
277 cj_size = nbnxn_kernel_to_ci_size(nb_kernel_type);
280 gmx_incons("unknown kernel type");
286 static int ci_to_cj(int na_cj_2log, int ci)
290 case 2: return ci; break;
291 case 1: return (ci<<1); break;
292 case 3: return (ci>>1); break;
298 gmx_bool nbnxn_kernel_pairlist_simple(int nb_kernel_type)
300 if (nb_kernel_type == nbnxnkNotSet)
302 gmx_fatal(FARGS, "Non-bonded kernel type not set for Verlet-style pair-list.");
305 switch (nb_kernel_type)
307 case nbnxnk8x8x8_GPU:
308 case nbnxnk8x8x8_PlainC:
311 case nbnxnk4x4_PlainC:
312 case nbnxnk4xN_SIMD_4xN:
313 case nbnxnk4xN_SIMD_2xNN:
317 gmx_incons("Invalid nonbonded kernel type passed!");
322 /* Initializes a single nbnxn_pairlist_t data structure */
323 static void nbnxn_init_pairlist_fep(t_nblist *nl)
325 nl->type = GMX_NBLIST_INTERACTION_FREE_ENERGY;
326 nl->igeometry = GMX_NBLIST_GEOMETRY_PARTICLE_PARTICLE;
327 /* The interaction functions are set in the free energy kernel fuction */
346 void nbnxn_init_search(nbnxn_search_t * nbs_ptr,
348 gmx_domdec_zones_t *zones,
360 nbs->DomDec = (n_dd_cells != NULL);
362 clear_ivec(nbs->dd_dim);
368 for (d = 0; d < DIM; d++)
370 if ((*n_dd_cells)[d] > 1)
373 /* Each grid matches a DD zone */
379 snew(nbs->grid, nbs->ngrid);
380 for (g = 0; g < nbs->ngrid; g++)
382 nbnxn_grid_init(&nbs->grid[g]);
385 nbs->cell_nalloc = 0;
389 nbs->nthread_max = nthread_max;
391 /* Initialize the work data structures for each thread */
392 snew(nbs->work, nbs->nthread_max);
393 for (t = 0; t < nbs->nthread_max; t++)
395 nbs->work[t].cxy_na = NULL;
396 nbs->work[t].cxy_na_nalloc = 0;
397 nbs->work[t].sort_work = NULL;
398 nbs->work[t].sort_work_nalloc = 0;
400 snew(nbs->work[t].nbl_fep, 1);
401 nbnxn_init_pairlist_fep(nbs->work[t].nbl_fep);
404 /* Initialize detailed nbsearch cycle counting */
405 nbs->print_cycles = (getenv("GMX_NBNXN_CYCLE") != 0);
406 nbs->search_count = 0;
407 nbs_cycle_clear(nbs->cc);
408 for (t = 0; t < nbs->nthread_max; t++)
410 nbs_cycle_clear(nbs->work[t].cc);
414 static real grid_atom_density(int n, rvec corner0, rvec corner1)
420 /* To avoid zero density we use a minimum of 1 atom */
424 rvec_sub(corner1, corner0, size);
426 return n/(size[XX]*size[YY]*size[ZZ]);
429 static int set_grid_size_xy(const nbnxn_search_t nbs,
432 int n, rvec corner0, rvec corner1,
437 real adens, tlen, tlen_x, tlen_y, nc_max;
440 rvec_sub(corner1, corner0, size);
444 assert(atom_density > 0);
446 /* target cell length */
449 /* To minimize the zero interactions, we should make
450 * the largest of the i/j cell cubic.
452 na_c = max(grid->na_c, grid->na_cj);
454 /* Approximately cubic cells */
455 tlen = pow(na_c/atom_density, 1.0/3.0);
461 /* Approximately cubic sub cells */
462 tlen = pow(grid->na_c/atom_density, 1.0/3.0);
463 tlen_x = tlen*GPU_NSUBCELL_X;
464 tlen_y = tlen*GPU_NSUBCELL_Y;
466 /* We round ncx and ncy down, because we get less cell pairs
467 * in the nbsist when the fixed cell dimensions (x,y) are
468 * larger than the variable one (z) than the other way around.
470 grid->ncx = max(1, (int)(size[XX]/tlen_x));
471 grid->ncy = max(1, (int)(size[YY]/tlen_y));
479 grid->sx = size[XX]/grid->ncx;
480 grid->sy = size[YY]/grid->ncy;
481 grid->inv_sx = 1/grid->sx;
482 grid->inv_sy = 1/grid->sy;
486 /* This is a non-home zone, add an extra row of cells
487 * for particles communicated for bonded interactions.
488 * These can be beyond the cut-off. It doesn't matter where
489 * they end up on the grid, but for performance it's better
490 * if they don't end up in cells that can be within cut-off range.
496 /* We need one additional cell entry for particles moved by DD */
497 if (grid->ncx*grid->ncy+1 > grid->cxy_nalloc)
499 grid->cxy_nalloc = over_alloc_large(grid->ncx*grid->ncy+1);
500 srenew(grid->cxy_na, grid->cxy_nalloc);
501 srenew(grid->cxy_ind, grid->cxy_nalloc+1);
503 for (t = 0; t < nbs->nthread_max; t++)
505 if (grid->ncx*grid->ncy+1 > nbs->work[t].cxy_na_nalloc)
507 nbs->work[t].cxy_na_nalloc = over_alloc_large(grid->ncx*grid->ncy+1);
508 srenew(nbs->work[t].cxy_na, nbs->work[t].cxy_na_nalloc);
512 /* Worst case scenario of 1 atom in each last cell */
513 if (grid->na_cj <= grid->na_c)
515 nc_max = n/grid->na_sc + grid->ncx*grid->ncy;
519 nc_max = n/grid->na_sc + grid->ncx*grid->ncy*grid->na_cj/grid->na_c;
522 if (nc_max > grid->nc_nalloc)
524 grid->nc_nalloc = over_alloc_large(nc_max);
525 srenew(grid->nsubc, grid->nc_nalloc);
526 srenew(grid->bbcz, grid->nc_nalloc*NNBSBB_D);
528 sfree_aligned(grid->bb);
529 /* This snew also zeros the contents, this avoid possible
530 * floating exceptions in SIMD with the unused bb elements.
534 snew_aligned(grid->bb, grid->nc_nalloc, 16);
541 pbb_nalloc = grid->nc_nalloc*GPU_NSUBCELL/STRIDE_PBB*NNBSBB_XXXX;
542 snew_aligned(grid->pbb, pbb_nalloc, 16);
544 snew_aligned(grid->bb, grid->nc_nalloc*GPU_NSUBCELL, 16);
550 if (grid->na_cj == grid->na_c)
552 grid->bbj = grid->bb;
556 sfree_aligned(grid->bbj);
557 snew_aligned(grid->bbj, grid->nc_nalloc*grid->na_c/grid->na_cj, 16);
561 srenew(grid->flags, grid->nc_nalloc);
564 srenew(grid->fep, grid->nc_nalloc*grid->na_sc/grid->na_c);
568 copy_rvec(corner0, grid->c0);
569 copy_rvec(corner1, grid->c1);
570 copy_rvec(size, grid->size);
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_grid_t *grid)
1013 for (c = 0; c < grid->nc; c++)
1015 for (d = 0; d < DIM; d++)
1017 ba[d] += grid->bb[c].upper[d] - grid->bb[c].lower[d];
1020 dsvmul(1.0/grid->nc, ba, ba);
1022 fprintf(fp, "ns bb: grid %4.2f %4.2f %4.2f abs %4.2f %4.2f %4.2f rel %4.2f %4.2f %4.2f\n",
1025 grid->atom_density > 0 ?
1026 grid->na_c/(grid->atom_density*grid->sx*grid->sy) : 0.0,
1027 ba[XX], ba[YY], ba[ZZ],
1030 grid->atom_density > 0 ?
1031 ba[ZZ]/(grid->na_c/(grid->atom_density*grid->sx*grid->sy)) : 0.0);
1034 /* Prints the average bb size, used for debug output */
1035 static void print_bbsizes_supersub(FILE *fp,
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: grid %4.2f %4.2f %4.2f abs %4.2f %4.2f %4.2f rel %4.2f %4.2f %4.2f\n",
1078 grid->sx/GPU_NSUBCELL_X,
1079 grid->sy/GPU_NSUBCELL_Y,
1080 grid->atom_density > 0 ?
1081 grid->na_sc/(grid->atom_density*grid->sx*grid->sy*GPU_NSUBCELL_Z) : 0.0,
1082 ba[XX], ba[YY], ba[ZZ],
1083 ba[XX]*GPU_NSUBCELL_X/grid->sx,
1084 ba[YY]*GPU_NSUBCELL_Y/grid->sy,
1085 grid->atom_density > 0 ?
1086 ba[ZZ]/(grid->na_sc/(grid->atom_density*grid->sx*grid->sy*GPU_NSUBCELL_Z)) : 0.0);
1089 /* Potentially sorts atoms on LJ coefficients !=0 and ==0.
1090 * Also sets interaction flags.
1092 void sort_on_lj(int na_c,
1093 int a0, int a1, const int *atinfo,
1097 int subc, s, a, n1, n2, a_lj_max, i, j;
1098 int sort1[NBNXN_NA_SC_MAX/GPU_NSUBCELL];
1099 int sort2[NBNXN_NA_SC_MAX/GPU_NSUBCELL];
1100 gmx_bool haveQ, bFEP;
1105 for (s = a0; s < a1; s += na_c)
1107 /* Make lists for this (sub-)cell on atoms with and without LJ */
1112 for (a = s; a < min(s+na_c, a1); a++)
1114 haveQ = haveQ || GET_CGINFO_HAS_Q(atinfo[order[a]]);
1116 if (GET_CGINFO_HAS_VDW(atinfo[order[a]]))
1118 sort1[n1++] = order[a];
1123 sort2[n2++] = order[a];
1127 /* If we don't have atoms with LJ, there's nothing to sort */
1130 *flags |= NBNXN_CI_DO_LJ(subc);
1134 /* Only sort when strictly necessary. Ordering particles
1135 * Ordering particles can lead to less accurate summation
1136 * due to rounding, both for LJ and Coulomb interactions.
1138 if (2*(a_lj_max - s) >= na_c)
1140 for (i = 0; i < n1; i++)
1142 order[a0+i] = sort1[i];
1144 for (j = 0; j < n2; j++)
1146 order[a0+n1+j] = sort2[j];
1150 *flags |= NBNXN_CI_HALF_LJ(subc);
1155 *flags |= NBNXN_CI_DO_COUL(subc);
1161 /* Fill a pair search cell with atoms.
1162 * Potentially sorts atoms and sets the interaction flags.
1164 void fill_cell(const nbnxn_search_t nbs,
1166 nbnxn_atomdata_t *nbat,
1170 int sx, int sy, int sz,
1171 nbnxn_bb_t gmx_unused *bb_work_aligned)
1184 sort_on_lj(grid->na_c, a0, a1, atinfo, nbs->a,
1185 grid->flags+(a0>>grid->na_c_2log)-grid->cell0);
1190 /* Set the fep flag for perturbed atoms in this (sub-)cell */
1193 /* The grid-local cluster/(sub-)cell index */
1194 c = (a0 >> grid->na_c_2log) - grid->cell0*(grid->bSimple ? 1 : GPU_NSUBCELL);
1196 for (at = a0; at < a1; at++)
1198 if (nbs->a[at] >= 0 && GET_CGINFO_FEP(atinfo[nbs->a[at]]))
1200 grid->fep[c] |= (1 << (at - a0));
1205 /* Now we have sorted the atoms, set the cell indices */
1206 for (a = a0; a < a1; a++)
1208 nbs->cell[nbs->a[a]] = a;
1211 copy_rvec_to_nbat_real(nbs->a+a0, a1-a0, grid->na_c, x,
1212 nbat->XFormat, nbat->x, a0,
1215 if (nbat->XFormat == nbatX4)
1217 /* Store the bounding boxes as xyz.xyz. */
1218 offset = (a0 - grid->cell0*grid->na_sc) >> grid->na_c_2log;
1219 bb_ptr = grid->bb + offset;
1221 #if defined GMX_NBNXN_SIMD && GMX_SIMD_REAL_WIDTH == 2
1222 if (2*grid->na_cj == grid->na_c)
1224 calc_bounding_box_x_x4_halves(na, nbat->x+X4_IND_A(a0), bb_ptr,
1225 grid->bbj+offset*2);
1230 calc_bounding_box_x_x4(na, nbat->x+X4_IND_A(a0), bb_ptr);
1233 else if (nbat->XFormat == nbatX8)
1235 /* Store the bounding boxes as xyz.xyz. */
1236 offset = (a0 - grid->cell0*grid->na_sc) >> grid->na_c_2log;
1237 bb_ptr = grid->bb + offset;
1239 calc_bounding_box_x_x8(na, nbat->x+X8_IND_A(a0), bb_ptr);
1242 else if (!grid->bSimple)
1244 /* Store the bounding boxes in a format convenient
1245 * for SIMD4 calculations: xxxxyyyyzzzz...
1249 ((a0-grid->cell0*grid->na_sc)>>(grid->na_c_2log+STRIDE_PBB_2LOG))*NNBSBB_XXXX +
1250 (((a0-grid->cell0*grid->na_sc)>>grid->na_c_2log) & (STRIDE_PBB-1));
1252 #ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
1253 if (nbat->XFormat == nbatXYZQ)
1255 calc_bounding_box_xxxx_simd4(na, nbat->x+a0*nbat->xstride,
1256 bb_work_aligned, pbb_ptr);
1261 calc_bounding_box_xxxx(na, nbat->xstride, nbat->x+a0*nbat->xstride,
1266 fprintf(debug, "%2d %2d %2d bb %5.2f %5.2f %5.2f %5.2f %5.2f %5.2f\n",
1268 pbb_ptr[0*STRIDE_PBB], pbb_ptr[3*STRIDE_PBB],
1269 pbb_ptr[1*STRIDE_PBB], pbb_ptr[4*STRIDE_PBB],
1270 pbb_ptr[2*STRIDE_PBB], pbb_ptr[5*STRIDE_PBB]);
1276 /* Store the bounding boxes as xyz.xyz. */
1277 bb_ptr = grid->bb+((a0-grid->cell0*grid->na_sc)>>grid->na_c_2log);
1279 calc_bounding_box(na, nbat->xstride, nbat->x+a0*nbat->xstride,
1285 bbo = (a0 - grid->cell0*grid->na_sc)/grid->na_c;
1286 fprintf(debug, "%2d %2d %2d bb %5.2f %5.2f %5.2f %5.2f %5.2f %5.2f\n",
1288 grid->bb[bbo].lower[BB_X],
1289 grid->bb[bbo].lower[BB_Y],
1290 grid->bb[bbo].lower[BB_Z],
1291 grid->bb[bbo].upper[BB_X],
1292 grid->bb[bbo].upper[BB_Y],
1293 grid->bb[bbo].upper[BB_Z]);
1298 /* Spatially sort the atoms within one grid column */
1299 static void sort_columns_simple(const nbnxn_search_t nbs,
1305 nbnxn_atomdata_t *nbat,
1306 int cxy_start, int cxy_end,
1310 int cx, cy, cz, ncz, cfilled, c;
1311 int na, ash, ind, a;
1316 fprintf(debug, "cell0 %d sorting columns %d - %d, atoms %d - %d\n",
1317 grid->cell0, cxy_start, cxy_end, a0, a1);
1320 /* Sort the atoms within each x,y column in 3 dimensions */
1321 for (cxy = cxy_start; cxy < cxy_end; cxy++)
1324 cy = cxy - cx*grid->ncy;
1326 na = grid->cxy_na[cxy];
1327 ncz = grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy];
1328 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1330 /* Sort the atoms within each x,y column on z coordinate */
1331 sort_atoms(ZZ, FALSE, dd_zone,
1334 1.0/grid->size[ZZ], ncz*grid->na_sc,
1337 /* Fill the ncz cells in this column */
1338 cfilled = grid->cxy_ind[cxy];
1339 for (cz = 0; cz < ncz; cz++)
1341 c = grid->cxy_ind[cxy] + cz;
1343 ash_c = ash + cz*grid->na_sc;
1344 na_c = min(grid->na_sc, na-(ash_c-ash));
1346 fill_cell(nbs, grid, nbat,
1347 ash_c, ash_c+na_c, atinfo, x,
1348 grid->na_sc*cx + (dd_zone >> 2),
1349 grid->na_sc*cy + (dd_zone & 3),
1353 /* This copy to bbcz is not really necessary.
1354 * But it allows to use the same grid search code
1355 * for the simple and supersub cell setups.
1361 grid->bbcz[c*NNBSBB_D ] = grid->bb[cfilled].lower[BB_Z];
1362 grid->bbcz[c*NNBSBB_D+1] = grid->bb[cfilled].upper[BB_Z];
1365 /* Set the unused atom indices to -1 */
1366 for (ind = na; ind < ncz*grid->na_sc; ind++)
1368 nbs->a[ash+ind] = -1;
1373 /* Spatially sort the atoms within one grid column */
1374 static void sort_columns_supersub(const nbnxn_search_t nbs,
1380 nbnxn_atomdata_t *nbat,
1381 int cxy_start, int cxy_end,
1385 int cx, cy, cz = -1, c = -1, ncz;
1386 int na, ash, na_c, ind, a;
1387 int subdiv_z, sub_z, na_z, ash_z;
1388 int subdiv_y, sub_y, na_y, ash_y;
1389 int subdiv_x, sub_x, na_x, ash_x;
1391 nbnxn_bb_t bb_work_array[2], *bb_work_aligned;
1393 bb_work_aligned = (nbnxn_bb_t *)(((size_t)(bb_work_array+1)) & (~((size_t)15)));
1397 fprintf(debug, "cell0 %d sorting columns %d - %d, atoms %d - %d\n",
1398 grid->cell0, cxy_start, cxy_end, a0, a1);
1401 subdiv_x = grid->na_c;
1402 subdiv_y = GPU_NSUBCELL_X*subdiv_x;
1403 subdiv_z = GPU_NSUBCELL_Y*subdiv_y;
1405 /* Sort the atoms within each x,y column in 3 dimensions */
1406 for (cxy = cxy_start; cxy < cxy_end; cxy++)
1409 cy = cxy - cx*grid->ncy;
1411 na = grid->cxy_na[cxy];
1412 ncz = grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy];
1413 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1415 /* Sort the atoms within each x,y column on z coordinate */
1416 sort_atoms(ZZ, FALSE, dd_zone,
1419 1.0/grid->size[ZZ], ncz*grid->na_sc,
1422 /* This loop goes over the supercells and subcells along z at once */
1423 for (sub_z = 0; sub_z < ncz*GPU_NSUBCELL_Z; sub_z++)
1425 ash_z = ash + sub_z*subdiv_z;
1426 na_z = min(subdiv_z, na-(ash_z-ash));
1428 /* We have already sorted on z */
1430 if (sub_z % GPU_NSUBCELL_Z == 0)
1432 cz = sub_z/GPU_NSUBCELL_Z;
1433 c = grid->cxy_ind[cxy] + cz;
1435 /* The number of atoms in this supercell */
1436 na_c = min(grid->na_sc, na-(ash_z-ash));
1438 grid->nsubc[c] = min(GPU_NSUBCELL, (na_c+grid->na_c-1)/grid->na_c);
1440 /* Store the z-boundaries of the super cell */
1441 grid->bbcz[c*NNBSBB_D ] = x[nbs->a[ash_z]][ZZ];
1442 grid->bbcz[c*NNBSBB_D+1] = x[nbs->a[ash_z+na_c-1]][ZZ];
1445 #if GPU_NSUBCELL_Y > 1
1446 /* Sort the atoms along y */
1447 sort_atoms(YY, (sub_z & 1), dd_zone,
1448 nbs->a+ash_z, na_z, x,
1449 grid->c0[YY]+cy*grid->sy,
1450 grid->inv_sy, subdiv_z,
1454 for (sub_y = 0; sub_y < GPU_NSUBCELL_Y; sub_y++)
1456 ash_y = ash_z + sub_y*subdiv_y;
1457 na_y = min(subdiv_y, na-(ash_y-ash));
1459 #if GPU_NSUBCELL_X > 1
1460 /* Sort the atoms along x */
1461 sort_atoms(XX, ((cz*GPU_NSUBCELL_Y + sub_y) & 1), dd_zone,
1462 nbs->a+ash_y, na_y, x,
1463 grid->c0[XX]+cx*grid->sx,
1464 grid->inv_sx, subdiv_y,
1468 for (sub_x = 0; sub_x < GPU_NSUBCELL_X; sub_x++)
1470 ash_x = ash_y + sub_x*subdiv_x;
1471 na_x = min(subdiv_x, na-(ash_x-ash));
1473 fill_cell(nbs, grid, nbat,
1474 ash_x, ash_x+na_x, atinfo, x,
1475 grid->na_c*(cx*GPU_NSUBCELL_X+sub_x) + (dd_zone >> 2),
1476 grid->na_c*(cy*GPU_NSUBCELL_Y+sub_y) + (dd_zone & 3),
1483 /* Set the unused atom indices to -1 */
1484 for (ind = na; ind < ncz*grid->na_sc; ind++)
1486 nbs->a[ash+ind] = -1;
1491 /* Determine in which grid column atoms should go */
1492 static void calc_column_indices(nbnxn_grid_t *grid,
1495 int dd_zone, const int *move,
1496 int thread, int nthread,
1503 /* We add one extra cell for particles which moved during DD */
1504 for (i = 0; i < grid->ncx*grid->ncy+1; i++)
1509 n0 = a0 + (int)((thread+0)*(a1 - a0))/nthread;
1510 n1 = a0 + (int)((thread+1)*(a1 - a0))/nthread;
1514 for (i = n0; i < n1; i++)
1516 if (move == NULL || move[i] >= 0)
1518 /* We need to be careful with rounding,
1519 * particles might be a few bits outside the local zone.
1520 * The int cast takes care of the lower bound,
1521 * we will explicitly take care of the upper bound.
1523 cx = (int)((x[i][XX] - grid->c0[XX])*grid->inv_sx);
1524 cy = (int)((x[i][YY] - grid->c0[YY])*grid->inv_sy);
1527 if (cx < 0 || cx > grid->ncx ||
1528 cy < 0 || cy > grid->ncy)
1531 "grid cell cx %d cy %d out of range (max %d %d)\n"
1532 "atom %f %f %f, grid->c0 %f %f",
1533 cx, cy, grid->ncx, grid->ncy,
1534 x[i][XX], x[i][YY], x[i][ZZ], grid->c0[XX], grid->c0[YY]);
1537 /* Take care of potential rouding issues */
1538 cx = min(cx, grid->ncx - 1);
1539 cy = min(cy, grid->ncy - 1);
1541 /* For the moment cell will contain only the, grid local,
1542 * x and y indices, not z.
1544 cell[i] = cx*grid->ncy + cy;
1548 /* Put this moved particle after the end of the grid,
1549 * so we can process it later without using conditionals.
1551 cell[i] = grid->ncx*grid->ncy;
1560 for (i = n0; i < n1; i++)
1562 cx = (int)((x[i][XX] - grid->c0[XX])*grid->inv_sx);
1563 cy = (int)((x[i][YY] - grid->c0[YY])*grid->inv_sy);
1565 /* For non-home zones there could be particles outside
1566 * the non-bonded cut-off range, which have been communicated
1567 * for bonded interactions only. For the result it doesn't
1568 * matter where these end up on the grid. For performance
1569 * we put them in an extra row at the border.
1572 cx = min(cx, grid->ncx - 1);
1574 cy = min(cy, grid->ncy - 1);
1576 /* For the moment cell will contain only the, grid local,
1577 * x and y indices, not z.
1579 cell[i] = cx*grid->ncy + cy;
1586 /* Determine in which grid cells the atoms should go */
1587 static void calc_cell_indices(const nbnxn_search_t nbs,
1594 nbnxn_atomdata_t *nbat)
1597 int cx, cy, cxy, ncz_max, ncz;
1598 int nthread, thread;
1599 int *cxy_na, cxy_na_i;
1601 nthread = gmx_omp_nthreads_get(emntPairsearch);
1603 #pragma omp parallel for num_threads(nthread) schedule(static)
1604 for (thread = 0; thread < nthread; thread++)
1606 calc_column_indices(grid, a0, a1, x, dd_zone, move, thread, nthread,
1607 nbs->cell, nbs->work[thread].cxy_na);
1610 /* Make the cell index as a function of x and y */
1613 grid->cxy_ind[0] = 0;
1614 for (i = 0; i < grid->ncx*grid->ncy+1; i++)
1616 /* We set ncz_max at the beginning of the loop iso at the end
1617 * to skip i=grid->ncx*grid->ncy which are moved particles
1618 * that do not need to be ordered on the grid.
1624 cxy_na_i = nbs->work[0].cxy_na[i];
1625 for (thread = 1; thread < nthread; thread++)
1627 cxy_na_i += nbs->work[thread].cxy_na[i];
1629 ncz = (cxy_na_i + grid->na_sc - 1)/grid->na_sc;
1630 if (nbat->XFormat == nbatX8)
1632 /* Make the number of cell a multiple of 2 */
1633 ncz = (ncz + 1) & ~1;
1635 grid->cxy_ind[i+1] = grid->cxy_ind[i] + ncz;
1636 /* Clear cxy_na, so we can reuse the array below */
1637 grid->cxy_na[i] = 0;
1639 grid->nc = grid->cxy_ind[grid->ncx*grid->ncy] - grid->cxy_ind[0];
1641 nbat->natoms = (grid->cell0 + grid->nc)*grid->na_sc;
1645 fprintf(debug, "ns na_sc %d na_c %d super-cells: %d x %d y %d z %.1f maxz %d\n",
1646 grid->na_sc, grid->na_c, grid->nc,
1647 grid->ncx, grid->ncy, grid->nc/((double)(grid->ncx*grid->ncy)),
1652 for (cy = 0; cy < grid->ncy; cy++)
1654 for (cx = 0; cx < grid->ncx; cx++)
1656 fprintf(debug, " %2d", grid->cxy_ind[i+1]-grid->cxy_ind[i]);
1659 fprintf(debug, "\n");
1664 /* Make sure the work array for sorting is large enough */
1665 if (ncz_max*grid->na_sc*SGSF > nbs->work[0].sort_work_nalloc)
1667 for (thread = 0; thread < nbs->nthread_max; thread++)
1669 nbs->work[thread].sort_work_nalloc =
1670 over_alloc_large(ncz_max*grid->na_sc*SGSF);
1671 srenew(nbs->work[thread].sort_work,
1672 nbs->work[thread].sort_work_nalloc);
1673 /* When not in use, all elements should be -1 */
1674 for (i = 0; i < nbs->work[thread].sort_work_nalloc; i++)
1676 nbs->work[thread].sort_work[i] = -1;
1681 /* Now we know the dimensions we can fill the grid.
1682 * This is the first, unsorted fill. We sort the columns after this.
1684 for (i = a0; i < a1; i++)
1686 /* At this point nbs->cell contains the local grid x,y indices */
1688 nbs->a[(grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc + grid->cxy_na[cxy]++] = i;
1693 /* Set the cell indices for the moved particles */
1694 n0 = grid->nc*grid->na_sc;
1695 n1 = grid->nc*grid->na_sc+grid->cxy_na[grid->ncx*grid->ncy];
1698 for (i = n0; i < n1; i++)
1700 nbs->cell[nbs->a[i]] = i;
1705 /* Sort the super-cell columns along z into the sub-cells. */
1706 #pragma omp parallel for num_threads(nthread) schedule(static)
1707 for (thread = 0; thread < nthread; thread++)
1711 sort_columns_simple(nbs, dd_zone, grid, a0, a1, atinfo, x, nbat,
1712 ((thread+0)*grid->ncx*grid->ncy)/nthread,
1713 ((thread+1)*grid->ncx*grid->ncy)/nthread,
1714 nbs->work[thread].sort_work);
1718 sort_columns_supersub(nbs, dd_zone, grid, a0, a1, atinfo, x, nbat,
1719 ((thread+0)*grid->ncx*grid->ncy)/nthread,
1720 ((thread+1)*grid->ncx*grid->ncy)/nthread,
1721 nbs->work[thread].sort_work);
1725 if (grid->bSimple && nbat->XFormat == nbatX8)
1727 combine_bounding_box_pairs(grid, grid->bb);
1732 grid->nsubc_tot = 0;
1733 for (i = 0; i < grid->nc; i++)
1735 grid->nsubc_tot += grid->nsubc[i];
1743 print_bbsizes_simple(debug, grid);
1747 fprintf(debug, "ns non-zero sub-cells: %d average atoms %.2f\n",
1748 grid->nsubc_tot, (a1-a0)/(double)grid->nsubc_tot);
1750 print_bbsizes_supersub(debug, grid);
1755 static void init_buffer_flags(nbnxn_buffer_flags_t *flags,
1760 flags->nflag = (natoms + NBNXN_BUFFERFLAG_SIZE - 1)/NBNXN_BUFFERFLAG_SIZE;
1761 if (flags->nflag > flags->flag_nalloc)
1763 flags->flag_nalloc = over_alloc_large(flags->nflag);
1764 srenew(flags->flag, flags->flag_nalloc);
1766 for (b = 0; b < flags->nflag; b++)
1768 bitmask_clear(&(flags->flag[b]));
1772 /* Sets up a grid and puts the atoms on the grid.
1773 * This function only operates on one domain of the domain decompostion.
1774 * Note that without domain decomposition there is only one domain.
1776 void nbnxn_put_on_grid(nbnxn_search_t nbs,
1777 int ePBC, matrix box,
1779 rvec corner0, rvec corner1,
1784 int nmoved, int *move,
1786 nbnxn_atomdata_t *nbat)
1790 int nc_max_grid, nc_max;
1792 grid = &nbs->grid[dd_zone];
1794 nbs_cycle_start(&nbs->cc[enbsCCgrid]);
1796 grid->bSimple = nbnxn_kernel_pairlist_simple(nb_kernel_type);
1798 grid->na_c = nbnxn_kernel_to_ci_size(nb_kernel_type);
1799 grid->na_cj = nbnxn_kernel_to_cj_size(nb_kernel_type);
1800 grid->na_sc = (grid->bSimple ? 1 : GPU_NSUBCELL)*grid->na_c;
1801 grid->na_c_2log = get_2log(grid->na_c);
1803 nbat->na_c = grid->na_c;
1812 (nbs->grid[dd_zone-1].cell0 + nbs->grid[dd_zone-1].nc)*
1813 nbs->grid[dd_zone-1].na_sc/grid->na_sc;
1821 copy_mat(box, nbs->box);
1823 /* Avoid zero density */
1824 if (atom_density > 0)
1826 grid->atom_density = atom_density;
1830 grid->atom_density = grid_atom_density(n-nmoved, corner0, corner1);
1835 nbs->natoms_local = a1 - nmoved;
1836 /* We assume that nbnxn_put_on_grid is called first
1837 * for the local atoms (dd_zone=0).
1839 nbs->natoms_nonlocal = a1 - nmoved;
1843 fprintf(debug, "natoms_local = %5d atom_density = %5.1f\n",
1844 nbs->natoms_local, grid->atom_density);
1849 nbs->natoms_nonlocal = max(nbs->natoms_nonlocal, a1);
1852 /* We always use the home zone (grid[0]) for setting the cell size,
1853 * since determining densities for non-local zones is difficult.
1855 nc_max_grid = set_grid_size_xy(nbs, grid,
1856 dd_zone, n-nmoved, corner0, corner1,
1857 nbs->grid[0].atom_density);
1859 nc_max = grid->cell0 + nc_max_grid;
1861 if (a1 > nbs->cell_nalloc)
1863 nbs->cell_nalloc = over_alloc_large(a1);
1864 srenew(nbs->cell, nbs->cell_nalloc);
1867 /* To avoid conditionals we store the moved particles at the end of a,
1868 * make sure we have enough space.
1870 if (nc_max*grid->na_sc + nmoved > nbs->a_nalloc)
1872 nbs->a_nalloc = over_alloc_large(nc_max*grid->na_sc + nmoved);
1873 srenew(nbs->a, nbs->a_nalloc);
1876 /* We need padding up to a multiple of the buffer flag size: simply add */
1877 if (nc_max*grid->na_sc + NBNXN_BUFFERFLAG_SIZE > nbat->nalloc)
1879 nbnxn_atomdata_realloc(nbat, nc_max*grid->na_sc+NBNXN_BUFFERFLAG_SIZE);
1882 calc_cell_indices(nbs, dd_zone, grid, a0, a1, atinfo, x, move, nbat);
1886 nbat->natoms_local = nbat->natoms;
1889 nbs_cycle_stop(&nbs->cc[enbsCCgrid]);
1892 /* Calls nbnxn_put_on_grid for all non-local domains */
1893 void nbnxn_put_on_grid_nonlocal(nbnxn_search_t nbs,
1894 const gmx_domdec_zones_t *zones,
1898 nbnxn_atomdata_t *nbat)
1903 for (zone = 1; zone < zones->n; zone++)
1905 for (d = 0; d < DIM; d++)
1907 c0[d] = zones->size[zone].bb_x0[d];
1908 c1[d] = zones->size[zone].bb_x1[d];
1911 nbnxn_put_on_grid(nbs, nbs->ePBC, NULL,
1913 zones->cg_range[zone],
1914 zones->cg_range[zone+1],
1924 /* Add simple grid type information to the local super/sub grid */
1925 void nbnxn_grid_add_simple(nbnxn_search_t nbs,
1926 nbnxn_atomdata_t *nbat)
1932 int nthreads gmx_unused;
1934 grid = &nbs->grid[0];
1938 gmx_incons("nbnxn_grid_simple called with a simple grid");
1941 ncd = grid->na_sc/NBNXN_CPU_CLUSTER_I_SIZE;
1943 if (grid->nc*ncd > grid->nc_nalloc_simple)
1945 grid->nc_nalloc_simple = over_alloc_large(grid->nc*ncd);
1946 srenew(grid->bbcz_simple, grid->nc_nalloc_simple*NNBSBB_D);
1947 srenew(grid->bb_simple, grid->nc_nalloc_simple);
1948 srenew(grid->flags_simple, grid->nc_nalloc_simple);
1951 sfree_aligned(grid->bbj);
1952 snew_aligned(grid->bbj, grid->nc_nalloc_simple/2, 16);
1956 bbcz = grid->bbcz_simple;
1957 bb = grid->bb_simple;
1959 nthreads = gmx_omp_nthreads_get(emntPairsearch);
1960 #pragma omp parallel for num_threads(nthreads) schedule(static)
1961 for (sc = 0; sc < grid->nc; sc++)
1965 for (c = 0; c < ncd; c++)
1969 na = NBNXN_CPU_CLUSTER_I_SIZE;
1971 nbat->type[tx*NBNXN_CPU_CLUSTER_I_SIZE+na-1] == nbat->ntype-1)
1978 switch (nbat->XFormat)
1981 /* PACK_X4==NBNXN_CPU_CLUSTER_I_SIZE, so this is simple */
1982 calc_bounding_box_x_x4(na, nbat->x+tx*STRIDE_P4,
1986 /* PACK_X8>NBNXN_CPU_CLUSTER_I_SIZE, more complicated */
1987 calc_bounding_box_x_x8(na, nbat->x+X8_IND_A(tx*NBNXN_CPU_CLUSTER_I_SIZE),
1991 calc_bounding_box(na, nbat->xstride,
1992 nbat->x+tx*NBNXN_CPU_CLUSTER_I_SIZE*nbat->xstride,
1996 bbcz[tx*NNBSBB_D+0] = bb[tx].lower[BB_Z];
1997 bbcz[tx*NNBSBB_D+1] = bb[tx].upper[BB_Z];
1999 /* No interaction optimization yet here */
2000 grid->flags_simple[tx] = NBNXN_CI_DO_LJ(0) | NBNXN_CI_DO_COUL(0);
2004 grid->flags_simple[tx] = 0;
2009 if (grid->bSimple && nbat->XFormat == nbatX8)
2011 combine_bounding_box_pairs(grid, grid->bb_simple);
2015 void nbnxn_get_ncells(nbnxn_search_t nbs, int *ncx, int *ncy)
2017 *ncx = nbs->grid[0].ncx;
2018 *ncy = nbs->grid[0].ncy;
2021 void nbnxn_get_atomorder(nbnxn_search_t nbs, int **a, int *n)
2023 const nbnxn_grid_t *grid;
2025 grid = &nbs->grid[0];
2027 /* Return the atom order for the home cell (index 0) */
2030 *n = grid->cxy_ind[grid->ncx*grid->ncy]*grid->na_sc;
2033 void nbnxn_set_atomorder(nbnxn_search_t nbs)
2036 int ao, cx, cy, cxy, cz, j;
2038 /* Set the atom order for the home cell (index 0) */
2039 grid = &nbs->grid[0];
2042 for (cx = 0; cx < grid->ncx; cx++)
2044 for (cy = 0; cy < grid->ncy; cy++)
2046 cxy = cx*grid->ncy + cy;
2047 j = grid->cxy_ind[cxy]*grid->na_sc;
2048 for (cz = 0; cz < grid->cxy_na[cxy]; cz++)
2059 /* Determines the cell range along one dimension that
2060 * the bounding box b0 - b1 sees.
2062 static void get_cell_range(real b0, real b1,
2063 int nc, real c0, real s, real invs,
2064 real d2, real r2, int *cf, int *cl)
2066 *cf = max((int)((b0 - c0)*invs), 0);
2068 while (*cf > 0 && d2 + sqr((b0 - c0) - (*cf-1+1)*s) < r2)
2073 *cl = min((int)((b1 - c0)*invs), nc-1);
2074 while (*cl < nc-1 && d2 + sqr((*cl+1)*s - (b1 - c0)) < r2)
2080 /* Reference code calculating the distance^2 between two bounding boxes */
2081 static float box_dist2(float bx0, float bx1, float by0,
2082 float by1, float bz0, float bz1,
2083 const nbnxn_bb_t *bb)
2086 float dl, dh, dm, dm0;
2090 dl = bx0 - bb->upper[BB_X];
2091 dh = bb->lower[BB_X] - bx1;
2096 dl = by0 - bb->upper[BB_Y];
2097 dh = bb->lower[BB_Y] - by1;
2102 dl = bz0 - bb->upper[BB_Z];
2103 dh = bb->lower[BB_Z] - bz1;
2111 /* Plain C code calculating the distance^2 between two bounding boxes */
2112 static float subc_bb_dist2(int si, const nbnxn_bb_t *bb_i_ci,
2113 int csj, const nbnxn_bb_t *bb_j_all)
2115 const nbnxn_bb_t *bb_i, *bb_j;
2117 float dl, dh, dm, dm0;
2119 bb_i = bb_i_ci + si;
2120 bb_j = bb_j_all + csj;
2124 dl = bb_i->lower[BB_X] - bb_j->upper[BB_X];
2125 dh = bb_j->lower[BB_X] - bb_i->upper[BB_X];
2130 dl = bb_i->lower[BB_Y] - bb_j->upper[BB_Y];
2131 dh = bb_j->lower[BB_Y] - bb_i->upper[BB_Y];
2136 dl = bb_i->lower[BB_Z] - bb_j->upper[BB_Z];
2137 dh = bb_j->lower[BB_Z] - bb_i->upper[BB_Z];
2145 #ifdef NBNXN_SEARCH_BB_SIMD4
2147 /* 4-wide SIMD code for bb distance for bb format xyz0 */
2148 static float subc_bb_dist2_simd4(int si, const nbnxn_bb_t *bb_i_ci,
2149 int csj, const nbnxn_bb_t *bb_j_all)
2151 gmx_simd4_float_t bb_i_S0, bb_i_S1;
2152 gmx_simd4_float_t bb_j_S0, bb_j_S1;
2153 gmx_simd4_float_t dl_S;
2154 gmx_simd4_float_t dh_S;
2155 gmx_simd4_float_t dm_S;
2156 gmx_simd4_float_t dm0_S;
2158 bb_i_S0 = gmx_simd4_load_f(&bb_i_ci[si].lower[0]);
2159 bb_i_S1 = gmx_simd4_load_f(&bb_i_ci[si].upper[0]);
2160 bb_j_S0 = gmx_simd4_load_f(&bb_j_all[csj].lower[0]);
2161 bb_j_S1 = gmx_simd4_load_f(&bb_j_all[csj].upper[0]);
2163 dl_S = gmx_simd4_sub_f(bb_i_S0, bb_j_S1);
2164 dh_S = gmx_simd4_sub_f(bb_j_S0, bb_i_S1);
2166 dm_S = gmx_simd4_max_f(dl_S, dh_S);
2167 dm0_S = gmx_simd4_max_f(dm_S, gmx_simd4_setzero_f());
2169 return gmx_simd4_dotproduct3_f(dm0_S, dm0_S);
2172 /* Calculate bb bounding distances of bb_i[si,...,si+3] and store them in d2 */
2173 #define SUBC_BB_DIST2_SIMD4_XXXX_INNER(si, bb_i, d2) \
2177 gmx_simd4_float_t dx_0, dy_0, dz_0; \
2178 gmx_simd4_float_t dx_1, dy_1, dz_1; \
2180 gmx_simd4_float_t mx, my, mz; \
2181 gmx_simd4_float_t m0x, m0y, m0z; \
2183 gmx_simd4_float_t d2x, d2y, d2z; \
2184 gmx_simd4_float_t d2s, d2t; \
2186 shi = si*NNBSBB_D*DIM; \
2188 xi_l = gmx_simd4_load_f(bb_i+shi+0*STRIDE_PBB); \
2189 yi_l = gmx_simd4_load_f(bb_i+shi+1*STRIDE_PBB); \
2190 zi_l = gmx_simd4_load_f(bb_i+shi+2*STRIDE_PBB); \
2191 xi_h = gmx_simd4_load_f(bb_i+shi+3*STRIDE_PBB); \
2192 yi_h = gmx_simd4_load_f(bb_i+shi+4*STRIDE_PBB); \
2193 zi_h = gmx_simd4_load_f(bb_i+shi+5*STRIDE_PBB); \
2195 dx_0 = gmx_simd4_sub_f(xi_l, xj_h); \
2196 dy_0 = gmx_simd4_sub_f(yi_l, yj_h); \
2197 dz_0 = gmx_simd4_sub_f(zi_l, zj_h); \
2199 dx_1 = gmx_simd4_sub_f(xj_l, xi_h); \
2200 dy_1 = gmx_simd4_sub_f(yj_l, yi_h); \
2201 dz_1 = gmx_simd4_sub_f(zj_l, zi_h); \
2203 mx = gmx_simd4_max_f(dx_0, dx_1); \
2204 my = gmx_simd4_max_f(dy_0, dy_1); \
2205 mz = gmx_simd4_max_f(dz_0, dz_1); \
2207 m0x = gmx_simd4_max_f(mx, zero); \
2208 m0y = gmx_simd4_max_f(my, zero); \
2209 m0z = gmx_simd4_max_f(mz, zero); \
2211 d2x = gmx_simd4_mul_f(m0x, m0x); \
2212 d2y = gmx_simd4_mul_f(m0y, m0y); \
2213 d2z = gmx_simd4_mul_f(m0z, m0z); \
2215 d2s = gmx_simd4_add_f(d2x, d2y); \
2216 d2t = gmx_simd4_add_f(d2s, d2z); \
2218 gmx_simd4_store_f(d2+si, d2t); \
2221 /* 4-wide SIMD code for nsi bb distances for bb format xxxxyyyyzzzz */
2222 static void subc_bb_dist2_simd4_xxxx(const float *bb_j,
2223 int nsi, const float *bb_i,
2226 gmx_simd4_float_t xj_l, yj_l, zj_l;
2227 gmx_simd4_float_t xj_h, yj_h, zj_h;
2228 gmx_simd4_float_t xi_l, yi_l, zi_l;
2229 gmx_simd4_float_t xi_h, yi_h, zi_h;
2231 gmx_simd4_float_t zero;
2233 zero = gmx_simd4_setzero_f();
2235 xj_l = gmx_simd4_set1_f(bb_j[0*STRIDE_PBB]);
2236 yj_l = gmx_simd4_set1_f(bb_j[1*STRIDE_PBB]);
2237 zj_l = gmx_simd4_set1_f(bb_j[2*STRIDE_PBB]);
2238 xj_h = gmx_simd4_set1_f(bb_j[3*STRIDE_PBB]);
2239 yj_h = gmx_simd4_set1_f(bb_j[4*STRIDE_PBB]);
2240 zj_h = gmx_simd4_set1_f(bb_j[5*STRIDE_PBB]);
2242 /* Here we "loop" over si (0,STRIDE_PBB) from 0 to nsi with step STRIDE_PBB.
2243 * But as we know the number of iterations is 1 or 2, we unroll manually.
2245 SUBC_BB_DIST2_SIMD4_XXXX_INNER(0, bb_i, d2);
2246 if (STRIDE_PBB < nsi)
2248 SUBC_BB_DIST2_SIMD4_XXXX_INNER(STRIDE_PBB, bb_i, d2);
2252 #endif /* NBNXN_SEARCH_BB_SIMD4 */
2254 /* Plain C function which determines if any atom pair between two cells
2255 * is within distance sqrt(rl2).
2257 static gmx_bool subc_in_range_x(int na_c,
2258 int si, const real *x_i,
2259 int csj, int stride, const real *x_j,
2265 for (i = 0; i < na_c; i++)
2267 i0 = (si*na_c + i)*DIM;
2268 for (j = 0; j < na_c; j++)
2270 j0 = (csj*na_c + j)*stride;
2272 d2 = sqr(x_i[i0 ] - x_j[j0 ]) +
2273 sqr(x_i[i0+1] - x_j[j0+1]) +
2274 sqr(x_i[i0+2] - x_j[j0+2]);
2286 #ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
2288 /* 4-wide SIMD function which determines if any atom pair between two cells,
2289 * both with 8 atoms, is within distance sqrt(rl2).
2290 * Using 8-wide AVX is not faster on Intel Sandy Bridge.
2292 static gmx_bool subc_in_range_simd4(int na_c,
2293 int si, const real *x_i,
2294 int csj, int stride, const real *x_j,
2297 gmx_simd4_real_t ix_S0, iy_S0, iz_S0;
2298 gmx_simd4_real_t ix_S1, iy_S1, iz_S1;
2300 gmx_simd4_real_t rc2_S;
2305 rc2_S = gmx_simd4_set1_r(rl2);
2307 dim_stride = NBNXN_GPU_CLUSTER_SIZE/STRIDE_PBB*DIM;
2308 ix_S0 = gmx_simd4_load_r(x_i+(si*dim_stride+0)*STRIDE_PBB);
2309 iy_S0 = gmx_simd4_load_r(x_i+(si*dim_stride+1)*STRIDE_PBB);
2310 iz_S0 = gmx_simd4_load_r(x_i+(si*dim_stride+2)*STRIDE_PBB);
2311 ix_S1 = gmx_simd4_load_r(x_i+(si*dim_stride+3)*STRIDE_PBB);
2312 iy_S1 = gmx_simd4_load_r(x_i+(si*dim_stride+4)*STRIDE_PBB);
2313 iz_S1 = gmx_simd4_load_r(x_i+(si*dim_stride+5)*STRIDE_PBB);
2315 /* We loop from the outer to the inner particles to maximize
2316 * the chance that we find a pair in range quickly and return.
2322 gmx_simd4_real_t jx0_S, jy0_S, jz0_S;
2323 gmx_simd4_real_t jx1_S, jy1_S, jz1_S;
2325 gmx_simd4_real_t dx_S0, dy_S0, dz_S0;
2326 gmx_simd4_real_t dx_S1, dy_S1, dz_S1;
2327 gmx_simd4_real_t dx_S2, dy_S2, dz_S2;
2328 gmx_simd4_real_t dx_S3, dy_S3, dz_S3;
2330 gmx_simd4_real_t rsq_S0;
2331 gmx_simd4_real_t rsq_S1;
2332 gmx_simd4_real_t rsq_S2;
2333 gmx_simd4_real_t rsq_S3;
2335 gmx_simd4_bool_t wco_S0;
2336 gmx_simd4_bool_t wco_S1;
2337 gmx_simd4_bool_t wco_S2;
2338 gmx_simd4_bool_t wco_S3;
2339 gmx_simd4_bool_t wco_any_S01, wco_any_S23, wco_any_S;
2341 jx0_S = gmx_simd4_set1_r(x_j[j0*stride+0]);
2342 jy0_S = gmx_simd4_set1_r(x_j[j0*stride+1]);
2343 jz0_S = gmx_simd4_set1_r(x_j[j0*stride+2]);
2345 jx1_S = gmx_simd4_set1_r(x_j[j1*stride+0]);
2346 jy1_S = gmx_simd4_set1_r(x_j[j1*stride+1]);
2347 jz1_S = gmx_simd4_set1_r(x_j[j1*stride+2]);
2349 /* Calculate distance */
2350 dx_S0 = gmx_simd4_sub_r(ix_S0, jx0_S);
2351 dy_S0 = gmx_simd4_sub_r(iy_S0, jy0_S);
2352 dz_S0 = gmx_simd4_sub_r(iz_S0, jz0_S);
2353 dx_S1 = gmx_simd4_sub_r(ix_S1, jx0_S);
2354 dy_S1 = gmx_simd4_sub_r(iy_S1, jy0_S);
2355 dz_S1 = gmx_simd4_sub_r(iz_S1, jz0_S);
2356 dx_S2 = gmx_simd4_sub_r(ix_S0, jx1_S);
2357 dy_S2 = gmx_simd4_sub_r(iy_S0, jy1_S);
2358 dz_S2 = gmx_simd4_sub_r(iz_S0, jz1_S);
2359 dx_S3 = gmx_simd4_sub_r(ix_S1, jx1_S);
2360 dy_S3 = gmx_simd4_sub_r(iy_S1, jy1_S);
2361 dz_S3 = gmx_simd4_sub_r(iz_S1, jz1_S);
2363 /* rsq = dx*dx+dy*dy+dz*dz */
2364 rsq_S0 = gmx_simd4_calc_rsq_r(dx_S0, dy_S0, dz_S0);
2365 rsq_S1 = gmx_simd4_calc_rsq_r(dx_S1, dy_S1, dz_S1);
2366 rsq_S2 = gmx_simd4_calc_rsq_r(dx_S2, dy_S2, dz_S2);
2367 rsq_S3 = gmx_simd4_calc_rsq_r(dx_S3, dy_S3, dz_S3);
2369 wco_S0 = gmx_simd4_cmplt_r(rsq_S0, rc2_S);
2370 wco_S1 = gmx_simd4_cmplt_r(rsq_S1, rc2_S);
2371 wco_S2 = gmx_simd4_cmplt_r(rsq_S2, rc2_S);
2372 wco_S3 = gmx_simd4_cmplt_r(rsq_S3, rc2_S);
2374 wco_any_S01 = gmx_simd4_or_b(wco_S0, wco_S1);
2375 wco_any_S23 = gmx_simd4_or_b(wco_S2, wco_S3);
2376 wco_any_S = gmx_simd4_or_b(wco_any_S01, wco_any_S23);
2378 if (gmx_simd4_anytrue_b(wco_any_S))
2392 /* Returns the j sub-cell for index cj_ind */
2393 static int nbl_cj(const nbnxn_pairlist_t *nbl, int cj_ind)
2395 return nbl->cj4[cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG].cj[cj_ind & (NBNXN_GPU_JGROUP_SIZE - 1)];
2398 /* Returns the i-interaction mask of the j sub-cell for index cj_ind */
2399 static unsigned int nbl_imask0(const nbnxn_pairlist_t *nbl, int cj_ind)
2401 return nbl->cj4[cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG].imei[0].imask;
2404 /* Ensures there is enough space for extra extra exclusion masks */
2405 static void check_excl_space(nbnxn_pairlist_t *nbl, int extra)
2407 if (nbl->nexcl+extra > nbl->excl_nalloc)
2409 nbl->excl_nalloc = over_alloc_small(nbl->nexcl+extra);
2410 nbnxn_realloc_void((void **)&nbl->excl,
2411 nbl->nexcl*sizeof(*nbl->excl),
2412 nbl->excl_nalloc*sizeof(*nbl->excl),
2413 nbl->alloc, nbl->free);
2417 /* Ensures there is enough space for ncell extra j-cells in the list */
2418 static void check_subcell_list_space_simple(nbnxn_pairlist_t *nbl,
2423 cj_max = nbl->ncj + ncell;
2425 if (cj_max > nbl->cj_nalloc)
2427 nbl->cj_nalloc = over_alloc_small(cj_max);
2428 nbnxn_realloc_void((void **)&nbl->cj,
2429 nbl->ncj*sizeof(*nbl->cj),
2430 nbl->cj_nalloc*sizeof(*nbl->cj),
2431 nbl->alloc, nbl->free);
2435 /* Ensures there is enough space for ncell extra j-subcells in the list */
2436 static void check_subcell_list_space_supersub(nbnxn_pairlist_t *nbl,
2439 int ncj4_max, j4, j, w, t;
2442 #define WARP_SIZE 32
2444 /* We can have maximally nsupercell*GPU_NSUBCELL sj lists */
2445 /* We can store 4 j-subcell - i-supercell pairs in one struct.
2446 * since we round down, we need one extra entry.
2448 ncj4_max = ((nbl->work->cj_ind + nsupercell*GPU_NSUBCELL + NBNXN_GPU_JGROUP_SIZE - 1) >> NBNXN_GPU_JGROUP_SIZE_2LOG);
2450 if (ncj4_max > nbl->cj4_nalloc)
2452 nbl->cj4_nalloc = over_alloc_small(ncj4_max);
2453 nbnxn_realloc_void((void **)&nbl->cj4,
2454 nbl->work->cj4_init*sizeof(*nbl->cj4),
2455 nbl->cj4_nalloc*sizeof(*nbl->cj4),
2456 nbl->alloc, nbl->free);
2459 if (ncj4_max > nbl->work->cj4_init)
2461 for (j4 = nbl->work->cj4_init; j4 < ncj4_max; j4++)
2463 /* No i-subcells and no excl's in the list initially */
2464 for (w = 0; w < NWARP; w++)
2466 nbl->cj4[j4].imei[w].imask = 0U;
2467 nbl->cj4[j4].imei[w].excl_ind = 0;
2471 nbl->work->cj4_init = ncj4_max;
2475 /* Set all excl masks for one GPU warp no exclusions */
2476 static void set_no_excls(nbnxn_excl_t *excl)
2480 for (t = 0; t < WARP_SIZE; t++)
2482 /* Turn all interaction bits on */
2483 excl->pair[t] = NBNXN_INTERACTION_MASK_ALL;
2487 /* Initializes a single nbnxn_pairlist_t data structure */
2488 static void nbnxn_init_pairlist(nbnxn_pairlist_t *nbl,
2490 nbnxn_alloc_t *alloc,
2495 nbl->alloc = nbnxn_alloc_aligned;
2503 nbl->free = nbnxn_free_aligned;
2510 nbl->bSimple = bSimple;
2521 /* We need one element extra in sj, so alloc initially with 1 */
2522 nbl->cj4_nalloc = 0;
2529 nbl->excl_nalloc = 0;
2531 check_excl_space(nbl, 1);
2533 set_no_excls(&nbl->excl[0]);
2539 snew_aligned(nbl->work->bb_ci, 1, NBNXN_SEARCH_BB_MEM_ALIGN);
2544 snew_aligned(nbl->work->pbb_ci, GPU_NSUBCELL/STRIDE_PBB*NNBSBB_XXXX, NBNXN_SEARCH_BB_MEM_ALIGN);
2546 snew_aligned(nbl->work->bb_ci, GPU_NSUBCELL, NBNXN_SEARCH_BB_MEM_ALIGN);
2549 snew_aligned(nbl->work->x_ci, NBNXN_NA_SC_MAX*DIM, NBNXN_SEARCH_BB_MEM_ALIGN);
2550 #ifdef GMX_NBNXN_SIMD
2551 snew_aligned(nbl->work->x_ci_simd_4xn, 1, NBNXN_MEM_ALIGN);
2552 snew_aligned(nbl->work->x_ci_simd_2xnn, 1, NBNXN_MEM_ALIGN);
2554 snew_aligned(nbl->work->d2, GPU_NSUBCELL, NBNXN_SEARCH_BB_MEM_ALIGN);
2556 nbl->work->sort = NULL;
2557 nbl->work->sort_nalloc = 0;
2558 nbl->work->sci_sort = NULL;
2559 nbl->work->sci_sort_nalloc = 0;
2562 void nbnxn_init_pairlist_set(nbnxn_pairlist_set_t *nbl_list,
2563 gmx_bool bSimple, gmx_bool bCombined,
2564 nbnxn_alloc_t *alloc,
2569 nbl_list->bSimple = bSimple;
2570 nbl_list->bCombined = bCombined;
2572 nbl_list->nnbl = gmx_omp_nthreads_get(emntNonbonded);
2574 if (!nbl_list->bCombined &&
2575 nbl_list->nnbl > NBNXN_BUFFERFLAG_MAX_THREADS)
2577 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.",
2578 nbl_list->nnbl, NBNXN_BUFFERFLAG_MAX_THREADS, NBNXN_BUFFERFLAG_MAX_THREADS);
2581 snew(nbl_list->nbl, nbl_list->nnbl);
2582 snew(nbl_list->nbl_fep, nbl_list->nnbl);
2583 /* Execute in order to avoid memory interleaving between threads */
2584 #pragma omp parallel for num_threads(nbl_list->nnbl) schedule(static)
2585 for (i = 0; i < nbl_list->nnbl; i++)
2587 /* Allocate the nblist data structure locally on each thread
2588 * to optimize memory access for NUMA architectures.
2590 snew(nbl_list->nbl[i], 1);
2592 /* Only list 0 is used on the GPU, use normal allocation for i>0 */
2595 nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, alloc, free);
2599 nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, NULL, NULL);
2602 snew(nbl_list->nbl_fep[i], 1);
2603 nbnxn_init_pairlist_fep(nbl_list->nbl_fep[i]);
2607 /* Print statistics of a pair list, used for debug output */
2608 static void print_nblist_statistics_simple(FILE *fp, const nbnxn_pairlist_t *nbl,
2609 const nbnxn_search_t nbs, real rl)
2611 const nbnxn_grid_t *grid;
2616 /* This code only produces correct statistics with domain decomposition */
2617 grid = &nbs->grid[0];
2619 fprintf(fp, "nbl nci %d ncj %d\n",
2620 nbl->nci, nbl->ncj);
2621 fprintf(fp, "nbl na_sc %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
2622 nbl->na_sc, rl, nbl->ncj, nbl->ncj/(double)grid->nc,
2623 nbl->ncj/(double)grid->nc*grid->na_sc,
2624 nbl->ncj/(double)grid->nc*grid->na_sc/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid->nc*grid->na_sc/(grid->size[XX]*grid->size[YY]*grid->size[ZZ])));
2626 fprintf(fp, "nbl average j cell list length %.1f\n",
2627 0.25*nbl->ncj/(double)max(nbl->nci, 1));
2629 for (s = 0; s < SHIFTS; s++)
2634 for (i = 0; i < nbl->nci; i++)
2636 cs[nbl->ci[i].shift & NBNXN_CI_SHIFT] +=
2637 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start;
2639 j = nbl->ci[i].cj_ind_start;
2640 while (j < nbl->ci[i].cj_ind_end &&
2641 nbl->cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
2647 fprintf(fp, "nbl cell pairs, total: %d excl: %d %.1f%%\n",
2648 nbl->ncj, npexcl, 100*npexcl/(double)max(nbl->ncj, 1));
2649 for (s = 0; s < SHIFTS; s++)
2653 fprintf(fp, "nbl shift %2d ncj %3d\n", s, cs[s]);
2658 /* Print statistics of a pair lists, used for debug output */
2659 static void print_nblist_statistics_supersub(FILE *fp, const nbnxn_pairlist_t *nbl,
2660 const nbnxn_search_t nbs, real rl)
2662 const nbnxn_grid_t *grid;
2663 int i, j4, j, si, b;
2664 int c[GPU_NSUBCELL+1];
2665 double sum_nsp, sum_nsp2;
2668 /* This code only produces correct statistics with domain decomposition */
2669 grid = &nbs->grid[0];
2671 fprintf(fp, "nbl nsci %d ncj4 %d nsi %d excl4 %d\n",
2672 nbl->nsci, nbl->ncj4, nbl->nci_tot, nbl->nexcl);
2673 fprintf(fp, "nbl na_c %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
2674 nbl->na_ci, rl, nbl->nci_tot, nbl->nci_tot/(double)grid->nsubc_tot,
2675 nbl->nci_tot/(double)grid->nsubc_tot*grid->na_c,
2676 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/(grid->size[XX]*grid->size[YY]*grid->size[ZZ])));
2681 for (si = 0; si <= GPU_NSUBCELL; si++)
2685 for (i = 0; i < nbl->nsci; i++)
2690 for (j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
2692 for (j = 0; j < NBNXN_GPU_JGROUP_SIZE; j++)
2695 for (si = 0; si < GPU_NSUBCELL; si++)
2697 if (nbl->cj4[j4].imei[0].imask & (1U << (j*GPU_NSUBCELL + si)))
2707 sum_nsp2 += nsp*nsp;
2708 nsp_max = max(nsp_max, nsp);
2712 sum_nsp /= nbl->nsci;
2713 sum_nsp2 /= nbl->nsci;
2715 fprintf(fp, "nbl #cluster-pairs: av %.1f stddev %.1f max %d\n",
2716 sum_nsp, sqrt(sum_nsp2 - sum_nsp*sum_nsp), nsp_max);
2720 for (b = 0; b <= GPU_NSUBCELL; b++)
2722 fprintf(fp, "nbl j-list #i-subcell %d %7d %4.1f\n",
2724 100.0*c[b]/(double)(nbl->ncj4*NBNXN_GPU_JGROUP_SIZE));
2729 /* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp */
2730 static void low_get_nbl_exclusions(nbnxn_pairlist_t *nbl, int cj4,
2731 int warp, nbnxn_excl_t **excl)
2733 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
2735 /* No exclusions set, make a new list entry */
2736 nbl->cj4[cj4].imei[warp].excl_ind = nbl->nexcl;
2738 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
2739 set_no_excls(*excl);
2743 /* We already have some exclusions, new ones can be added to the list */
2744 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
2748 /* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp,
2749 * generates a new element and allocates extra memory, if necessary.
2751 static void get_nbl_exclusions_1(nbnxn_pairlist_t *nbl, int cj4,
2752 int warp, nbnxn_excl_t **excl)
2754 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
2756 /* We need to make a new list entry, check if we have space */
2757 check_excl_space(nbl, 1);
2759 low_get_nbl_exclusions(nbl, cj4, warp, excl);
2762 /* Returns pointers to the exclusion mask for cj4-unit cj4 for both warps,
2763 * generates a new element and allocates extra memory, if necessary.
2765 static void get_nbl_exclusions_2(nbnxn_pairlist_t *nbl, int cj4,
2766 nbnxn_excl_t **excl_w0,
2767 nbnxn_excl_t **excl_w1)
2769 /* Check for space we might need */
2770 check_excl_space(nbl, 2);
2772 low_get_nbl_exclusions(nbl, cj4, 0, excl_w0);
2773 low_get_nbl_exclusions(nbl, cj4, 1, excl_w1);
2776 /* Sets the self exclusions i=j and pair exclusions i>j */
2777 static void set_self_and_newton_excls_supersub(nbnxn_pairlist_t *nbl,
2778 int cj4_ind, int sj_offset,
2781 nbnxn_excl_t *excl[2];
2784 /* Here we only set the set self and double pair exclusions */
2786 get_nbl_exclusions_2(nbl, cj4_ind, &excl[0], &excl[1]);
2788 /* Only minor < major bits set */
2789 for (ej = 0; ej < nbl->na_ci; ej++)
2792 for (ei = ej; ei < nbl->na_ci; ei++)
2794 excl[w]->pair[(ej & (NBNXN_GPU_JGROUP_SIZE-1))*nbl->na_ci + ei] &=
2795 ~(1U << (sj_offset*GPU_NSUBCELL + si));
2800 /* Returns a diagonal or off-diagonal interaction mask for plain C lists */
2801 static unsigned int get_imask(gmx_bool rdiag, int ci, int cj)
2803 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
2806 /* Returns a diagonal or off-diagonal interaction mask for cj-size=2 */
2807 static unsigned int get_imask_simd_j2(gmx_bool rdiag, int ci, int cj)
2809 return (rdiag && ci*2 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_0 :
2810 (rdiag && ci*2+1 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_1 :
2811 NBNXN_INTERACTION_MASK_ALL));
2814 /* Returns a diagonal or off-diagonal interaction mask for cj-size=4 */
2815 static unsigned int get_imask_simd_j4(gmx_bool rdiag, int ci, int cj)
2817 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
2820 /* Returns a diagonal or off-diagonal interaction mask for cj-size=8 */
2821 static unsigned int get_imask_simd_j8(gmx_bool rdiag, int ci, int cj)
2823 return (rdiag && ci == cj*2 ? NBNXN_INTERACTION_MASK_DIAG_J8_0 :
2824 (rdiag && ci == cj*2+1 ? NBNXN_INTERACTION_MASK_DIAG_J8_1 :
2825 NBNXN_INTERACTION_MASK_ALL));
2828 #ifdef GMX_NBNXN_SIMD
2829 #if GMX_SIMD_REAL_WIDTH == 2
2830 #define get_imask_simd_4xn get_imask_simd_j2
2832 #if GMX_SIMD_REAL_WIDTH == 4
2833 #define get_imask_simd_4xn get_imask_simd_j4
2835 #if GMX_SIMD_REAL_WIDTH == 8
2836 #define get_imask_simd_4xn get_imask_simd_j8
2837 #define get_imask_simd_2xnn get_imask_simd_j4
2839 #if GMX_SIMD_REAL_WIDTH == 16
2840 #define get_imask_simd_2xnn get_imask_simd_j8
2844 /* Plain C code for making a pair list of cell ci vs cell cjf-cjl.
2845 * Checks bounding box distances and possibly atom pair distances.
2847 static void make_cluster_list_simple(const nbnxn_grid_t *gridj,
2848 nbnxn_pairlist_t *nbl,
2849 int ci, int cjf, int cjl,
2850 gmx_bool remove_sub_diag,
2852 real rl2, float rbb2,
2855 const nbnxn_list_work_t *work;
2857 const nbnxn_bb_t *bb_ci;
2862 int cjf_gl, cjl_gl, cj;
2866 bb_ci = nbl->work->bb_ci;
2867 x_ci = nbl->work->x_ci;
2870 while (!InRange && cjf <= cjl)
2872 d2 = subc_bb_dist2(0, bb_ci, cjf, gridj->bb);
2875 /* Check if the distance is within the distance where
2876 * we use only the bounding box distance rbb,
2877 * or within the cut-off and there is at least one atom pair
2878 * within the cut-off.
2888 cjf_gl = gridj->cell0 + cjf;
2889 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
2891 for (j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE; j++)
2893 InRange = InRange ||
2894 (sqr(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
2895 sqr(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
2896 sqr(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rl2);
2899 *ndistc += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
2912 while (!InRange && cjl > cjf)
2914 d2 = subc_bb_dist2(0, bb_ci, cjl, gridj->bb);
2917 /* Check if the distance is within the distance where
2918 * we use only the bounding box distance rbb,
2919 * or within the cut-off and there is at least one atom pair
2920 * within the cut-off.
2930 cjl_gl = gridj->cell0 + cjl;
2931 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
2933 for (j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE; j++)
2935 InRange = InRange ||
2936 (sqr(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
2937 sqr(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
2938 sqr(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rl2);
2941 *ndistc += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
2951 for (cj = cjf; cj <= cjl; cj++)
2953 /* Store cj and the interaction mask */
2954 nbl->cj[nbl->ncj].cj = gridj->cell0 + cj;
2955 nbl->cj[nbl->ncj].excl = get_imask(remove_sub_diag, ci, cj);
2958 /* Increase the closing index in i super-cell list */
2959 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
2963 #ifdef GMX_NBNXN_SIMD_4XN
2964 #include "gromacs/mdlib/nbnxn_search_simd_4xn.h"
2966 #ifdef GMX_NBNXN_SIMD_2XNN
2967 #include "gromacs/mdlib/nbnxn_search_simd_2xnn.h"
2970 /* Plain C or SIMD4 code for making a pair list of super-cell sci vs scj.
2971 * Checks bounding box distances and possibly atom pair distances.
2973 static void make_cluster_list_supersub(const nbnxn_grid_t *gridi,
2974 const nbnxn_grid_t *gridj,
2975 nbnxn_pairlist_t *nbl,
2977 gmx_bool sci_equals_scj,
2978 int stride, const real *x,
2979 real rl2, float rbb2,
2984 int cjo, ci1, ci, cj, cj_gl;
2985 int cj4_ind, cj_offset;
2989 const float *pbb_ci;
2991 const nbnxn_bb_t *bb_ci;
2996 #define PRUNE_LIST_CPU_ONE
2997 #ifdef PRUNE_LIST_CPU_ONE
3001 d2l = nbl->work->d2;
3004 pbb_ci = nbl->work->pbb_ci;
3006 bb_ci = nbl->work->bb_ci;
3008 x_ci = nbl->work->x_ci;
3012 for (cjo = 0; cjo < gridj->nsubc[scj]; cjo++)
3014 cj4_ind = (nbl->work->cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG);
3015 cj_offset = nbl->work->cj_ind - cj4_ind*NBNXN_GPU_JGROUP_SIZE;
3016 cj4 = &nbl->cj4[cj4_ind];
3018 cj = scj*GPU_NSUBCELL + cjo;
3020 cj_gl = gridj->cell0*GPU_NSUBCELL + cj;
3022 /* Initialize this j-subcell i-subcell list */
3023 cj4->cj[cj_offset] = cj_gl;
3032 ci1 = gridi->nsubc[sci];
3036 /* Determine all ci1 bb distances in one call with SIMD4 */
3037 subc_bb_dist2_simd4_xxxx(gridj->pbb+(cj>>STRIDE_PBB_2LOG)*NNBSBB_XXXX+(cj & (STRIDE_PBB-1)),
3043 /* We use a fixed upper-bound instead of ci1 to help optimization */
3044 for (ci = 0; ci < GPU_NSUBCELL; ci++)
3051 #ifndef NBNXN_BBXXXX
3052 /* Determine the bb distance between ci and cj */
3053 d2l[ci] = subc_bb_dist2(ci, bb_ci, cj, gridj->bb);
3058 #ifdef PRUNE_LIST_CPU_ALL
3059 /* Check if the distance is within the distance where
3060 * we use only the bounding box distance rbb,
3061 * or within the cut-off and there is at least one atom pair
3062 * within the cut-off. This check is very costly.
3064 *ndistc += na_c*na_c;
3067 #ifdef NBNXN_PBB_SIMD4
3072 (na_c, ci, x_ci, cj_gl, stride, x, rl2)))
3074 /* Check if the distance between the two bounding boxes
3075 * in within the pair-list cut-off.
3080 /* Flag this i-subcell to be taken into account */
3081 imask |= (1U << (cj_offset*GPU_NSUBCELL+ci));
3083 #ifdef PRUNE_LIST_CPU_ONE
3091 #ifdef PRUNE_LIST_CPU_ONE
3092 /* If we only found 1 pair, check if any atoms are actually
3093 * within the cut-off, so we could get rid of it.
3095 if (npair == 1 && d2l[ci_last] >= rbb2)
3097 /* Avoid using function pointers here, as it's slower */
3099 #ifdef NBNXN_PBB_SIMD4
3100 !subc_in_range_simd4
3104 (na_c, ci_last, x_ci, cj_gl, stride, x, rl2))
3106 imask &= ~(1U << (cj_offset*GPU_NSUBCELL+ci_last));
3114 /* We have a useful sj entry, close it now */
3116 /* Set the exclucions for the ci== sj entry.
3117 * Here we don't bother to check if this entry is actually flagged,
3118 * as it will nearly always be in the list.
3122 set_self_and_newton_excls_supersub(nbl, cj4_ind, cj_offset, cjo);
3125 /* Copy the cluster interaction mask to the list */
3126 for (w = 0; w < NWARP; w++)
3128 cj4->imei[w].imask |= imask;
3131 nbl->work->cj_ind++;
3133 /* Keep the count */
3134 nbl->nci_tot += npair;
3136 /* Increase the closing index in i super-cell list */
3137 nbl->sci[nbl->nsci].cj4_ind_end =
3138 ((nbl->work->cj_ind+NBNXN_GPU_JGROUP_SIZE-1) >> NBNXN_GPU_JGROUP_SIZE_2LOG);
3143 /* Set all atom-pair exclusions from the topology stored in excl
3144 * as masks in the pair-list for simple list i-entry nbl_ci
3146 static void set_ci_top_excls(const nbnxn_search_t nbs,
3147 nbnxn_pairlist_t *nbl,
3148 gmx_bool diagRemoved,
3151 const nbnxn_ci_t *nbl_ci,
3152 const t_blocka *excl)
3156 int cj_ind_first, cj_ind_last;
3157 int cj_first, cj_last;
3159 int i, ai, aj, si, eind, ge, se;
3160 int found, cj_ind_0, cj_ind_1, cj_ind_m;
3164 nbnxn_excl_t *nbl_excl;
3165 int inner_i, inner_e;
3169 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
3177 cj_ind_first = nbl_ci->cj_ind_start;
3178 cj_ind_last = nbl->ncj - 1;
3180 cj_first = nbl->cj[cj_ind_first].cj;
3181 cj_last = nbl->cj[cj_ind_last].cj;
3183 /* Determine how many contiguous j-cells we have starting
3184 * from the first i-cell. This number can be used to directly
3185 * calculate j-cell indices for excluded atoms.
3188 if (na_ci_2log == na_cj_2log)
3190 while (cj_ind_first + ndirect <= cj_ind_last &&
3191 nbl->cj[cj_ind_first+ndirect].cj == ci + ndirect)
3196 #ifdef NBNXN_SEARCH_BB_SIMD4
3199 while (cj_ind_first + ndirect <= cj_ind_last &&
3200 nbl->cj[cj_ind_first+ndirect].cj == ci_to_cj(na_cj_2log, ci) + ndirect)
3207 /* Loop over the atoms in the i super-cell */
3208 for (i = 0; i < nbl->na_sc; i++)
3210 ai = nbs->a[ci*nbl->na_sc+i];
3213 si = (i>>na_ci_2log);
3215 /* Loop over the topology-based exclusions for this i-atom */
3216 for (eind = excl->index[ai]; eind < excl->index[ai+1]; eind++)
3222 /* The self exclusion are already set, save some time */
3228 /* Without shifts we only calculate interactions j>i
3229 * for one-way pair-lists.
3231 if (diagRemoved && ge <= ci*nbl->na_sc + i)
3236 se = (ge >> na_cj_2log);
3238 /* Could the cluster se be in our list? */
3239 if (se >= cj_first && se <= cj_last)
3241 if (se < cj_first + ndirect)
3243 /* We can calculate cj_ind directly from se */
3244 found = cj_ind_first + se - cj_first;
3248 /* Search for se using bisection */
3250 cj_ind_0 = cj_ind_first + ndirect;
3251 cj_ind_1 = cj_ind_last + 1;
3252 while (found == -1 && cj_ind_0 < cj_ind_1)
3254 cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
3256 cj_m = nbl->cj[cj_ind_m].cj;
3264 cj_ind_1 = cj_ind_m;
3268 cj_ind_0 = cj_ind_m + 1;
3275 inner_i = i - (si << na_ci_2log);
3276 inner_e = ge - (se << na_cj_2log);
3278 nbl->cj[found].excl &= ~(1U<<((inner_i<<na_cj_2log) + inner_e));
3286 /* Add a new i-entry to the FEP list and copy the i-properties */
3287 static gmx_inline void fep_list_new_nri_copy(t_nblist *nlist)
3289 /* Add a new i-entry */
3292 assert(nlist->nri < nlist->maxnri);
3294 /* Duplicate the last i-entry, except for jindex, which continues */
3295 nlist->iinr[nlist->nri] = nlist->iinr[nlist->nri-1];
3296 nlist->shift[nlist->nri] = nlist->shift[nlist->nri-1];
3297 nlist->gid[nlist->nri] = nlist->gid[nlist->nri-1];
3298 nlist->jindex[nlist->nri] = nlist->nrj;
3301 /* For load balancing of the free-energy lists over threads, we set
3302 * the maximum nrj size of an i-entry to 40. This leads to good
3303 * load balancing in the worst case scenario of a single perturbed
3304 * particle on 16 threads, while not introducing significant overhead.
3305 * Note that half of the perturbed pairs will anyhow end up in very small lists,
3306 * since non perturbed i-particles will see few perturbed j-particles).
3308 const int max_nrj_fep = 40;
3310 /* Exclude the perturbed pairs from the Verlet list. This is only done to avoid
3311 * singularities for overlapping particles (0/0), since the charges and
3312 * LJ parameters have been zeroed in the nbnxn data structure.
3313 * Simultaneously make a group pair list for the perturbed pairs.
3315 static void make_fep_list(const nbnxn_search_t nbs,
3316 const nbnxn_atomdata_t *nbat,
3317 nbnxn_pairlist_t *nbl,
3318 gmx_bool bDiagRemoved,
3320 const nbnxn_grid_t *gridi,
3321 const nbnxn_grid_t *gridj,
3324 int ci, cj_ind_start, cj_ind_end, cj_ind, cja, cjr;
3326 int ngid, gid_i = 0, gid_j, gid;
3327 int egp_shift, egp_mask;
3329 int i, j, ind_i, ind_j, ai, aj;
3331 gmx_bool bFEP_i, bFEP_i_all;
3333 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
3341 cj_ind_start = nbl_ci->cj_ind_start;
3342 cj_ind_end = nbl_ci->cj_ind_end;
3344 /* In worst case we have alternating energy groups
3345 * and create #atom-pair lists, which means we need the size
3346 * of a cluster pair (na_ci*na_cj) times the number of cj's.
3348 nri_max = nbl->na_ci*nbl->na_cj*(cj_ind_end - cj_ind_start);
3349 if (nlist->nri + nri_max > nlist->maxnri)
3351 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
3352 reallocate_nblist(nlist);
3355 ngid = nbat->nenergrp;
3357 if (ngid*gridj->na_cj > sizeof(gid_cj)*8)
3359 gmx_fatal(FARGS, "The Verlet scheme with %dx%d kernels and free-energy only supports up to %d energy groups",
3360 gridi->na_c, gridj->na_cj, (sizeof(gid_cj)*8)/gridj->na_cj);
3363 egp_shift = nbat->neg_2log;
3364 egp_mask = (1<<nbat->neg_2log) - 1;
3366 /* Loop over the atoms in the i sub-cell */
3368 for (i = 0; i < nbl->na_ci; i++)
3370 ind_i = ci*nbl->na_ci + i;
3375 nlist->jindex[nri+1] = nlist->jindex[nri];
3376 nlist->iinr[nri] = ai;
3377 /* The actual energy group pair index is set later */
3378 nlist->gid[nri] = 0;
3379 nlist->shift[nri] = nbl_ci->shift & NBNXN_CI_SHIFT;
3381 bFEP_i = gridi->fep[ci - gridi->cell0] & (1 << i);
3383 bFEP_i_all = bFEP_i_all && bFEP_i;
3385 if ((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj > nlist->maxnrj)
3387 nlist->maxnrj = over_alloc_small((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj);
3388 srenew(nlist->jjnr, nlist->maxnrj);
3389 srenew(nlist->excl_fep, nlist->maxnrj);
3394 gid_i = (nbat->energrp[ci] >> (egp_shift*i)) & egp_mask;
3397 for (cj_ind = cj_ind_start; cj_ind < cj_ind_end; cj_ind++)
3399 unsigned int fep_cj;
3401 cja = nbl->cj[cj_ind].cj;
3403 if (gridj->na_cj == gridj->na_c)
3405 cjr = cja - gridj->cell0;
3406 fep_cj = gridj->fep[cjr];
3409 gid_cj = nbat->energrp[cja];
3412 else if (2*gridj->na_cj == gridj->na_c)
3414 cjr = cja - gridj->cell0*2;
3415 /* Extract half of the ci fep/energrp mask */
3416 fep_cj = (gridj->fep[cjr>>1] >> ((cjr&1)*gridj->na_cj)) & ((1<<gridj->na_cj) - 1);
3419 gid_cj = nbat->energrp[cja>>1] >> ((cja&1)*gridj->na_cj*egp_shift) & ((1<<(gridj->na_cj*egp_shift)) - 1);
3424 cjr = cja - (gridj->cell0>>1);
3425 /* Combine two ci fep masks/energrp */
3426 fep_cj = gridj->fep[cjr*2] + (gridj->fep[cjr*2+1] << gridj->na_c);
3429 gid_cj = nbat->energrp[cja*2] + (nbat->energrp[cja*2+1] << (gridj->na_c*egp_shift));
3433 if (bFEP_i || fep_cj != 0)
3435 for (j = 0; j < nbl->na_cj; j++)
3437 /* Is this interaction perturbed and not excluded? */
3438 ind_j = cja*nbl->na_cj + j;
3441 (bFEP_i || (fep_cj & (1 << j))) &&
3442 (!bDiagRemoved || ind_j >= ind_i))
3446 gid_j = (gid_cj >> (j*egp_shift)) & egp_mask;
3447 gid = GID(gid_i, gid_j, ngid);
3449 if (nlist->nrj > nlist->jindex[nri] &&
3450 nlist->gid[nri] != gid)
3452 /* Energy group pair changed: new list */
3453 fep_list_new_nri_copy(nlist);
3456 nlist->gid[nri] = gid;
3459 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
3461 fep_list_new_nri_copy(nlist);
3465 /* Add it to the FEP list */
3466 nlist->jjnr[nlist->nrj] = aj;
3467 nlist->excl_fep[nlist->nrj] = (nbl->cj[cj_ind].excl >> (i*nbl->na_cj + j)) & 1;
3470 /* Exclude it from the normal list.
3471 * Note that the charge has been set to zero,
3472 * but we need to avoid 0/0, as perturbed atoms
3473 * can be on top of each other.
3475 nbl->cj[cj_ind].excl &= ~(1U << (i*nbl->na_cj + j));
3481 if (nlist->nrj > nlist->jindex[nri])
3483 /* Actually add this new, non-empty, list */
3485 nlist->jindex[nlist->nri] = nlist->nrj;
3492 /* All interactions are perturbed, we can skip this entry */
3493 nbl_ci->cj_ind_end = cj_ind_start;
3497 /* Return the index of atom a within a cluster */
3498 static gmx_inline int cj_mod_cj4(int cj)
3500 return cj & (NBNXN_GPU_JGROUP_SIZE - 1);
3503 /* Convert a j-cluster to a cj4 group */
3504 static gmx_inline int cj_to_cj4(int cj)
3506 return cj >> NBNXN_GPU_JGROUP_SIZE_2LOG;
3509 /* Return the index of an j-atom within a warp */
3510 static gmx_inline int a_mod_wj(int a)
3512 return a & (NBNXN_GPU_CLUSTER_SIZE/2 - 1);
3515 /* As make_fep_list above, but for super/sub lists. */
3516 static void make_fep_list_supersub(const nbnxn_search_t nbs,
3517 const nbnxn_atomdata_t *nbat,
3518 nbnxn_pairlist_t *nbl,
3519 gmx_bool bDiagRemoved,
3520 const nbnxn_sci_t *nbl_sci,
3525 const nbnxn_grid_t *gridi,
3526 const nbnxn_grid_t *gridj,
3529 int sci, cj4_ind_start, cj4_ind_end, cj4_ind, gcj, cjr;
3532 int i, j, ind_i, ind_j, ai, aj;
3536 const nbnxn_cj4_t *cj4;
3538 if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
3546 cj4_ind_start = nbl_sci->cj4_ind_start;
3547 cj4_ind_end = nbl_sci->cj4_ind_end;
3549 /* Here we process one super-cell, max #atoms na_sc, versus a list
3550 * cj4 entries, each with max NBNXN_GPU_JGROUP_SIZE cj's, each
3551 * of size na_cj atoms.
3552 * On the GPU we don't support energy groups (yet).
3553 * So for each of the na_sc i-atoms, we need max one FEP list
3554 * for each max_nrj_fep j-atoms.
3556 nri_max = nbl->na_sc*nbl->na_cj*(1 + ((cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE)/max_nrj_fep);
3557 if (nlist->nri + nri_max > nlist->maxnri)
3559 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
3560 reallocate_nblist(nlist);
3563 /* Loop over the atoms in the i super-cluster */
3564 for (c = 0; c < GPU_NSUBCELL; c++)
3566 c_abs = sci*GPU_NSUBCELL + c;
3568 for (i = 0; i < nbl->na_ci; i++)
3570 ind_i = c_abs*nbl->na_ci + i;
3575 nlist->jindex[nri+1] = nlist->jindex[nri];
3576 nlist->iinr[nri] = ai;
3577 /* With GPUs, energy groups are not supported */
3578 nlist->gid[nri] = 0;
3579 nlist->shift[nri] = nbl_sci->shift & NBNXN_CI_SHIFT;
3581 bFEP_i = (gridi->fep[c_abs - gridi->cell0*GPU_NSUBCELL] & (1 << i));
3583 xi = nbat->x[ind_i*nbat->xstride+XX] + shx;
3584 yi = nbat->x[ind_i*nbat->xstride+YY] + shy;
3585 zi = nbat->x[ind_i*nbat->xstride+ZZ] + shz;
3587 if ((nlist->nrj + cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE*nbl->na_cj > nlist->maxnrj)
3589 nlist->maxnrj = over_alloc_small((nlist->nrj + cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE*nbl->na_cj);
3590 srenew(nlist->jjnr, nlist->maxnrj);
3591 srenew(nlist->excl_fep, nlist->maxnrj);
3594 for (cj4_ind = cj4_ind_start; cj4_ind < cj4_ind_end; cj4_ind++)
3596 cj4 = &nbl->cj4[cj4_ind];
3598 for (gcj = 0; gcj < NBNXN_GPU_JGROUP_SIZE; gcj++)
3600 unsigned int fep_cj;
3602 if ((cj4->imei[0].imask & (1U << (gcj*GPU_NSUBCELL + c))) == 0)
3604 /* Skip this ci for this cj */
3608 cjr = cj4->cj[gcj] - gridj->cell0*GPU_NSUBCELL;
3610 fep_cj = gridj->fep[cjr];
3612 if (bFEP_i || fep_cj != 0)
3614 for (j = 0; j < nbl->na_cj; j++)
3616 /* Is this interaction perturbed and not excluded? */
3617 ind_j = (gridj->cell0*GPU_NSUBCELL + cjr)*nbl->na_cj + j;
3620 (bFEP_i || (fep_cj & (1 << j))) &&
3621 (!bDiagRemoved || ind_j >= ind_i))
3625 unsigned int excl_bit;
3628 get_nbl_exclusions_1(nbl, cj4_ind, j>>2, &excl);
3630 excl_pair = a_mod_wj(j)*nbl->na_ci + i;
3631 excl_bit = (1U << (gcj*GPU_NSUBCELL + c));
3633 dx = nbat->x[ind_j*nbat->xstride+XX] - xi;
3634 dy = nbat->x[ind_j*nbat->xstride+YY] - yi;
3635 dz = nbat->x[ind_j*nbat->xstride+ZZ] - zi;
3637 /* The unpruned GPU list has more than 2/3
3638 * of the atom pairs beyond rlist. Using
3639 * this list will cause a lot of overhead
3640 * in the CPU FEP kernels, especially
3641 * relative to the fast GPU kernels.
3642 * So we prune the FEP list here.
3644 if (dx*dx + dy*dy + dz*dz < rlist_fep2)
3646 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
3648 fep_list_new_nri_copy(nlist);
3652 /* Add it to the FEP list */
3653 nlist->jjnr[nlist->nrj] = aj;
3654 nlist->excl_fep[nlist->nrj] = (excl->pair[excl_pair] & excl_bit) ? 1 : 0;
3658 /* Exclude it from the normal list.
3659 * Note that the charge and LJ parameters have
3660 * been set to zero, but we need to avoid 0/0,
3661 * as perturbed atoms can be on top of each other.
3663 excl->pair[excl_pair] &= ~excl_bit;
3667 /* Note that we could mask out this pair in imask
3668 * if all i- and/or all j-particles are perturbed.
3669 * But since the perturbed pairs on the CPU will
3670 * take an order of magnitude more time, the GPU
3671 * will finish before the CPU and there is no gain.
3677 if (nlist->nrj > nlist->jindex[nri])
3679 /* Actually add this new, non-empty, list */
3681 nlist->jindex[nlist->nri] = nlist->nrj;
3688 /* Set all atom-pair exclusions from the topology stored in excl
3689 * as masks in the pair-list for i-super-cell entry nbl_sci
3691 static void set_sci_top_excls(const nbnxn_search_t nbs,
3692 nbnxn_pairlist_t *nbl,
3693 gmx_bool diagRemoved,
3695 const nbnxn_sci_t *nbl_sci,
3696 const t_blocka *excl)
3701 int cj_ind_first, cj_ind_last;
3702 int cj_first, cj_last;
3704 int i, ai, aj, si, eind, ge, se;
3705 int found, cj_ind_0, cj_ind_1, cj_ind_m;
3709 nbnxn_excl_t *nbl_excl;
3710 int inner_i, inner_e, w;
3716 if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
3724 cj_ind_first = nbl_sci->cj4_ind_start*NBNXN_GPU_JGROUP_SIZE;
3725 cj_ind_last = nbl->work->cj_ind - 1;
3727 cj_first = nbl->cj4[nbl_sci->cj4_ind_start].cj[0];
3728 cj_last = nbl_cj(nbl, cj_ind_last);
3730 /* Determine how many contiguous j-clusters we have starting
3731 * from the first i-cluster. This number can be used to directly
3732 * calculate j-cluster indices for excluded atoms.
3735 while (cj_ind_first + ndirect <= cj_ind_last &&
3736 nbl_cj(nbl, cj_ind_first+ndirect) == sci*GPU_NSUBCELL + ndirect)
3741 /* Loop over the atoms in the i super-cell */
3742 for (i = 0; i < nbl->na_sc; i++)
3744 ai = nbs->a[sci*nbl->na_sc+i];
3747 si = (i>>na_c_2log);
3749 /* Loop over the topology-based exclusions for this i-atom */
3750 for (eind = excl->index[ai]; eind < excl->index[ai+1]; eind++)
3756 /* The self exclusion are already set, save some time */
3762 /* Without shifts we only calculate interactions j>i
3763 * for one-way pair-lists.
3765 if (diagRemoved && ge <= sci*nbl->na_sc + i)
3771 /* Could the cluster se be in our list? */
3772 if (se >= cj_first && se <= cj_last)
3774 if (se < cj_first + ndirect)
3776 /* We can calculate cj_ind directly from se */
3777 found = cj_ind_first + se - cj_first;
3781 /* Search for se using bisection */
3783 cj_ind_0 = cj_ind_first + ndirect;
3784 cj_ind_1 = cj_ind_last + 1;
3785 while (found == -1 && cj_ind_0 < cj_ind_1)
3787 cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
3789 cj_m = nbl_cj(nbl, cj_ind_m);
3797 cj_ind_1 = cj_ind_m;
3801 cj_ind_0 = cj_ind_m + 1;
3808 inner_i = i - si*na_c;
3809 inner_e = ge - se*na_c;
3811 if (nbl_imask0(nbl, found) & (1U << (cj_mod_cj4(found)*GPU_NSUBCELL + si)))
3815 get_nbl_exclusions_1(nbl, cj_to_cj4(found), w, &nbl_excl);
3817 nbl_excl->pair[a_mod_wj(inner_e)*nbl->na_ci+inner_i] &=
3818 ~(1U << (cj_mod_cj4(found)*GPU_NSUBCELL + si));
3827 /* Reallocate the simple ci list for at least n entries */
3828 static void nb_realloc_ci(nbnxn_pairlist_t *nbl, int n)
3830 nbl->ci_nalloc = over_alloc_small(n);
3831 nbnxn_realloc_void((void **)&nbl->ci,
3832 nbl->nci*sizeof(*nbl->ci),
3833 nbl->ci_nalloc*sizeof(*nbl->ci),
3834 nbl->alloc, nbl->free);
3837 /* Reallocate the super-cell sci list for at least n entries */
3838 static void nb_realloc_sci(nbnxn_pairlist_t *nbl, int n)
3840 nbl->sci_nalloc = over_alloc_small(n);
3841 nbnxn_realloc_void((void **)&nbl->sci,
3842 nbl->nsci*sizeof(*nbl->sci),
3843 nbl->sci_nalloc*sizeof(*nbl->sci),
3844 nbl->alloc, nbl->free);
3847 /* Make a new ci entry at index nbl->nci */
3848 static void new_ci_entry(nbnxn_pairlist_t *nbl, int ci, int shift, int flags)
3850 if (nbl->nci + 1 > nbl->ci_nalloc)
3852 nb_realloc_ci(nbl, nbl->nci+1);
3854 nbl->ci[nbl->nci].ci = ci;
3855 nbl->ci[nbl->nci].shift = shift;
3856 /* Store the interaction flags along with the shift */
3857 nbl->ci[nbl->nci].shift |= flags;
3858 nbl->ci[nbl->nci].cj_ind_start = nbl->ncj;
3859 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
3862 /* Make a new sci entry at index nbl->nsci */
3863 static void new_sci_entry(nbnxn_pairlist_t *nbl, int sci, int shift)
3865 if (nbl->nsci + 1 > nbl->sci_nalloc)
3867 nb_realloc_sci(nbl, nbl->nsci+1);
3869 nbl->sci[nbl->nsci].sci = sci;
3870 nbl->sci[nbl->nsci].shift = shift;
3871 nbl->sci[nbl->nsci].cj4_ind_start = nbl->ncj4;
3872 nbl->sci[nbl->nsci].cj4_ind_end = nbl->ncj4;
3875 /* Sort the simple j-list cj on exclusions.
3876 * Entries with exclusions will all be sorted to the beginning of the list.
3878 static void sort_cj_excl(nbnxn_cj_t *cj, int ncj,
3879 nbnxn_list_work_t *work)
3883 if (ncj > work->cj_nalloc)
3885 work->cj_nalloc = over_alloc_large(ncj);
3886 srenew(work->cj, work->cj_nalloc);
3889 /* Make a list of the j-cells involving exclusions */
3891 for (j = 0; j < ncj; j++)
3893 if (cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
3895 work->cj[jnew++] = cj[j];
3898 /* Check if there are exclusions at all or not just the first entry */
3899 if (!((jnew == 0) ||
3900 (jnew == 1 && cj[0].excl != NBNXN_INTERACTION_MASK_ALL)))
3902 for (j = 0; j < ncj; j++)
3904 if (cj[j].excl == NBNXN_INTERACTION_MASK_ALL)
3906 work->cj[jnew++] = cj[j];
3909 for (j = 0; j < ncj; j++)
3911 cj[j] = work->cj[j];
3916 /* Close this simple list i entry */
3917 static void close_ci_entry_simple(nbnxn_pairlist_t *nbl)
3921 /* All content of the new ci entry have already been filled correctly,
3922 * we only need to increase the count here (for non empty lists).
3924 jlen = nbl->ci[nbl->nci].cj_ind_end - nbl->ci[nbl->nci].cj_ind_start;
3927 sort_cj_excl(nbl->cj+nbl->ci[nbl->nci].cj_ind_start, jlen, nbl->work);
3929 /* The counts below are used for non-bonded pair/flop counts
3930 * and should therefore match the available kernel setups.
3932 if (!(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_COUL(0)))
3934 nbl->work->ncj_noq += jlen;
3936 else if ((nbl->ci[nbl->nci].shift & NBNXN_CI_HALF_LJ(0)) ||
3937 !(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_LJ(0)))
3939 nbl->work->ncj_hlj += jlen;
3946 /* Split sci entry for load balancing on the GPU.
3947 * Splitting ensures we have enough lists to fully utilize the whole GPU.
3948 * With progBal we generate progressively smaller lists, which improves
3949 * load balancing. As we only know the current count on our own thread,
3950 * we will need to estimate the current total amount of i-entries.
3951 * As the lists get concatenated later, this estimate depends
3952 * both on nthread and our own thread index.
3954 static void split_sci_entry(nbnxn_pairlist_t *nbl,
3956 gmx_bool progBal, int nsp_tot_est,
3957 int thread, int nthread)
3961 int cj4_start, cj4_end, j4len, cj4;
3963 int nsp, nsp_sci, nsp_cj4, nsp_cj4_e, nsp_cj4_p;
3968 /* Estimate the total numbers of ci's of the nblist combined
3969 * over all threads using the target number of ci's.
3971 nsp_est = (nsp_tot_est*thread)/nthread + nbl->nci_tot;
3973 /* The first ci blocks should be larger, to avoid overhead.
3974 * The last ci blocks should be smaller, to improve load balancing.
3975 * The factor 3/2 makes the first block 3/2 times the target average
3976 * and ensures that the total number of blocks end up equal to
3977 * that with of equally sized blocks of size nsp_target_av.
3979 nsp_max = nsp_target_av*nsp_tot_est*3/(2*(nsp_est + nsp_tot_est));
3983 nsp_max = nsp_target_av;
3986 /* Since nsp_max is a maximum/cut-off (this avoids high outliers,
3987 * which lead to load imbalance), not an average, we add half the
3988 * number of pairs in a cj4 block to get the average about right.
3990 nsp_max += GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE/2;
3992 cj4_start = nbl->sci[nbl->nsci-1].cj4_ind_start;
3993 cj4_end = nbl->sci[nbl->nsci-1].cj4_ind_end;
3994 j4len = cj4_end - cj4_start;
3996 if (j4len > 1 && j4len*GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE > nsp_max)
3998 /* Remove the last ci entry and process the cj4's again */
4006 for (cj4 = cj4_start; cj4 < cj4_end; cj4++)
4008 nsp_cj4_p = nsp_cj4;
4009 /* Count the number of cluster pairs in this cj4 group */
4011 for (p = 0; p < GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE; p++)
4013 nsp_cj4 += (nbl->cj4[cj4].imei[0].imask >> p) & 1;
4016 if (nsp_cj4 > 0 && nsp + nsp_cj4 > nsp_max)
4018 /* Split the list at cj4 */
4019 nbl->sci[sci].cj4_ind_end = cj4;
4020 /* Create a new sci entry */
4023 if (nbl->nsci+1 > nbl->sci_nalloc)
4025 nb_realloc_sci(nbl, nbl->nsci+1);
4027 nbl->sci[sci].sci = nbl->sci[nbl->nsci-1].sci;
4028 nbl->sci[sci].shift = nbl->sci[nbl->nsci-1].shift;
4029 nbl->sci[sci].cj4_ind_start = cj4;
4031 nsp_cj4_e = nsp_cj4_p;
4037 /* Put the remaining cj4's in the last sci entry */
4038 nbl->sci[sci].cj4_ind_end = cj4_end;
4040 /* Possibly balance out the last two sci's
4041 * by moving the last cj4 of the second last sci.
4043 if (nsp_sci - nsp_cj4_e >= nsp + nsp_cj4_e)
4045 nbl->sci[sci-1].cj4_ind_end--;
4046 nbl->sci[sci].cj4_ind_start--;
4053 /* Clost this super/sub list i entry */
4054 static void close_ci_entry_supersub(nbnxn_pairlist_t *nbl,
4056 gmx_bool progBal, int nsp_tot_est,
4057 int thread, int nthread)
4062 /* All content of the new ci entry have already been filled correctly,
4063 * we only need to increase the count here (for non empty lists).
4065 j4len = nbl->sci[nbl->nsci].cj4_ind_end - nbl->sci[nbl->nsci].cj4_ind_start;
4068 /* We can only have complete blocks of 4 j-entries in a list,
4069 * so round the count up before closing.
4071 nbl->ncj4 = ((nbl->work->cj_ind + NBNXN_GPU_JGROUP_SIZE - 1) >> NBNXN_GPU_JGROUP_SIZE_2LOG);
4072 nbl->work->cj_ind = nbl->ncj4*NBNXN_GPU_JGROUP_SIZE;
4078 /* Measure the size of the new entry and potentially split it */
4079 split_sci_entry(nbl, nsp_max_av, progBal, nsp_tot_est,
4085 /* Syncs the working array before adding another grid pair to the list */
4086 static void sync_work(nbnxn_pairlist_t *nbl)
4090 nbl->work->cj_ind = nbl->ncj4*NBNXN_GPU_JGROUP_SIZE;
4091 nbl->work->cj4_init = nbl->ncj4;
4095 /* Clears an nbnxn_pairlist_t data structure */
4096 static void clear_pairlist(nbnxn_pairlist_t *nbl)
4105 nbl->work->ncj_noq = 0;
4106 nbl->work->ncj_hlj = 0;
4109 /* Clears a group scheme pair list */
4110 static void clear_pairlist_fep(t_nblist *nl)
4114 if (nl->jindex == NULL)
4116 snew(nl->jindex, 1);
4121 /* Sets a simple list i-cell bounding box, including PBC shift */
4122 static gmx_inline void set_icell_bb_simple(const nbnxn_bb_t *bb, int ci,
4123 real shx, real shy, real shz,
4126 bb_ci->lower[BB_X] = bb[ci].lower[BB_X] + shx;
4127 bb_ci->lower[BB_Y] = bb[ci].lower[BB_Y] + shy;
4128 bb_ci->lower[BB_Z] = bb[ci].lower[BB_Z] + shz;
4129 bb_ci->upper[BB_X] = bb[ci].upper[BB_X] + shx;
4130 bb_ci->upper[BB_Y] = bb[ci].upper[BB_Y] + shy;
4131 bb_ci->upper[BB_Z] = bb[ci].upper[BB_Z] + shz;
4135 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
4136 static void set_icell_bbxxxx_supersub(const float *bb, int ci,
4137 real shx, real shy, real shz,
4142 ia = ci*(GPU_NSUBCELL>>STRIDE_PBB_2LOG)*NNBSBB_XXXX;
4143 for (m = 0; m < (GPU_NSUBCELL>>STRIDE_PBB_2LOG)*NNBSBB_XXXX; m += NNBSBB_XXXX)
4145 for (i = 0; i < STRIDE_PBB; i++)
4147 bb_ci[m+0*STRIDE_PBB+i] = bb[ia+m+0*STRIDE_PBB+i] + shx;
4148 bb_ci[m+1*STRIDE_PBB+i] = bb[ia+m+1*STRIDE_PBB+i] + shy;
4149 bb_ci[m+2*STRIDE_PBB+i] = bb[ia+m+2*STRIDE_PBB+i] + shz;
4150 bb_ci[m+3*STRIDE_PBB+i] = bb[ia+m+3*STRIDE_PBB+i] + shx;
4151 bb_ci[m+4*STRIDE_PBB+i] = bb[ia+m+4*STRIDE_PBB+i] + shy;
4152 bb_ci[m+5*STRIDE_PBB+i] = bb[ia+m+5*STRIDE_PBB+i] + shz;
4158 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
4159 static void set_icell_bb_supersub(const nbnxn_bb_t *bb, int ci,
4160 real shx, real shy, real shz,
4165 for (i = 0; i < GPU_NSUBCELL; i++)
4167 set_icell_bb_simple(bb, ci*GPU_NSUBCELL+i,
4173 /* Copies PBC shifted i-cell atom coordinates x,y,z to working array */
4174 static void icell_set_x_simple(int ci,
4175 real shx, real shy, real shz,
4176 int gmx_unused na_c,
4177 int stride, const real *x,
4178 nbnxn_list_work_t *work)
4182 ia = ci*NBNXN_CPU_CLUSTER_I_SIZE;
4184 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE; i++)
4186 work->x_ci[i*STRIDE_XYZ+XX] = x[(ia+i)*stride+XX] + shx;
4187 work->x_ci[i*STRIDE_XYZ+YY] = x[(ia+i)*stride+YY] + shy;
4188 work->x_ci[i*STRIDE_XYZ+ZZ] = x[(ia+i)*stride+ZZ] + shz;
4192 /* Copies PBC shifted super-cell atom coordinates x,y,z to working array */
4193 static void icell_set_x_supersub(int ci,
4194 real shx, real shy, real shz,
4196 int stride, const real *x,
4197 nbnxn_list_work_t *work)
4204 ia = ci*GPU_NSUBCELL*na_c;
4205 for (i = 0; i < GPU_NSUBCELL*na_c; i++)
4207 x_ci[i*DIM + XX] = x[(ia+i)*stride + XX] + shx;
4208 x_ci[i*DIM + YY] = x[(ia+i)*stride + YY] + shy;
4209 x_ci[i*DIM + ZZ] = x[(ia+i)*stride + ZZ] + shz;
4213 #ifdef NBNXN_SEARCH_BB_SIMD4
4214 /* Copies PBC shifted super-cell packed atom coordinates to working array */
4215 static void icell_set_x_supersub_simd4(int ci,
4216 real shx, real shy, real shz,
4218 int stride, const real *x,
4219 nbnxn_list_work_t *work)
4221 int si, io, ia, i, j;
4226 for (si = 0; si < GPU_NSUBCELL; si++)
4228 for (i = 0; i < na_c; i += STRIDE_PBB)
4231 ia = ci*GPU_NSUBCELL*na_c + io;
4232 for (j = 0; j < STRIDE_PBB; j++)
4234 x_ci[io*DIM + j + XX*STRIDE_PBB] = x[(ia+j)*stride+XX] + shx;
4235 x_ci[io*DIM + j + YY*STRIDE_PBB] = x[(ia+j)*stride+YY] + shy;
4236 x_ci[io*DIM + j + ZZ*STRIDE_PBB] = x[(ia+j)*stride+ZZ] + shz;
4243 static real minimum_subgrid_size_xy(const nbnxn_grid_t *grid)
4247 return min(grid->sx, grid->sy);
4251 return min(grid->sx/GPU_NSUBCELL_X, grid->sy/GPU_NSUBCELL_Y);
4255 static real effective_buffer_1x1_vs_MxN(const nbnxn_grid_t *gridi,
4256 const nbnxn_grid_t *gridj)
4258 const real eff_1x1_buffer_fac_overest = 0.1;
4260 /* Determine an atom-pair list cut-off buffer size for atom pairs,
4261 * to be added to rlist (including buffer) used for MxN.
4262 * This is for converting an MxN list to a 1x1 list. This means we can't
4263 * use the normal buffer estimate, as we have an MxN list in which
4264 * some atom pairs beyond rlist are missing. We want to capture
4265 * the beneficial effect of buffering by extra pairs just outside rlist,
4266 * while removing the useless pairs that are further away from rlist.
4267 * (Also the buffer could have been set manually not using the estimate.)
4268 * This buffer size is an overestimate.
4269 * We add 10% of the smallest grid sub-cell dimensions.
4270 * Note that the z-size differs per cell and we don't use this,
4271 * so we overestimate.
4272 * With PME, the 10% value gives a buffer that is somewhat larger
4273 * than the effective buffer with a tolerance of 0.005 kJ/mol/ps.
4274 * Smaller tolerances or using RF lead to a smaller effective buffer,
4275 * so 10% gives a safe overestimate.
4277 return eff_1x1_buffer_fac_overest*(minimum_subgrid_size_xy(gridi) +
4278 minimum_subgrid_size_xy(gridj));
4281 /* Clusters at the cut-off only increase rlist by 60% of their size */
4282 static real nbnxn_rlist_inc_outside_fac = 0.6;
4284 /* Due to the cluster size the effective pair-list is longer than
4285 * that of a simple atom pair-list. This function gives the extra distance.
4287 real nbnxn_get_rlist_effective_inc(int cluster_size_j, real atom_density)
4290 real vol_inc_i, vol_inc_j;
4292 /* We should get this from the setup, but currently it's the same for
4293 * all setups, including GPUs.
4295 cluster_size_i = NBNXN_CPU_CLUSTER_I_SIZE;
4297 vol_inc_i = (cluster_size_i - 1)/atom_density;
4298 vol_inc_j = (cluster_size_j - 1)/atom_density;
4300 return nbnxn_rlist_inc_outside_fac*pow(vol_inc_i + vol_inc_j, 1.0/3.0);
4303 /* Estimates the interaction volume^2 for non-local interactions */
4304 static real nonlocal_vol2(const gmx_domdec_zones_t *zones, rvec ls, real r)
4313 /* Here we simply add up the volumes of 1, 2 or 3 1D decomposition
4314 * not home interaction volume^2. As these volumes are not additive,
4315 * this is an overestimate, but it would only be significant in the limit
4316 * of small cells, where we anyhow need to split the lists into
4317 * as small parts as possible.
4320 for (z = 0; z < zones->n; z++)
4322 if (zones->shift[z][XX] + zones->shift[z][YY] + zones->shift[z][ZZ] == 1)
4327 for (d = 0; d < DIM; d++)
4329 if (zones->shift[z][d] == 0)
4333 za *= zones->size[z].x1[d] - zones->size[z].x0[d];
4337 /* 4 octants of a sphere */
4338 vold_est = 0.25*M_PI*r*r*r*r;
4339 /* 4 quarter pie slices on the edges */
4340 vold_est += 4*cl*M_PI/6.0*r*r*r;
4341 /* One rectangular volume on a face */
4342 vold_est += ca*0.5*r*r;
4344 vol2_est_tot += vold_est*za;
4348 return vol2_est_tot;
4351 /* Estimates the average size of a full j-list for super/sub setup */
4352 static void get_nsubpair_target(const nbnxn_search_t nbs,
4355 int min_ci_balanced,
4356 int *nsubpair_target,
4357 int *nsubpair_tot_est)
4359 /* The target value of 36 seems to be the optimum for Kepler.
4360 * Maxwell is less sensitive to the exact value.
4362 const int nsubpair_target_min = 36;
4363 const nbnxn_grid_t *grid;
4365 real xy_diag2, r_eff_sup, vol_est, nsp_est, nsp_est_nl;
4368 grid = &nbs->grid[0];
4370 if (min_ci_balanced <= 0 || grid->nc >= min_ci_balanced || grid->nc == 0)
4372 /* We don't need to balance the list sizes */
4373 *nsubpair_target = 0;
4374 *nsubpair_tot_est = 0;
4379 ls[XX] = (grid->c1[XX] - grid->c0[XX])/(grid->ncx*GPU_NSUBCELL_X);
4380 ls[YY] = (grid->c1[YY] - grid->c0[YY])/(grid->ncy*GPU_NSUBCELL_Y);
4381 ls[ZZ] = (grid->c1[ZZ] - grid->c0[ZZ])*grid->ncx*grid->ncy/(grid->nc*GPU_NSUBCELL_Z);
4383 /* The average squared length of the diagonal of a sub cell */
4384 xy_diag2 = ls[XX]*ls[XX] + ls[YY]*ls[YY] + ls[ZZ]*ls[ZZ];
4386 /* The formulas below are a heuristic estimate of the average nsj per si*/
4387 r_eff_sup = rlist + nbnxn_rlist_inc_outside_fac*sqr((grid->na_c - 1.0)/grid->na_c)*sqrt(xy_diag2/3);
4389 if (!nbs->DomDec || nbs->zones->n == 1)
4396 sqr(grid->atom_density/grid->na_c)*
4397 nonlocal_vol2(nbs->zones, ls, r_eff_sup);
4402 /* Sub-cell interacts with itself */
4403 vol_est = ls[XX]*ls[YY]*ls[ZZ];
4404 /* 6/2 rectangular volume on the faces */
4405 vol_est += (ls[XX]*ls[YY] + ls[XX]*ls[ZZ] + ls[YY]*ls[ZZ])*r_eff_sup;
4406 /* 12/2 quarter pie slices on the edges */
4407 vol_est += 2*(ls[XX] + ls[YY] + ls[ZZ])*0.25*M_PI*sqr(r_eff_sup);
4408 /* 4 octants of a sphere */
4409 vol_est += 0.5*4.0/3.0*M_PI*pow(r_eff_sup, 3);
4411 nsp_est = grid->nsubc_tot*vol_est*grid->atom_density/grid->na_c;
4413 /* Subtract the non-local pair count */
4414 nsp_est -= nsp_est_nl;
4418 fprintf(debug, "nsp_est local %5.1f non-local %5.1f\n",
4419 nsp_est, nsp_est_nl);
4424 nsp_est = nsp_est_nl;
4427 /* Thus the (average) maximum j-list size should be as follows.
4428 * Since there is overhead, we shouldn't make the lists too small
4429 * (and we can't chop up j-groups) so we use a minimum target size of 36.
4431 *nsubpair_target = max(nsubpair_target_min,
4432 (int)(nsp_est/min_ci_balanced + 0.5));
4433 *nsubpair_tot_est = (int)nsp_est;
4437 fprintf(debug, "nbl nsp estimate %.1f, nsubpair_target %d\n",
4438 nsp_est, *nsubpair_target);
4442 /* Debug list print function */
4443 static void print_nblist_ci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
4447 for (i = 0; i < nbl->nci; i++)
4449 fprintf(fp, "ci %4d shift %2d ncj %3d\n",
4450 nbl->ci[i].ci, nbl->ci[i].shift,
4451 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start);
4453 for (j = nbl->ci[i].cj_ind_start; j < nbl->ci[i].cj_ind_end; j++)
4455 fprintf(fp, " cj %5d imask %x\n",
4462 /* Debug list print function */
4463 static void print_nblist_sci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
4465 int i, j4, j, ncp, si;
4467 for (i = 0; i < nbl->nsci; i++)
4469 fprintf(fp, "ci %4d shift %2d ncj4 %2d\n",
4470 nbl->sci[i].sci, nbl->sci[i].shift,
4471 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start);
4474 for (j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
4476 for (j = 0; j < NBNXN_GPU_JGROUP_SIZE; j++)
4478 fprintf(fp, " sj %5d imask %x\n",
4480 nbl->cj4[j4].imei[0].imask);
4481 for (si = 0; si < GPU_NSUBCELL; si++)
4483 if (nbl->cj4[j4].imei[0].imask & (1U << (j*GPU_NSUBCELL + si)))
4490 fprintf(fp, "ci %4d shift %2d ncj4 %2d ncp %3d\n",
4491 nbl->sci[i].sci, nbl->sci[i].shift,
4492 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start,
4497 /* Combine pair lists *nbl generated on multiple threads nblc */
4498 static void combine_nblists(int nnbl, nbnxn_pairlist_t **nbl,
4499 nbnxn_pairlist_t *nblc)
4501 int nsci, ncj4, nexcl;
4503 int nthreads gmx_unused;
4507 gmx_incons("combine_nblists does not support simple lists");
4512 nexcl = nblc->nexcl;
4513 for (i = 0; i < nnbl; i++)
4515 nsci += nbl[i]->nsci;
4516 ncj4 += nbl[i]->ncj4;
4517 nexcl += nbl[i]->nexcl;
4520 if (nsci > nblc->sci_nalloc)
4522 nb_realloc_sci(nblc, nsci);
4524 if (ncj4 > nblc->cj4_nalloc)
4526 nblc->cj4_nalloc = over_alloc_small(ncj4);
4527 nbnxn_realloc_void((void **)&nblc->cj4,
4528 nblc->ncj4*sizeof(*nblc->cj4),
4529 nblc->cj4_nalloc*sizeof(*nblc->cj4),
4530 nblc->alloc, nblc->free);
4532 if (nexcl > nblc->excl_nalloc)
4534 nblc->excl_nalloc = over_alloc_small(nexcl);
4535 nbnxn_realloc_void((void **)&nblc->excl,
4536 nblc->nexcl*sizeof(*nblc->excl),
4537 nblc->excl_nalloc*sizeof(*nblc->excl),
4538 nblc->alloc, nblc->free);
4541 /* Each thread should copy its own data to the combined arrays,
4542 * as otherwise data will go back and forth between different caches.
4544 nthreads = gmx_omp_nthreads_get(emntPairsearch);
4545 #pragma omp parallel for num_threads(nthreads) schedule(static)
4546 for (n = 0; n < nnbl; n++)
4553 const nbnxn_pairlist_t *nbli;
4555 /* Determine the offset in the combined data for our thread */
4556 sci_offset = nblc->nsci;
4557 cj4_offset = nblc->ncj4;
4558 ci_offset = nblc->nci_tot;
4559 excl_offset = nblc->nexcl;
4561 for (i = 0; i < n; i++)
4563 sci_offset += nbl[i]->nsci;
4564 cj4_offset += nbl[i]->ncj4;
4565 ci_offset += nbl[i]->nci_tot;
4566 excl_offset += nbl[i]->nexcl;
4571 for (i = 0; i < nbli->nsci; i++)
4573 nblc->sci[sci_offset+i] = nbli->sci[i];
4574 nblc->sci[sci_offset+i].cj4_ind_start += cj4_offset;
4575 nblc->sci[sci_offset+i].cj4_ind_end += cj4_offset;
4578 for (j4 = 0; j4 < nbli->ncj4; j4++)
4580 nblc->cj4[cj4_offset+j4] = nbli->cj4[j4];
4581 nblc->cj4[cj4_offset+j4].imei[0].excl_ind += excl_offset;
4582 nblc->cj4[cj4_offset+j4].imei[1].excl_ind += excl_offset;
4585 for (j4 = 0; j4 < nbli->nexcl; j4++)
4587 nblc->excl[excl_offset+j4] = nbli->excl[j4];
4591 for (n = 0; n < nnbl; n++)
4593 nblc->nsci += nbl[n]->nsci;
4594 nblc->ncj4 += nbl[n]->ncj4;
4595 nblc->nci_tot += nbl[n]->nci_tot;
4596 nblc->nexcl += nbl[n]->nexcl;
4600 static void balance_fep_lists(const nbnxn_search_t nbs,
4601 nbnxn_pairlist_set_t *nbl_lists)
4604 int nri_tot, nrj_tot, nrj_target;
4608 nnbl = nbl_lists->nnbl;
4612 /* Nothing to balance */
4616 /* Count the total i-lists and pairs */
4619 for (th = 0; th < nnbl; th++)
4621 nri_tot += nbl_lists->nbl_fep[th]->nri;
4622 nrj_tot += nbl_lists->nbl_fep[th]->nrj;
4625 nrj_target = (nrj_tot + nnbl - 1)/nnbl;
4627 assert(gmx_omp_nthreads_get(emntNonbonded) == nnbl);
4629 #pragma omp parallel for schedule(static) num_threads(nnbl)
4630 for (th = 0; th < nnbl; th++)
4634 nbl = nbs->work[th].nbl_fep;
4636 /* Note that here we allocate for the total size, instead of
4637 * a per-thread esimate (which is hard to obtain).
4639 if (nri_tot > nbl->maxnri)
4641 nbl->maxnri = over_alloc_large(nri_tot);
4642 reallocate_nblist(nbl);
4644 if (nri_tot > nbl->maxnri || nrj_tot > nbl->maxnrj)
4646 nbl->maxnrj = over_alloc_small(nrj_tot);
4647 srenew(nbl->jjnr, nbl->maxnrj);
4648 srenew(nbl->excl_fep, nbl->maxnrj);
4651 clear_pairlist_fep(nbl);
4654 /* Loop over the source lists and assign and copy i-entries */
4656 nbld = nbs->work[th_dest].nbl_fep;
4657 for (th = 0; th < nnbl; th++)
4662 nbls = nbl_lists->nbl_fep[th];
4664 for (i = 0; i < nbls->nri; i++)
4668 /* The number of pairs in this i-entry */
4669 nrj = nbls->jindex[i+1] - nbls->jindex[i];
4671 /* Decide if list th_dest is too large and we should procede
4672 * to the next destination list.
4674 if (th_dest+1 < nnbl && nbld->nrj > 0 &&
4675 nbld->nrj + nrj - nrj_target > nrj_target - nbld->nrj)
4678 nbld = nbs->work[th_dest].nbl_fep;
4681 nbld->iinr[nbld->nri] = nbls->iinr[i];
4682 nbld->gid[nbld->nri] = nbls->gid[i];
4683 nbld->shift[nbld->nri] = nbls->shift[i];
4685 for (j = nbls->jindex[i]; j < nbls->jindex[i+1]; j++)
4687 nbld->jjnr[nbld->nrj] = nbls->jjnr[j];
4688 nbld->excl_fep[nbld->nrj] = nbls->excl_fep[j];
4692 nbld->jindex[nbld->nri] = nbld->nrj;
4696 /* Swap the list pointers */
4697 for (th = 0; th < nnbl; th++)
4701 nbl_tmp = nbl_lists->nbl_fep[th];
4702 nbl_lists->nbl_fep[th] = nbs->work[th].nbl_fep;
4703 nbs->work[th].nbl_fep = nbl_tmp;
4707 fprintf(debug, "nbl_fep[%d] nri %4d nrj %4d\n",
4709 nbl_lists->nbl_fep[th]->nri,
4710 nbl_lists->nbl_fep[th]->nrj);
4715 /* Returns the next ci to be processes by our thread */
4716 static gmx_bool next_ci(const nbnxn_grid_t *grid,
4718 int nth, int ci_block,
4719 int *ci_x, int *ci_y,
4725 if (*ci_b == ci_block)
4727 /* Jump to the next block assigned to this task */
4728 *ci += (nth - 1)*ci_block;
4732 if (*ci >= grid->nc*conv)
4737 while (*ci >= grid->cxy_ind[*ci_x*grid->ncy + *ci_y + 1]*conv)
4740 if (*ci_y == grid->ncy)
4750 /* Returns the distance^2 for which we put cell pairs in the list
4751 * without checking atom pair distances. This is usually < rlist^2.
4753 static float boundingbox_only_distance2(const nbnxn_grid_t *gridi,
4754 const nbnxn_grid_t *gridj,
4758 /* If the distance between two sub-cell bounding boxes is less
4759 * than this distance, do not check the distance between
4760 * all particle pairs in the sub-cell, since then it is likely
4761 * that the box pair has atom pairs within the cut-off.
4762 * We use the nblist cut-off minus 0.5 times the average x/y diagonal
4763 * spacing of the sub-cells. Around 40% of the checked pairs are pruned.
4764 * Using more than 0.5 gains at most 0.5%.
4765 * If forces are calculated more than twice, the performance gain
4766 * in the force calculation outweighs the cost of checking.
4767 * Note that with subcell lists, the atom-pair distance check
4768 * is only performed when only 1 out of 8 sub-cells in within range,
4769 * this is because the GPU is much faster than the cpu.
4774 bbx = 0.5*(gridi->sx + gridj->sx);
4775 bby = 0.5*(gridi->sy + gridj->sy);
4778 bbx /= GPU_NSUBCELL_X;
4779 bby /= GPU_NSUBCELL_Y;
4782 rbb2 = sqr(max(0, rlist - 0.5*sqrt(bbx*bbx + bby*bby)));
4787 return (float)((1+GMX_FLOAT_EPS)*rbb2);
4791 static int get_ci_block_size(const nbnxn_grid_t *gridi,
4792 gmx_bool bDomDec, int nth)
4794 const int ci_block_enum = 5;
4795 const int ci_block_denom = 11;
4796 const int ci_block_min_atoms = 16;
4799 /* Here we decide how to distribute the blocks over the threads.
4800 * We use prime numbers to try to avoid that the grid size becomes
4801 * a multiple of the number of threads, which would lead to some
4802 * threads getting "inner" pairs and others getting boundary pairs,
4803 * which in turns will lead to load imbalance between threads.
4804 * Set the block size as 5/11/ntask times the average number of cells
4805 * in a y,z slab. This should ensure a quite uniform distribution
4806 * of the grid parts of the different thread along all three grid
4807 * zone boundaries with 3D domain decomposition. At the same time
4808 * the blocks will not become too small.
4810 ci_block = (gridi->nc*ci_block_enum)/(ci_block_denom*gridi->ncx*nth);
4812 /* Ensure the blocks are not too small: avoids cache invalidation */
4813 if (ci_block*gridi->na_sc < ci_block_min_atoms)
4815 ci_block = (ci_block_min_atoms + gridi->na_sc - 1)/gridi->na_sc;
4818 /* Without domain decomposition
4819 * or with less than 3 blocks per task, divide in nth blocks.
4821 if (!bDomDec || ci_block*3*nth > gridi->nc)
4823 ci_block = (gridi->nc + nth - 1)/nth;
4829 /* Generates the part of pair-list nbl assigned to our thread */
4830 static void nbnxn_make_pairlist_part(const nbnxn_search_t nbs,
4831 const nbnxn_grid_t *gridi,
4832 const nbnxn_grid_t *gridj,
4833 nbnxn_search_work_t *work,
4834 const nbnxn_atomdata_t *nbat,
4835 const t_blocka *excl,
4839 gmx_bool bFBufferFlag,
4842 int nsubpair_tot_est,
4844 nbnxn_pairlist_t *nbl,
4849 real rl2, rl_fep2 = 0;
4852 int ci_b, ci, ci_x, ci_y, ci_xy, cj;
4858 int conv_i, cell0_i;
4859 const nbnxn_bb_t *bb_i = NULL;
4861 const float *pbb_i = NULL;
4863 const float *bbcz_i, *bbcz_j;
4865 real bx0, bx1, by0, by1, bz0, bz1;
4867 real d2cx, d2z, d2z_cx, d2z_cy, d2zx, d2zxy, d2xy;
4868 int cxf, cxl, cyf, cyf_x, cyl;
4870 int c0, c1, cs, cf, cl;
4873 int gridi_flag_shift = 0, gridj_flag_shift = 0;
4874 gmx_bitmask_t *gridj_flag = NULL;
4875 int ncj_old_i, ncj_old_j;
4877 nbs_cycle_start(&work->cc[enbsCCsearch]);
4879 if (gridj->bSimple != nbl->bSimple)
4881 gmx_incons("Grid incompatible with pair-list");
4885 nbl->na_sc = gridj->na_sc;
4886 nbl->na_ci = gridj->na_c;
4887 nbl->na_cj = nbnxn_kernel_to_cj_size(nb_kernel_type);
4888 na_cj_2log = get_2log(nbl->na_cj);
4894 /* Determine conversion of clusters to flag blocks */
4895 gridi_flag_shift = 0;
4896 while ((nbl->na_ci<<gridi_flag_shift) < NBNXN_BUFFERFLAG_SIZE)
4900 gridj_flag_shift = 0;
4901 while ((nbl->na_cj<<gridj_flag_shift) < NBNXN_BUFFERFLAG_SIZE)
4906 gridj_flag = work->buffer_flags.flag;
4909 copy_mat(nbs->box, box);
4911 rl2 = nbl->rlist*nbl->rlist;
4913 if (nbs->bFEP && !nbl->bSimple)
4915 /* Determine an atom-pair list cut-off distance for FEP atom pairs.
4916 * We should not simply use rlist, since then we would not have
4917 * the small, effective buffering of the NxN lists.
4918 * The buffer is on overestimate, but the resulting cost for pairs
4919 * beyond rlist is neglible compared to the FEP pairs within rlist.
4921 rl_fep2 = nbl->rlist + effective_buffer_1x1_vs_MxN(gridi, gridj);
4925 fprintf(debug, "nbl_fep atom-pair rlist %f\n", rl_fep2);
4927 rl_fep2 = rl_fep2*rl_fep2;
4930 rbb2 = boundingbox_only_distance2(gridi, gridj, nbl->rlist, nbl->bSimple);
4934 fprintf(debug, "nbl bounding box only distance %f\n", sqrt(rbb2));
4937 /* Set the shift range */
4938 for (d = 0; d < DIM; d++)
4940 /* Check if we need periodicity shifts.
4941 * Without PBC or with domain decomposition we don't need them.
4943 if (d >= ePBC2npbcdim(nbs->ePBC) || nbs->dd_dim[d])
4950 box[XX][XX] - fabs(box[YY][XX]) - fabs(box[ZZ][XX]) < sqrt(rl2))
4961 if (nbl->bSimple && !gridi->bSimple)
4963 conv_i = gridi->na_sc/gridj->na_sc;
4964 bb_i = gridi->bb_simple;
4965 bbcz_i = gridi->bbcz_simple;
4966 flags_i = gridi->flags_simple;
4981 /* We use the normal bounding box format for both grid types */
4984 bbcz_i = gridi->bbcz;
4985 flags_i = gridi->flags;
4987 cell0_i = gridi->cell0*conv_i;
4989 bbcz_j = gridj->bbcz;
4993 /* Blocks of the conversion factor - 1 give a large repeat count
4994 * combined with a small block size. This should result in good
4995 * load balancing for both small and large domains.
4997 ci_block = conv_i - 1;
5001 fprintf(debug, "nbl nc_i %d col.av. %.1f ci_block %d\n",
5002 gridi->nc, gridi->nc/(double)(gridi->ncx*gridi->ncy), ci_block);
5008 /* Initially ci_b and ci to 1 before where we want them to start,
5009 * as they will both be incremented in next_ci.
5012 ci = th*ci_block - 1;
5015 while (next_ci(gridi, conv_i, nth, ci_block, &ci_x, &ci_y, &ci_b, &ci))
5017 if (nbl->bSimple && flags_i[ci] == 0)
5022 ncj_old_i = nbl->ncj;
5025 if (gridj != gridi && shp[XX] == 0)
5029 bx1 = bb_i[ci].upper[BB_X];
5033 bx1 = gridi->c0[XX] + (ci_x+1)*gridi->sx;
5035 if (bx1 < gridj->c0[XX])
5037 d2cx = sqr(gridj->c0[XX] - bx1);
5046 ci_xy = ci_x*gridi->ncy + ci_y;
5048 /* Loop over shift vectors in three dimensions */
5049 for (tz = -shp[ZZ]; tz <= shp[ZZ]; tz++)
5051 shz = tz*box[ZZ][ZZ];
5053 bz0 = bbcz_i[ci*NNBSBB_D ] + shz;
5054 bz1 = bbcz_i[ci*NNBSBB_D+1] + shz;
5066 d2z = sqr(bz0 - box[ZZ][ZZ]);
5069 d2z_cx = d2z + d2cx;
5077 bz1/((real)(gridi->cxy_ind[ci_xy+1] - gridi->cxy_ind[ci_xy]));
5082 /* The check with bz1_frac close to or larger than 1 comes later */
5084 for (ty = -shp[YY]; ty <= shp[YY]; ty++)
5086 shy = ty*box[YY][YY] + tz*box[ZZ][YY];
5090 by0 = bb_i[ci].lower[BB_Y] + shy;
5091 by1 = bb_i[ci].upper[BB_Y] + shy;
5095 by0 = gridi->c0[YY] + (ci_y )*gridi->sy + shy;
5096 by1 = gridi->c0[YY] + (ci_y+1)*gridi->sy + shy;
5099 get_cell_range(by0, by1,
5100 gridj->ncy, gridj->c0[YY], gridj->sy, gridj->inv_sy,
5110 if (by1 < gridj->c0[YY])
5112 d2z_cy += sqr(gridj->c0[YY] - by1);
5114 else if (by0 > gridj->c1[YY])
5116 d2z_cy += sqr(by0 - gridj->c1[YY]);
5119 for (tx = -shp[XX]; tx <= shp[XX]; tx++)
5121 shift = XYZ2IS(tx, ty, tz);
5123 #ifdef NBNXN_SHIFT_BACKWARD
5124 if (gridi == gridj && shift > CENTRAL)
5130 shx = tx*box[XX][XX] + ty*box[YY][XX] + tz*box[ZZ][XX];
5134 bx0 = bb_i[ci].lower[BB_X] + shx;
5135 bx1 = bb_i[ci].upper[BB_X] + shx;
5139 bx0 = gridi->c0[XX] + (ci_x )*gridi->sx + shx;
5140 bx1 = gridi->c0[XX] + (ci_x+1)*gridi->sx + shx;
5143 get_cell_range(bx0, bx1,
5144 gridj->ncx, gridj->c0[XX], gridj->sx, gridj->inv_sx,
5155 new_ci_entry(nbl, cell0_i+ci, shift, flags_i[ci]);
5159 new_sci_entry(nbl, cell0_i+ci, shift);
5162 #ifndef NBNXN_SHIFT_BACKWARD
5165 if (shift == CENTRAL && gridi == gridj &&
5169 /* Leave the pairs with i > j.
5170 * x is the major index, so skip half of it.
5177 set_icell_bb_simple(bb_i, ci, shx, shy, shz,
5183 set_icell_bbxxxx_supersub(pbb_i, ci, shx, shy, shz,
5186 set_icell_bb_supersub(bb_i, ci, shx, shy, shz,
5191 nbs->icell_set_x(cell0_i+ci, shx, shy, shz,
5192 gridi->na_c, nbat->xstride, nbat->x,
5195 for (cx = cxf; cx <= cxl; cx++)
5198 if (gridj->c0[XX] + cx*gridj->sx > bx1)
5200 d2zx += sqr(gridj->c0[XX] + cx*gridj->sx - bx1);
5202 else if (gridj->c0[XX] + (cx+1)*gridj->sx < bx0)
5204 d2zx += sqr(gridj->c0[XX] + (cx+1)*gridj->sx - bx0);
5207 #ifndef NBNXN_SHIFT_BACKWARD
5208 if (gridi == gridj &&
5209 cx == 0 && cyf < ci_y)
5211 if (gridi == gridj &&
5212 cx == 0 && shift == CENTRAL && cyf < ci_y)
5215 /* Leave the pairs with i > j.
5216 * Skip half of y when i and j have the same x.
5225 for (cy = cyf_x; cy <= cyl; cy++)
5227 c0 = gridj->cxy_ind[cx*gridj->ncy+cy];
5228 c1 = gridj->cxy_ind[cx*gridj->ncy+cy+1];
5229 #ifdef NBNXN_SHIFT_BACKWARD
5230 if (gridi == gridj &&
5231 shift == CENTRAL && c0 < ci)
5238 if (gridj->c0[YY] + cy*gridj->sy > by1)
5240 d2zxy += sqr(gridj->c0[YY] + cy*gridj->sy - by1);
5242 else if (gridj->c0[YY] + (cy+1)*gridj->sy < by0)
5244 d2zxy += sqr(gridj->c0[YY] + (cy+1)*gridj->sy - by0);
5246 if (c1 > c0 && d2zxy < rl2)
5248 cs = c0 + (int)(bz1_frac*(c1 - c0));
5256 /* Find the lowest cell that can possibly
5261 (bbcz_j[cf*NNBSBB_D+1] >= bz0 ||
5262 d2xy + sqr(bbcz_j[cf*NNBSBB_D+1] - bz0) < rl2))
5267 /* Find the highest cell that can possibly
5272 (bbcz_j[cl*NNBSBB_D] <= bz1 ||
5273 d2xy + sqr(bbcz_j[cl*NNBSBB_D] - bz1) < rl2))
5278 #ifdef NBNXN_REFCODE
5280 /* Simple reference code, for debugging,
5281 * overrides the more complex code above.
5286 for (k = c0; k < c1; k++)
5288 if (box_dist2(bx0, bx1, by0, by1, bz0, bz1, bb+k) < rl2 &&
5293 if (box_dist2(bx0, bx1, by0, by1, bz0, bz1, bb+k) < rl2 &&
5304 /* We want each atom/cell pair only once,
5305 * only use cj >= ci.
5307 #ifndef NBNXN_SHIFT_BACKWARD
5310 if (shift == CENTRAL)
5319 /* For f buffer flags with simple lists */
5320 ncj_old_j = nbl->ncj;
5322 switch (nb_kernel_type)
5324 case nbnxnk4x4_PlainC:
5325 check_subcell_list_space_simple(nbl, cl-cf+1);
5327 make_cluster_list_simple(gridj,
5329 (gridi == gridj && shift == CENTRAL),
5334 #ifdef GMX_NBNXN_SIMD_4XN
5335 case nbnxnk4xN_SIMD_4xN:
5336 check_subcell_list_space_simple(nbl, ci_to_cj(na_cj_2log, cl-cf)+2);
5337 make_cluster_list_simd_4xn(gridj,
5339 (gridi == gridj && shift == CENTRAL),
5345 #ifdef GMX_NBNXN_SIMD_2XNN
5346 case nbnxnk4xN_SIMD_2xNN:
5347 check_subcell_list_space_simple(nbl, ci_to_cj(na_cj_2log, cl-cf)+2);
5348 make_cluster_list_simd_2xnn(gridj,
5350 (gridi == gridj && shift == CENTRAL),
5356 case nbnxnk8x8x8_PlainC:
5357 case nbnxnk8x8x8_GPU:
5358 check_subcell_list_space_supersub(nbl, cl-cf+1);
5359 for (cj = cf; cj <= cl; cj++)
5361 make_cluster_list_supersub(gridi, gridj,
5363 (gridi == gridj && shift == CENTRAL && ci == cj),
5364 nbat->xstride, nbat->x,
5370 ncpcheck += cl - cf + 1;
5372 if (bFBufferFlag && nbl->ncj > ncj_old_j)
5376 cbf = nbl->cj[ncj_old_j].cj >> gridj_flag_shift;
5377 cbl = nbl->cj[nbl->ncj-1].cj >> gridj_flag_shift;
5378 for (cb = cbf; cb <= cbl; cb++)
5380 bitmask_init_bit(&gridj_flag[cb], th);
5388 /* Set the exclusions for this ci list */
5391 set_ci_top_excls(nbs,
5393 shift == CENTRAL && gridi == gridj,
5396 &(nbl->ci[nbl->nci]),
5401 make_fep_list(nbs, nbat, nbl,
5402 shift == CENTRAL && gridi == gridj,
5403 &(nbl->ci[nbl->nci]),
5404 gridi, gridj, nbl_fep);
5409 set_sci_top_excls(nbs,
5411 shift == CENTRAL && gridi == gridj,
5413 &(nbl->sci[nbl->nsci]),
5418 make_fep_list_supersub(nbs, nbat, nbl,
5419 shift == CENTRAL && gridi == gridj,
5420 &(nbl->sci[nbl->nsci]),
5423 gridi, gridj, nbl_fep);
5427 /* Close this ci list */
5430 close_ci_entry_simple(nbl);
5434 close_ci_entry_supersub(nbl,
5436 progBal, nsubpair_tot_est,
5443 if (bFBufferFlag && nbl->ncj > ncj_old_i)
5445 bitmask_init_bit(&(work->buffer_flags.flag[(gridi->cell0+ci)>>gridi_flag_shift]), th);
5449 work->ndistc = ndistc;
5451 nbs_cycle_stop(&work->cc[enbsCCsearch]);
5455 fprintf(debug, "number of distance checks %d\n", ndistc);
5456 fprintf(debug, "ncpcheck %s %d\n", gridi == gridj ? "local" : "non-local",
5461 print_nblist_statistics_simple(debug, nbl, nbs, rlist);
5465 print_nblist_statistics_supersub(debug, nbl, nbs, rlist);
5470 fprintf(debug, "nbl FEP list pairs: %d\n", nbl_fep->nrj);
5475 static void reduce_buffer_flags(const nbnxn_search_t nbs,
5477 const nbnxn_buffer_flags_t *dest)
5480 gmx_bitmask_t *flag;
5482 for (s = 0; s < nsrc; s++)
5484 flag = nbs->work[s].buffer_flags.flag;
5486 for (b = 0; b < dest->nflag; b++)
5488 bitmask_union(&(dest->flag[b]), flag[b]);
5493 static void print_reduction_cost(const nbnxn_buffer_flags_t *flags, int nout)
5495 int nelem, nkeep, ncopy, nred, b, c, out;
5496 gmx_bitmask_t mask_0;
5502 bitmask_init_bit(&mask_0, 0);
5503 for (b = 0; b < flags->nflag; b++)
5505 if (bitmask_is_equal(flags->flag[b], mask_0))
5507 /* Only flag 0 is set, no copy of reduction required */
5511 else if (!bitmask_is_zero(flags->flag[b]))
5514 for (out = 0; out < nout; out++)
5516 if (bitmask_is_set(flags->flag[b], out))
5533 fprintf(debug, "nbnxn reduction: #flag %d #list %d elem %4.2f, keep %4.2f copy %4.2f red %4.2f\n",
5535 nelem/(double)(flags->nflag),
5536 nkeep/(double)(flags->nflag),
5537 ncopy/(double)(flags->nflag),
5538 nred/(double)(flags->nflag));
5541 /* Perform a count (linear) sort to sort the smaller lists to the end.
5542 * This avoids load imbalance on the GPU, as large lists will be
5543 * scheduled and executed first and the smaller lists later.
5544 * Load balancing between multi-processors only happens at the end
5545 * and there smaller lists lead to more effective load balancing.
5546 * The sorting is done on the cj4 count, not on the actual pair counts.
5547 * Not only does this make the sort faster, but it also results in
5548 * better load balancing than using a list sorted on exact load.
5549 * This function swaps the pointer in the pair list to avoid a copy operation.
5551 static void sort_sci(nbnxn_pairlist_t *nbl)
5553 nbnxn_list_work_t *work;
5554 int m, i, s, s0, s1;
5555 nbnxn_sci_t *sci_sort;
5557 if (nbl->ncj4 <= nbl->nsci)
5559 /* nsci = 0 or all sci have size 1, sorting won't change the order */
5565 /* We will distinguish differences up to double the average */
5566 m = (2*nbl->ncj4)/nbl->nsci;
5568 if (m + 1 > work->sort_nalloc)
5570 work->sort_nalloc = over_alloc_large(m + 1);
5571 srenew(work->sort, work->sort_nalloc);
5574 if (work->sci_sort_nalloc != nbl->sci_nalloc)
5576 work->sci_sort_nalloc = nbl->sci_nalloc;
5577 nbnxn_realloc_void((void **)&work->sci_sort,
5579 work->sci_sort_nalloc*sizeof(*work->sci_sort),
5580 nbl->alloc, nbl->free);
5583 /* Count the entries of each size */
5584 for (i = 0; i <= m; i++)
5588 for (s = 0; s < nbl->nsci; s++)
5590 i = min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start);
5593 /* Calculate the offset for each count */
5596 for (i = m - 1; i >= 0; i--)
5599 work->sort[i] = work->sort[i + 1] + s0;
5603 /* Sort entries directly into place */
5604 sci_sort = work->sci_sort;
5605 for (s = 0; s < nbl->nsci; s++)
5607 i = min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start);
5608 sci_sort[work->sort[i]++] = nbl->sci[s];
5611 /* Swap the sci pointers so we use the new, sorted list */
5612 work->sci_sort = nbl->sci;
5613 nbl->sci = sci_sort;
5616 /* Make a local or non-local pair-list, depending on iloc */
5617 void nbnxn_make_pairlist(const nbnxn_search_t nbs,
5618 nbnxn_atomdata_t *nbat,
5619 const t_blocka *excl,
5621 int min_ci_balanced,
5622 nbnxn_pairlist_set_t *nbl_list,
5627 nbnxn_grid_t *gridi, *gridj;
5629 int nzi, zi, zj0, zj1, zj;
5630 int nsubpair_target, nsubpair_tot_est;
5633 nbnxn_pairlist_t **nbl;
5635 gmx_bool CombineNBLists;
5637 int np_tot, np_noq, np_hlj, nap;
5639 /* Check if we are running hybrid GPU + CPU nbnxn mode */
5640 bGPUCPU = (!nbs->grid[0].bSimple && nbl_list->bSimple);
5642 nnbl = nbl_list->nnbl;
5643 nbl = nbl_list->nbl;
5644 CombineNBLists = nbl_list->bCombined;
5648 fprintf(debug, "ns making %d nblists\n", nnbl);
5651 nbat->bUseBufferFlags = (nbat->nout > 1);
5652 /* We should re-init the flags before making the first list */
5653 if (nbat->bUseBufferFlags && (LOCAL_I(iloc) || bGPUCPU))
5655 init_buffer_flags(&nbat->buffer_flags, nbat->natoms);
5658 if (nbl_list->bSimple)
5660 switch (nb_kernel_type)
5662 #ifdef GMX_NBNXN_SIMD_4XN
5663 case nbnxnk4xN_SIMD_4xN:
5664 nbs->icell_set_x = icell_set_x_simd_4xn;
5667 #ifdef GMX_NBNXN_SIMD_2XNN
5668 case nbnxnk4xN_SIMD_2xNN:
5669 nbs->icell_set_x = icell_set_x_simd_2xnn;
5673 nbs->icell_set_x = icell_set_x_simple;
5679 #ifdef NBNXN_SEARCH_BB_SIMD4
5680 nbs->icell_set_x = icell_set_x_supersub_simd4;
5682 nbs->icell_set_x = icell_set_x_supersub;
5688 /* Only zone (grid) 0 vs 0 */
5695 nzi = nbs->zones->nizone;
5698 if (!nbl_list->bSimple && min_ci_balanced > 0)
5700 get_nsubpair_target(nbs, iloc, rlist, min_ci_balanced,
5701 &nsubpair_target, &nsubpair_tot_est);
5705 nsubpair_target = 0;
5706 nsubpair_tot_est = 0;
5709 /* Clear all pair-lists */
5710 for (th = 0; th < nnbl; th++)
5712 clear_pairlist(nbl[th]);
5716 clear_pairlist_fep(nbl_list->nbl_fep[th]);
5720 for (zi = 0; zi < nzi; zi++)
5722 gridi = &nbs->grid[zi];
5724 if (NONLOCAL_I(iloc))
5726 zj0 = nbs->zones->izone[zi].j0;
5727 zj1 = nbs->zones->izone[zi].j1;
5733 for (zj = zj0; zj < zj1; zj++)
5735 gridj = &nbs->grid[zj];
5739 fprintf(debug, "ns search grid %d vs %d\n", zi, zj);
5742 nbs_cycle_start(&nbs->cc[enbsCCsearch]);
5744 if (nbl[0]->bSimple && !gridi->bSimple)
5746 /* Hybrid list, determine blocking later */
5751 ci_block = get_ci_block_size(gridi, nbs->DomDec, nnbl);
5754 /* With GPU: generate progressively smaller lists for
5755 * load balancing for local only or non-local with 2 zones.
5757 progBal = (LOCAL_I(iloc) || nbs->zones->n <= 2);
5759 #pragma omp parallel for num_threads(nnbl) schedule(static)
5760 for (th = 0; th < nnbl; th++)
5762 /* Re-init the thread-local work flag data before making
5763 * the first list (not an elegant conditional).
5765 if (nbat->bUseBufferFlags && ((zi == 0 && zj == 0) ||
5766 (bGPUCPU && zi == 0 && zj == 1)))
5768 init_buffer_flags(&nbs->work[th].buffer_flags, nbat->natoms);
5771 if (CombineNBLists && th > 0)
5773 clear_pairlist(nbl[th]);
5776 /* Divide the i super cell equally over the nblists */
5777 nbnxn_make_pairlist_part(nbs, gridi, gridj,
5778 &nbs->work[th], nbat, excl,
5782 nbat->bUseBufferFlags,
5784 progBal, nsubpair_tot_est,
5787 nbl_list->nbl_fep[th]);
5789 nbs_cycle_stop(&nbs->cc[enbsCCsearch]);
5794 for (th = 0; th < nnbl; th++)
5796 inc_nrnb(nrnb, eNR_NBNXN_DIST2, nbs->work[th].ndistc);
5798 if (nbl_list->bSimple)
5800 np_tot += nbl[th]->ncj;
5801 np_noq += nbl[th]->work->ncj_noq;
5802 np_hlj += nbl[th]->work->ncj_hlj;
5806 /* This count ignores potential subsequent pair pruning */
5807 np_tot += nbl[th]->nci_tot;
5810 nap = nbl[0]->na_ci*nbl[0]->na_cj;
5811 nbl_list->natpair_ljq = (np_tot - np_noq)*nap - np_hlj*nap/2;
5812 nbl_list->natpair_lj = np_noq*nap;
5813 nbl_list->natpair_q = np_hlj*nap/2;
5815 if (CombineNBLists && nnbl > 1)
5817 nbs_cycle_start(&nbs->cc[enbsCCcombine]);
5819 combine_nblists(nnbl-1, nbl+1, nbl[0]);
5821 nbs_cycle_stop(&nbs->cc[enbsCCcombine]);
5826 if (!nbl_list->bSimple)
5828 /* Sort the entries on size, large ones first */
5829 if (CombineNBLists || nnbl == 1)
5835 #pragma omp parallel for num_threads(nnbl) schedule(static)
5836 for (th = 0; th < nnbl; th++)
5843 if (nbat->bUseBufferFlags)
5845 reduce_buffer_flags(nbs, nnbl, &nbat->buffer_flags);
5850 /* Balance the free-energy lists over all the threads */
5851 balance_fep_lists(nbs, nbl_list);
5854 /* Special performance logging stuff (env.var. GMX_NBNXN_CYCLE) */
5857 nbs->search_count++;
5859 if (nbs->print_cycles &&
5860 (!nbs->DomDec || (nbs->DomDec && !LOCAL_I(iloc))) &&
5861 nbs->search_count % 100 == 0)
5863 nbs_cycle_print(stderr, nbs);
5866 if (debug && (CombineNBLists && nnbl > 1))
5868 if (nbl[0]->bSimple)
5870 print_nblist_statistics_simple(debug, nbl[0], nbs, rlist);
5874 print_nblist_statistics_supersub(debug, nbl[0], nbs, rlist);
5882 if (nbl[0]->bSimple)
5884 print_nblist_ci_cj(debug, nbl[0]);
5888 print_nblist_sci_cj(debug, nbl[0]);
5892 if (nbat->bUseBufferFlags)
5894 print_reduction_cost(&nbat->buffer_flags, nnbl);