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/pbcutil/ishift.h"
55 #include "gromacs/pbcutil/pbc.h"
56 #include "gromacs/utility/smalloc.h"
58 /* nbnxn_internal.h included gromacs/simd/macros.h */
59 #include "gromacs/mdlib/nbnxn_internal.h"
61 #include "gromacs/simd/vector_operations.h"
64 #ifdef NBNXN_SEARCH_BB_SIMD4
65 /* Always use 4-wide SIMD for bounding box calculations */
68 /* Single precision BBs + coordinates, we can also load coordinates with SIMD */
69 # define NBNXN_SEARCH_SIMD4_FLOAT_X_BB
72 # if defined NBNXN_SEARCH_SIMD4_FLOAT_X_BB && (GPU_NSUBCELL == 4 || GPU_NSUBCELL == 8)
73 /* Store bounding boxes with x, y and z coordinates in packs of 4 */
74 # define NBNXN_PBB_SIMD4
77 /* The packed bounding box coordinate stride is always set to 4.
78 * With AVX we could use 8, but that turns out not to be faster.
81 # define STRIDE_PBB_2LOG 2
83 #endif /* NBNXN_SEARCH_BB_SIMD4 */
87 /* The functions below are macros as they are performance sensitive */
89 /* 4x4 list, pack=4: no complex conversion required */
90 /* i-cluster to j-cluster conversion */
91 #define CI_TO_CJ_J4(ci) (ci)
92 /* cluster index to coordinate array index conversion */
93 #define X_IND_CI_J4(ci) ((ci)*STRIDE_P4)
94 #define X_IND_CJ_J4(cj) ((cj)*STRIDE_P4)
96 /* 4x2 list, pack=4: j-cluster size is half the packing width */
97 /* i-cluster to j-cluster conversion */
98 #define CI_TO_CJ_J2(ci) ((ci)<<1)
99 /* cluster index to coordinate array index conversion */
100 #define X_IND_CI_J2(ci) ((ci)*STRIDE_P4)
101 #define X_IND_CJ_J2(cj) (((cj)>>1)*STRIDE_P4 + ((cj) & 1)*(PACK_X4>>1))
103 /* 4x8 list, pack=8: i-cluster size is half the packing width */
104 /* i-cluster to j-cluster conversion */
105 #define CI_TO_CJ_J8(ci) ((ci)>>1)
106 /* cluster index to coordinate array index conversion */
107 #define X_IND_CI_J8(ci) (((ci)>>1)*STRIDE_P8 + ((ci) & 1)*(PACK_X8>>1))
108 #define X_IND_CJ_J8(cj) ((cj)*STRIDE_P8)
110 /* The j-cluster size is matched to the SIMD width */
111 #if GMX_SIMD_REAL_WIDTH == 2
112 #define CI_TO_CJ_SIMD_4XN(ci) CI_TO_CJ_J2(ci)
113 #define X_IND_CI_SIMD_4XN(ci) X_IND_CI_J2(ci)
114 #define X_IND_CJ_SIMD_4XN(cj) X_IND_CJ_J2(cj)
116 #if GMX_SIMD_REAL_WIDTH == 4
117 #define CI_TO_CJ_SIMD_4XN(ci) CI_TO_CJ_J4(ci)
118 #define X_IND_CI_SIMD_4XN(ci) X_IND_CI_J4(ci)
119 #define X_IND_CJ_SIMD_4XN(cj) X_IND_CJ_J4(cj)
121 #if GMX_SIMD_REAL_WIDTH == 8
122 #define CI_TO_CJ_SIMD_4XN(ci) CI_TO_CJ_J8(ci)
123 #define X_IND_CI_SIMD_4XN(ci) X_IND_CI_J8(ci)
124 #define X_IND_CJ_SIMD_4XN(cj) X_IND_CJ_J8(cj)
125 /* Half SIMD with j-cluster size */
126 #define CI_TO_CJ_SIMD_2XNN(ci) CI_TO_CJ_J4(ci)
127 #define X_IND_CI_SIMD_2XNN(ci) X_IND_CI_J4(ci)
128 #define X_IND_CJ_SIMD_2XNN(cj) X_IND_CJ_J4(cj)
130 #if GMX_SIMD_REAL_WIDTH == 16
131 #define CI_TO_CJ_SIMD_2XNN(ci) CI_TO_CJ_J8(ci)
132 #define X_IND_CI_SIMD_2XNN(ci) X_IND_CI_J8(ci)
133 #define X_IND_CJ_SIMD_2XNN(cj) X_IND_CJ_J8(cj)
135 #error "unsupported GMX_SIMD_REAL_WIDTH"
141 #endif /* GMX_NBNXN_SIMD */
144 #ifdef NBNXN_SEARCH_BB_SIMD4
145 /* Store bounding boxes corners as quadruplets: xxxxyyyyzzzz */
147 /* Size of bounding box corners quadruplet */
148 #define NNBSBB_XXXX (NNBSBB_D*DIM*STRIDE_PBB)
151 /* We shift the i-particles backward for PBC.
152 * This leads to more conditionals than shifting forward.
153 * We do this to get more balanced pair lists.
155 #define NBNXN_SHIFT_BACKWARD
158 /* This define is a lazy way to avoid interdependence of the grid
159 * and searching data structures.
161 #define NBNXN_NA_SC_MAX (GPU_NSUBCELL*NBNXN_GPU_CLUSTER_SIZE)
164 static void nbs_cycle_clear(nbnxn_cycle_t *cc)
168 for (i = 0; i < enbsCCnr; i++)
175 static double Mcyc_av(const nbnxn_cycle_t *cc)
177 return (double)cc->c*1e-6/cc->count;
180 static void nbs_cycle_print(FILE *fp, const nbnxn_search_t nbs)
186 fprintf(fp, "ns %4d grid %4.1f search %4.1f red.f %5.3f",
187 nbs->cc[enbsCCgrid].count,
188 Mcyc_av(&nbs->cc[enbsCCgrid]),
189 Mcyc_av(&nbs->cc[enbsCCsearch]),
190 Mcyc_av(&nbs->cc[enbsCCreducef]));
192 if (nbs->nthread_max > 1)
194 if (nbs->cc[enbsCCcombine].count > 0)
196 fprintf(fp, " comb %5.2f",
197 Mcyc_av(&nbs->cc[enbsCCcombine]));
199 fprintf(fp, " s. th");
200 for (t = 0; t < nbs->nthread_max; t++)
202 fprintf(fp, " %4.1f",
203 Mcyc_av(&nbs->work[t].cc[enbsCCsearch]));
209 static void nbnxn_grid_init(nbnxn_grid_t * grid)
212 grid->cxy_ind = NULL;
213 grid->cxy_nalloc = 0;
219 static int get_2log(int n)
224 while ((1<<log2) < n)
230 gmx_fatal(FARGS, "nbnxn na_c (%d) is not a power of 2", n);
236 static int nbnxn_kernel_to_ci_size(int nb_kernel_type)
238 switch (nb_kernel_type)
240 case nbnxnk4x4_PlainC:
241 case nbnxnk4xN_SIMD_4xN:
242 case nbnxnk4xN_SIMD_2xNN:
243 return NBNXN_CPU_CLUSTER_I_SIZE;
244 case nbnxnk8x8x8_CUDA:
245 case nbnxnk8x8x8_PlainC:
246 /* The cluster size for super/sub lists is only set here.
247 * Any value should work for the pair-search and atomdata code.
248 * The kernels, of course, might require a particular value.
250 return NBNXN_GPU_CLUSTER_SIZE;
252 gmx_incons("unknown kernel type");
258 int nbnxn_kernel_to_cj_size(int nb_kernel_type)
260 int nbnxn_simd_width = 0;
263 #ifdef GMX_NBNXN_SIMD
264 nbnxn_simd_width = GMX_SIMD_REAL_WIDTH;
267 switch (nb_kernel_type)
269 case nbnxnk4x4_PlainC:
270 cj_size = NBNXN_CPU_CLUSTER_I_SIZE;
272 case nbnxnk4xN_SIMD_4xN:
273 cj_size = nbnxn_simd_width;
275 case nbnxnk4xN_SIMD_2xNN:
276 cj_size = nbnxn_simd_width/2;
278 case nbnxnk8x8x8_CUDA:
279 case nbnxnk8x8x8_PlainC:
280 cj_size = nbnxn_kernel_to_ci_size(nb_kernel_type);
283 gmx_incons("unknown kernel type");
289 static int ci_to_cj(int na_cj_2log, int ci)
293 case 2: return ci; break;
294 case 1: return (ci<<1); break;
295 case 3: return (ci>>1); break;
301 gmx_bool nbnxn_kernel_pairlist_simple(int nb_kernel_type)
303 if (nb_kernel_type == nbnxnkNotSet)
305 gmx_fatal(FARGS, "Non-bonded kernel type not set for Verlet-style pair-list.");
308 switch (nb_kernel_type)
310 case nbnxnk8x8x8_CUDA:
311 case nbnxnk8x8x8_PlainC:
314 case nbnxnk4x4_PlainC:
315 case nbnxnk4xN_SIMD_4xN:
316 case nbnxnk4xN_SIMD_2xNN:
320 gmx_incons("Invalid nonbonded kernel type passed!");
325 /* Initializes a single nbnxn_pairlist_t data structure */
326 static void nbnxn_init_pairlist_fep(t_nblist *nl)
328 nl->type = GMX_NBLIST_INTERACTION_FREE_ENERGY;
329 nl->igeometry = GMX_NBLIST_GEOMETRY_PARTICLE_PARTICLE;
330 /* The interaction functions are set in the free energy kernel fuction */
349 void nbnxn_init_search(nbnxn_search_t * nbs_ptr,
351 gmx_domdec_zones_t *zones,
363 nbs->DomDec = (n_dd_cells != NULL);
365 clear_ivec(nbs->dd_dim);
371 for (d = 0; d < DIM; d++)
373 if ((*n_dd_cells)[d] > 1)
376 /* Each grid matches a DD zone */
382 snew(nbs->grid, nbs->ngrid);
383 for (g = 0; g < nbs->ngrid; g++)
385 nbnxn_grid_init(&nbs->grid[g]);
388 nbs->cell_nalloc = 0;
392 nbs->nthread_max = nthread_max;
394 /* Initialize the work data structures for each thread */
395 snew(nbs->work, nbs->nthread_max);
396 for (t = 0; t < nbs->nthread_max; t++)
398 nbs->work[t].cxy_na = NULL;
399 nbs->work[t].cxy_na_nalloc = 0;
400 nbs->work[t].sort_work = NULL;
401 nbs->work[t].sort_work_nalloc = 0;
403 snew(nbs->work[t].nbl_fep, 1);
404 nbnxn_init_pairlist_fep(nbs->work[t].nbl_fep);
407 /* Initialize detailed nbsearch cycle counting */
408 nbs->print_cycles = (getenv("GMX_NBNXN_CYCLE") != 0);
409 nbs->search_count = 0;
410 nbs_cycle_clear(nbs->cc);
411 for (t = 0; t < nbs->nthread_max; t++)
413 nbs_cycle_clear(nbs->work[t].cc);
417 static real grid_atom_density(int n, rvec corner0, rvec corner1)
423 /* To avoid zero density we use a minimum of 1 atom */
427 rvec_sub(corner1, corner0, size);
429 return n/(size[XX]*size[YY]*size[ZZ]);
432 static int set_grid_size_xy(const nbnxn_search_t nbs,
435 int n, rvec corner0, rvec corner1,
440 real adens, tlen, tlen_x, tlen_y, nc_max;
443 rvec_sub(corner1, corner0, size);
447 assert(atom_density > 0);
449 /* target cell length */
452 /* To minimize the zero interactions, we should make
453 * the largest of the i/j cell cubic.
455 na_c = max(grid->na_c, grid->na_cj);
457 /* Approximately cubic cells */
458 tlen = pow(na_c/atom_density, 1.0/3.0);
464 /* Approximately cubic sub cells */
465 tlen = pow(grid->na_c/atom_density, 1.0/3.0);
466 tlen_x = tlen*GPU_NSUBCELL_X;
467 tlen_y = tlen*GPU_NSUBCELL_Y;
469 /* We round ncx and ncy down, because we get less cell pairs
470 * in the nbsist when the fixed cell dimensions (x,y) are
471 * larger than the variable one (z) than the other way around.
473 grid->ncx = max(1, (int)(size[XX]/tlen_x));
474 grid->ncy = max(1, (int)(size[YY]/tlen_y));
482 grid->sx = size[XX]/grid->ncx;
483 grid->sy = size[YY]/grid->ncy;
484 grid->inv_sx = 1/grid->sx;
485 grid->inv_sy = 1/grid->sy;
489 /* This is a non-home zone, add an extra row of cells
490 * for particles communicated for bonded interactions.
491 * These can be beyond the cut-off. It doesn't matter where
492 * they end up on the grid, but for performance it's better
493 * if they don't end up in cells that can be within cut-off range.
499 /* We need one additional cell entry for particles moved by DD */
500 if (grid->ncx*grid->ncy+1 > grid->cxy_nalloc)
502 grid->cxy_nalloc = over_alloc_large(grid->ncx*grid->ncy+1);
503 srenew(grid->cxy_na, grid->cxy_nalloc);
504 srenew(grid->cxy_ind, grid->cxy_nalloc+1);
506 for (t = 0; t < nbs->nthread_max; t++)
508 if (grid->ncx*grid->ncy+1 > nbs->work[t].cxy_na_nalloc)
510 nbs->work[t].cxy_na_nalloc = over_alloc_large(grid->ncx*grid->ncy+1);
511 srenew(nbs->work[t].cxy_na, nbs->work[t].cxy_na_nalloc);
515 /* Worst case scenario of 1 atom in each last cell */
516 if (grid->na_cj <= grid->na_c)
518 nc_max = n/grid->na_sc + grid->ncx*grid->ncy;
522 nc_max = n/grid->na_sc + grid->ncx*grid->ncy*grid->na_cj/grid->na_c;
525 if (nc_max > grid->nc_nalloc)
527 grid->nc_nalloc = over_alloc_large(nc_max);
528 srenew(grid->nsubc, grid->nc_nalloc);
529 srenew(grid->bbcz, grid->nc_nalloc*NNBSBB_D);
531 sfree_aligned(grid->bb);
532 /* This snew also zeros the contents, this avoid possible
533 * floating exceptions in SIMD with the unused bb elements.
537 snew_aligned(grid->bb, grid->nc_nalloc, 16);
544 pbb_nalloc = grid->nc_nalloc*GPU_NSUBCELL/STRIDE_PBB*NNBSBB_XXXX;
545 snew_aligned(grid->pbb, pbb_nalloc, 16);
547 snew_aligned(grid->bb, grid->nc_nalloc*GPU_NSUBCELL, 16);
553 if (grid->na_cj == grid->na_c)
555 grid->bbj = grid->bb;
559 sfree_aligned(grid->bbj);
560 snew_aligned(grid->bbj, grid->nc_nalloc*grid->na_c/grid->na_cj, 16);
564 srenew(grid->flags, grid->nc_nalloc);
567 srenew(grid->fep, grid->nc_nalloc*grid->na_sc/grid->na_c);
571 copy_rvec(corner0, grid->c0);
572 copy_rvec(corner1, grid->c1);
573 copy_rvec(size, grid->size);
578 /* We need to sort paricles in grid columns on z-coordinate.
579 * As particle are very often distributed homogeneously, we a sorting
580 * algorithm similar to pigeonhole sort. We multiply the z-coordinate
581 * by a factor, cast to an int and try to store in that hole. If the hole
582 * is full, we move this or another particle. A second pass is needed to make
583 * contiguous elements. SORT_GRID_OVERSIZE is the ratio of holes to particles.
584 * 4 is the optimal value for homogeneous particle distribution and allows
585 * for an O(#particles) sort up till distributions were all particles are
586 * concentrated in 1/4 of the space. No NlogN fallback is implemented,
587 * as it can be expensive to detect imhomogeneous particle distributions.
588 * SGSF is the maximum ratio of holes used, in the worst case all particles
589 * end up in the last hole and we need #particles extra holes at the end.
591 #define SORT_GRID_OVERSIZE 4
592 #define SGSF (SORT_GRID_OVERSIZE + 1)
594 /* Sort particle index a on coordinates x along dim.
595 * Backwards tells if we want decreasing iso increasing coordinates.
596 * h0 is the minimum of the coordinate range.
597 * invh is the 1/length of the sorting range.
598 * n_per_h (>=n) is the expected average number of particles per 1/invh
599 * sort is the sorting work array.
600 * sort should have a size of at least n_per_h*SORT_GRID_OVERSIZE + n,
601 * or easier, allocate at least n*SGSF elements.
603 static void sort_atoms(int dim, gmx_bool Backwards,
604 int gmx_unused dd_zone,
605 int *a, int n, rvec *x,
606 real h0, real invh, int n_per_h,
610 int zi, zim, zi_min, zi_max;
622 gmx_incons("n > n_per_h");
626 /* Transform the inverse range height into the inverse hole height */
627 invh *= n_per_h*SORT_GRID_OVERSIZE;
629 /* Set nsort to the maximum possible number of holes used.
630 * In worst case all n elements end up in the last bin.
632 nsort = n_per_h*SORT_GRID_OVERSIZE + n;
634 /* Determine the index range used, so we can limit it for the second pass */
638 /* Sort the particles using a simple index sort */
639 for (i = 0; i < n; i++)
641 /* The cast takes care of float-point rounding effects below zero.
642 * This code assumes particles are less than 1/SORT_GRID_OVERSIZE
643 * times the box height out of the box.
645 zi = (int)((x[a[i]][dim] - h0)*invh);
648 /* As we can have rounding effect, we use > iso >= here */
649 if (zi < 0 || (dd_zone == 0 && zi > n_per_h*SORT_GRID_OVERSIZE))
651 gmx_fatal(FARGS, "(int)((x[%d][%c]=%f - %f)*%f) = %d, not in 0 - %d*%d\n",
652 a[i], 'x'+dim, x[a[i]][dim], h0, invh, zi,
653 n_per_h, SORT_GRID_OVERSIZE);
657 /* In a non-local domain, particles communcated for bonded interactions
658 * can be far beyond the grid size, which is set by the non-bonded
659 * cut-off distance. We sort such particles into the last cell.
661 if (zi > n_per_h*SORT_GRID_OVERSIZE)
663 zi = n_per_h*SORT_GRID_OVERSIZE;
666 /* Ideally this particle should go in sort cell zi,
667 * but that might already be in use,
668 * in that case find the first empty cell higher up
673 zi_min = min(zi_min, zi);
674 zi_max = max(zi_max, zi);
678 /* We have multiple atoms in the same sorting slot.
679 * Sort on real z for minimal bounding box size.
680 * There is an extra check for identical z to ensure
681 * well-defined output order, independent of input order
682 * to ensure binary reproducibility after restarts.
684 while (sort[zi] >= 0 && ( x[a[i]][dim] > x[sort[zi]][dim] ||
685 (x[a[i]][dim] == x[sort[zi]][dim] &&
693 /* Shift all elements by one slot until we find an empty slot */
696 while (sort[zim] >= 0)
704 zi_max = max(zi_max, zim);
707 zi_max = max(zi_max, zi);
714 for (zi = 0; zi < nsort; zi++)
725 for (zi = zi_max; zi >= zi_min; zi--)
736 gmx_incons("Lost particles while sorting");
741 #define R2F_D(x) ((float)((x) >= 0 ? ((1-GMX_FLOAT_EPS)*(x)) : ((1+GMX_FLOAT_EPS)*(x))))
742 #define R2F_U(x) ((float)((x) >= 0 ? ((1+GMX_FLOAT_EPS)*(x)) : ((1-GMX_FLOAT_EPS)*(x))))
748 /* Coordinate order x,y,z, bb order xyz0 */
749 static void calc_bounding_box(int na, int stride, const real *x, nbnxn_bb_t *bb)
752 real xl, xh, yl, yh, zl, zh;
762 for (j = 1; j < na; j++)
764 xl = min(xl, x[i+XX]);
765 xh = max(xh, x[i+XX]);
766 yl = min(yl, x[i+YY]);
767 yh = max(yh, x[i+YY]);
768 zl = min(zl, x[i+ZZ]);
769 zh = max(zh, x[i+ZZ]);
772 /* Note: possible double to float conversion here */
773 bb->lower[BB_X] = R2F_D(xl);
774 bb->lower[BB_Y] = R2F_D(yl);
775 bb->lower[BB_Z] = R2F_D(zl);
776 bb->upper[BB_X] = R2F_U(xh);
777 bb->upper[BB_Y] = R2F_U(yh);
778 bb->upper[BB_Z] = R2F_U(zh);
781 /* Packed coordinates, bb order xyz0 */
782 static void calc_bounding_box_x_x4(int na, const real *x, nbnxn_bb_t *bb)
785 real xl, xh, yl, yh, zl, zh;
793 for (j = 1; j < na; j++)
795 xl = min(xl, x[j+XX*PACK_X4]);
796 xh = max(xh, x[j+XX*PACK_X4]);
797 yl = min(yl, x[j+YY*PACK_X4]);
798 yh = max(yh, x[j+YY*PACK_X4]);
799 zl = min(zl, x[j+ZZ*PACK_X4]);
800 zh = max(zh, x[j+ZZ*PACK_X4]);
802 /* Note: possible double to float conversion here */
803 bb->lower[BB_X] = R2F_D(xl);
804 bb->lower[BB_Y] = R2F_D(yl);
805 bb->lower[BB_Z] = R2F_D(zl);
806 bb->upper[BB_X] = R2F_U(xh);
807 bb->upper[BB_Y] = R2F_U(yh);
808 bb->upper[BB_Z] = R2F_U(zh);
811 /* Packed coordinates, bb order xyz0 */
812 static void calc_bounding_box_x_x8(int na, const real *x, nbnxn_bb_t *bb)
815 real xl, xh, yl, yh, zl, zh;
823 for (j = 1; j < na; j++)
825 xl = min(xl, x[j+XX*PACK_X8]);
826 xh = max(xh, x[j+XX*PACK_X8]);
827 yl = min(yl, x[j+YY*PACK_X8]);
828 yh = max(yh, x[j+YY*PACK_X8]);
829 zl = min(zl, x[j+ZZ*PACK_X8]);
830 zh = max(zh, x[j+ZZ*PACK_X8]);
832 /* Note: possible double to float conversion here */
833 bb->lower[BB_X] = R2F_D(xl);
834 bb->lower[BB_Y] = R2F_D(yl);
835 bb->lower[BB_Z] = R2F_D(zl);
836 bb->upper[BB_X] = R2F_U(xh);
837 bb->upper[BB_Y] = R2F_U(yh);
838 bb->upper[BB_Z] = R2F_U(zh);
841 /* Packed coordinates, bb order xyz0 */
842 static void calc_bounding_box_x_x4_halves(int na, const real *x,
843 nbnxn_bb_t *bb, nbnxn_bb_t *bbj)
845 calc_bounding_box_x_x4(min(na, 2), x, bbj);
849 calc_bounding_box_x_x4(min(na-2, 2), x+(PACK_X4>>1), bbj+1);
853 /* Set the "empty" bounding box to the same as the first one,
854 * so we don't need to treat special cases in the rest of the code.
856 #ifdef NBNXN_SEARCH_BB_SIMD4
857 gmx_simd4_store_f(&bbj[1].lower[0], gmx_simd4_load_f(&bbj[0].lower[0]));
858 gmx_simd4_store_f(&bbj[1].upper[0], gmx_simd4_load_f(&bbj[0].upper[0]));
864 #ifdef NBNXN_SEARCH_BB_SIMD4
865 gmx_simd4_store_f(&bb->lower[0],
866 gmx_simd4_min_f(gmx_simd4_load_f(&bbj[0].lower[0]),
867 gmx_simd4_load_f(&bbj[1].lower[0])));
868 gmx_simd4_store_f(&bb->upper[0],
869 gmx_simd4_max_f(gmx_simd4_load_f(&bbj[0].upper[0]),
870 gmx_simd4_load_f(&bbj[1].upper[0])));
875 for (i = 0; i < NNBSBB_C; i++)
877 bb->lower[i] = min(bbj[0].lower[i], bbj[1].lower[i]);
878 bb->upper[i] = max(bbj[0].upper[i], bbj[1].upper[i]);
884 #ifdef NBNXN_SEARCH_BB_SIMD4
886 /* Coordinate order xyz, bb order xxxxyyyyzzzz */
887 static void calc_bounding_box_xxxx(int na, int stride, const real *x, float *bb)
890 real xl, xh, yl, yh, zl, zh;
900 for (j = 1; j < na; j++)
902 xl = min(xl, x[i+XX]);
903 xh = max(xh, x[i+XX]);
904 yl = min(yl, x[i+YY]);
905 yh = max(yh, x[i+YY]);
906 zl = min(zl, x[i+ZZ]);
907 zh = max(zh, x[i+ZZ]);
910 /* Note: possible double to float conversion here */
911 bb[0*STRIDE_PBB] = R2F_D(xl);
912 bb[1*STRIDE_PBB] = R2F_D(yl);
913 bb[2*STRIDE_PBB] = R2F_D(zl);
914 bb[3*STRIDE_PBB] = R2F_U(xh);
915 bb[4*STRIDE_PBB] = R2F_U(yh);
916 bb[5*STRIDE_PBB] = R2F_U(zh);
919 #endif /* NBNXN_SEARCH_BB_SIMD4 */
921 #ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
923 /* Coordinate order xyz?, bb order xyz0 */
924 static void calc_bounding_box_simd4(int na, const float *x, nbnxn_bb_t *bb)
926 gmx_simd4_float_t bb_0_S, bb_1_S;
927 gmx_simd4_float_t x_S;
931 bb_0_S = gmx_simd4_load_f(x);
934 for (i = 1; i < na; i++)
936 x_S = gmx_simd4_load_f(x+i*NNBSBB_C);
937 bb_0_S = gmx_simd4_min_f(bb_0_S, x_S);
938 bb_1_S = gmx_simd4_max_f(bb_1_S, x_S);
941 gmx_simd4_store_f(&bb->lower[0], bb_0_S);
942 gmx_simd4_store_f(&bb->upper[0], bb_1_S);
945 /* Coordinate order xyz?, bb order xxxxyyyyzzzz */
946 static void calc_bounding_box_xxxx_simd4(int na, const float *x,
947 nbnxn_bb_t *bb_work_aligned,
950 calc_bounding_box_simd4(na, x, bb_work_aligned);
952 bb[0*STRIDE_PBB] = bb_work_aligned->lower[BB_X];
953 bb[1*STRIDE_PBB] = bb_work_aligned->lower[BB_Y];
954 bb[2*STRIDE_PBB] = bb_work_aligned->lower[BB_Z];
955 bb[3*STRIDE_PBB] = bb_work_aligned->upper[BB_X];
956 bb[4*STRIDE_PBB] = bb_work_aligned->upper[BB_Y];
957 bb[5*STRIDE_PBB] = bb_work_aligned->upper[BB_Z];
960 #endif /* NBNXN_SEARCH_SIMD4_FLOAT_X_BB */
963 /* Combines pairs of consecutive bounding boxes */
964 static void combine_bounding_box_pairs(nbnxn_grid_t *grid, const nbnxn_bb_t *bb)
966 int i, j, sc2, nc2, c2;
968 for (i = 0; i < grid->ncx*grid->ncy; i++)
970 /* Starting bb in a column is expected to be 2-aligned */
971 sc2 = grid->cxy_ind[i]>>1;
972 /* For odd numbers skip the last bb here */
973 nc2 = (grid->cxy_na[i]+3)>>(2+1);
974 for (c2 = sc2; c2 < sc2+nc2; c2++)
976 #ifdef NBNXN_SEARCH_BB_SIMD4
977 gmx_simd4_float_t min_S, max_S;
979 min_S = gmx_simd4_min_f(gmx_simd4_load_f(&bb[c2*2+0].lower[0]),
980 gmx_simd4_load_f(&bb[c2*2+1].lower[0]));
981 max_S = gmx_simd4_max_f(gmx_simd4_load_f(&bb[c2*2+0].upper[0]),
982 gmx_simd4_load_f(&bb[c2*2+1].upper[0]));
983 gmx_simd4_store_f(&grid->bbj[c2].lower[0], min_S);
984 gmx_simd4_store_f(&grid->bbj[c2].upper[0], max_S);
986 for (j = 0; j < NNBSBB_C; j++)
988 grid->bbj[c2].lower[j] = min(bb[c2*2+0].lower[j],
989 bb[c2*2+1].lower[j]);
990 grid->bbj[c2].upper[j] = max(bb[c2*2+0].upper[j],
991 bb[c2*2+1].upper[j]);
995 if (((grid->cxy_na[i]+3)>>2) & 1)
997 /* The bb count in this column is odd: duplicate the last bb */
998 for (j = 0; j < NNBSBB_C; j++)
1000 grid->bbj[c2].lower[j] = bb[c2*2].lower[j];
1001 grid->bbj[c2].upper[j] = bb[c2*2].upper[j];
1008 /* Prints the average bb size, used for debug output */
1009 static void print_bbsizes_simple(FILE *fp,
1010 const nbnxn_grid_t *grid)
1016 for (c = 0; c < grid->nc; c++)
1018 for (d = 0; d < DIM; d++)
1020 ba[d] += grid->bb[c].upper[d] - grid->bb[c].lower[d];
1023 dsvmul(1.0/grid->nc, ba, ba);
1025 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",
1028 grid->atom_density > 0 ?
1029 grid->na_c/(grid->atom_density*grid->sx*grid->sy) : 0.0,
1030 ba[XX], ba[YY], ba[ZZ],
1033 grid->atom_density > 0 ?
1034 ba[ZZ]/(grid->na_c/(grid->atom_density*grid->sx*grid->sy)) : 0.0);
1037 /* Prints the average bb size, used for debug output */
1038 static void print_bbsizes_supersub(FILE *fp,
1039 const nbnxn_grid_t *grid)
1046 for (c = 0; c < grid->nc; c++)
1049 for (s = 0; s < grid->nsubc[c]; s += STRIDE_PBB)
1053 cs_w = (c*GPU_NSUBCELL + s)/STRIDE_PBB;
1054 for (i = 0; i < STRIDE_PBB; i++)
1056 for (d = 0; d < DIM; d++)
1059 grid->pbb[cs_w*NNBSBB_XXXX+(DIM+d)*STRIDE_PBB+i] -
1060 grid->pbb[cs_w*NNBSBB_XXXX+ d *STRIDE_PBB+i];
1065 for (s = 0; s < grid->nsubc[c]; s++)
1069 cs = c*GPU_NSUBCELL + s;
1070 for (d = 0; d < DIM; d++)
1072 ba[d] += grid->bb[cs].upper[d] - grid->bb[cs].lower[d];
1076 ns += grid->nsubc[c];
1078 dsvmul(1.0/ns, ba, ba);
1080 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",
1081 grid->sx/GPU_NSUBCELL_X,
1082 grid->sy/GPU_NSUBCELL_Y,
1083 grid->atom_density > 0 ?
1084 grid->na_sc/(grid->atom_density*grid->sx*grid->sy*GPU_NSUBCELL_Z) : 0.0,
1085 ba[XX], ba[YY], ba[ZZ],
1086 ba[XX]*GPU_NSUBCELL_X/grid->sx,
1087 ba[YY]*GPU_NSUBCELL_Y/grid->sy,
1088 grid->atom_density > 0 ?
1089 ba[ZZ]/(grid->na_sc/(grid->atom_density*grid->sx*grid->sy*GPU_NSUBCELL_Z)) : 0.0);
1092 /* Potentially sorts atoms on LJ coefficients !=0 and ==0.
1093 * Also sets interaction flags.
1095 void sort_on_lj(int na_c,
1096 int a0, int a1, const int *atinfo,
1100 int subc, s, a, n1, n2, a_lj_max, i, j;
1101 int sort1[NBNXN_NA_SC_MAX/GPU_NSUBCELL];
1102 int sort2[NBNXN_NA_SC_MAX/GPU_NSUBCELL];
1103 gmx_bool haveQ, bFEP;
1108 for (s = a0; s < a1; s += na_c)
1110 /* Make lists for this (sub-)cell on atoms with and without LJ */
1115 for (a = s; a < min(s+na_c, a1); a++)
1117 haveQ = haveQ || GET_CGINFO_HAS_Q(atinfo[order[a]]);
1119 if (GET_CGINFO_HAS_VDW(atinfo[order[a]]))
1121 sort1[n1++] = order[a];
1126 sort2[n2++] = order[a];
1130 /* If we don't have atoms with LJ, there's nothing to sort */
1133 *flags |= NBNXN_CI_DO_LJ(subc);
1137 /* Only sort when strictly necessary. Ordering particles
1138 * Ordering particles can lead to less accurate summation
1139 * due to rounding, both for LJ and Coulomb interactions.
1141 if (2*(a_lj_max - s) >= na_c)
1143 for (i = 0; i < n1; i++)
1145 order[a0+i] = sort1[i];
1147 for (j = 0; j < n2; j++)
1149 order[a0+n1+j] = sort2[j];
1153 *flags |= NBNXN_CI_HALF_LJ(subc);
1158 *flags |= NBNXN_CI_DO_COUL(subc);
1164 /* Fill a pair search cell with atoms.
1165 * Potentially sorts atoms and sets the interaction flags.
1167 void fill_cell(const nbnxn_search_t nbs,
1169 nbnxn_atomdata_t *nbat,
1173 int sx, int sy, int sz,
1174 nbnxn_bb_t gmx_unused *bb_work_aligned)
1187 sort_on_lj(grid->na_c, a0, a1, atinfo, nbs->a,
1188 grid->flags+(a0>>grid->na_c_2log)-grid->cell0);
1193 /* Set the fep flag for perturbed atoms in this (sub-)cell */
1196 /* The grid-local cluster/(sub-)cell index */
1197 c = (a0 >> grid->na_c_2log) - grid->cell0*(grid->bSimple ? 1 : GPU_NSUBCELL);
1199 for (at = a0; at < a1; at++)
1201 if (nbs->a[at] >= 0 && GET_CGINFO_FEP(atinfo[nbs->a[at]]))
1203 grid->fep[c] |= (1 << (at - a0));
1208 /* Now we have sorted the atoms, set the cell indices */
1209 for (a = a0; a < a1; a++)
1211 nbs->cell[nbs->a[a]] = a;
1214 copy_rvec_to_nbat_real(nbs->a+a0, a1-a0, grid->na_c, x,
1215 nbat->XFormat, nbat->x, a0,
1218 if (nbat->XFormat == nbatX4)
1220 /* Store the bounding boxes as xyz.xyz. */
1221 offset = (a0 - grid->cell0*grid->na_sc) >> grid->na_c_2log;
1222 bb_ptr = grid->bb + offset;
1224 #if defined GMX_NBNXN_SIMD && GMX_SIMD_REAL_WIDTH == 2
1225 if (2*grid->na_cj == grid->na_c)
1227 calc_bounding_box_x_x4_halves(na, nbat->x+X4_IND_A(a0), bb_ptr,
1228 grid->bbj+offset*2);
1233 calc_bounding_box_x_x4(na, nbat->x+X4_IND_A(a0), bb_ptr);
1236 else if (nbat->XFormat == nbatX8)
1238 /* Store the bounding boxes as xyz.xyz. */
1239 offset = (a0 - grid->cell0*grid->na_sc) >> grid->na_c_2log;
1240 bb_ptr = grid->bb + offset;
1242 calc_bounding_box_x_x8(na, nbat->x+X8_IND_A(a0), bb_ptr);
1245 else if (!grid->bSimple)
1247 /* Store the bounding boxes in a format convenient
1248 * for SIMD4 calculations: xxxxyyyyzzzz...
1252 ((a0-grid->cell0*grid->na_sc)>>(grid->na_c_2log+STRIDE_PBB_2LOG))*NNBSBB_XXXX +
1253 (((a0-grid->cell0*grid->na_sc)>>grid->na_c_2log) & (STRIDE_PBB-1));
1255 #ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
1256 if (nbat->XFormat == nbatXYZQ)
1258 calc_bounding_box_xxxx_simd4(na, nbat->x+a0*nbat->xstride,
1259 bb_work_aligned, pbb_ptr);
1264 calc_bounding_box_xxxx(na, nbat->xstride, nbat->x+a0*nbat->xstride,
1269 fprintf(debug, "%2d %2d %2d bb %5.2f %5.2f %5.2f %5.2f %5.2f %5.2f\n",
1271 pbb_ptr[0*STRIDE_PBB], pbb_ptr[3*STRIDE_PBB],
1272 pbb_ptr[1*STRIDE_PBB], pbb_ptr[4*STRIDE_PBB],
1273 pbb_ptr[2*STRIDE_PBB], pbb_ptr[5*STRIDE_PBB]);
1279 /* Store the bounding boxes as xyz.xyz. */
1280 bb_ptr = grid->bb+((a0-grid->cell0*grid->na_sc)>>grid->na_c_2log);
1282 calc_bounding_box(na, nbat->xstride, nbat->x+a0*nbat->xstride,
1288 bbo = (a0 - grid->cell0*grid->na_sc)/grid->na_c;
1289 fprintf(debug, "%2d %2d %2d bb %5.2f %5.2f %5.2f %5.2f %5.2f %5.2f\n",
1291 grid->bb[bbo].lower[BB_X],
1292 grid->bb[bbo].lower[BB_Y],
1293 grid->bb[bbo].lower[BB_Z],
1294 grid->bb[bbo].upper[BB_X],
1295 grid->bb[bbo].upper[BB_Y],
1296 grid->bb[bbo].upper[BB_Z]);
1301 /* Spatially sort the atoms within one grid column */
1302 static void sort_columns_simple(const nbnxn_search_t nbs,
1308 nbnxn_atomdata_t *nbat,
1309 int cxy_start, int cxy_end,
1313 int cx, cy, cz, ncz, cfilled, c;
1314 int na, ash, ind, a;
1319 fprintf(debug, "cell0 %d sorting columns %d - %d, atoms %d - %d\n",
1320 grid->cell0, cxy_start, cxy_end, a0, a1);
1323 /* Sort the atoms within each x,y column in 3 dimensions */
1324 for (cxy = cxy_start; cxy < cxy_end; cxy++)
1327 cy = cxy - cx*grid->ncy;
1329 na = grid->cxy_na[cxy];
1330 ncz = grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy];
1331 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1333 /* Sort the atoms within each x,y column on z coordinate */
1334 sort_atoms(ZZ, FALSE, dd_zone,
1337 1.0/grid->size[ZZ], ncz*grid->na_sc,
1340 /* Fill the ncz cells in this column */
1341 cfilled = grid->cxy_ind[cxy];
1342 for (cz = 0; cz < ncz; cz++)
1344 c = grid->cxy_ind[cxy] + cz;
1346 ash_c = ash + cz*grid->na_sc;
1347 na_c = min(grid->na_sc, na-(ash_c-ash));
1349 fill_cell(nbs, grid, nbat,
1350 ash_c, ash_c+na_c, atinfo, x,
1351 grid->na_sc*cx + (dd_zone >> 2),
1352 grid->na_sc*cy + (dd_zone & 3),
1356 /* This copy to bbcz is not really necessary.
1357 * But it allows to use the same grid search code
1358 * for the simple and supersub cell setups.
1364 grid->bbcz[c*NNBSBB_D ] = grid->bb[cfilled].lower[BB_Z];
1365 grid->bbcz[c*NNBSBB_D+1] = grid->bb[cfilled].upper[BB_Z];
1368 /* Set the unused atom indices to -1 */
1369 for (ind = na; ind < ncz*grid->na_sc; ind++)
1371 nbs->a[ash+ind] = -1;
1376 /* Spatially sort the atoms within one grid column */
1377 static void sort_columns_supersub(const nbnxn_search_t nbs,
1383 nbnxn_atomdata_t *nbat,
1384 int cxy_start, int cxy_end,
1388 int cx, cy, cz = -1, c = -1, ncz;
1389 int na, ash, na_c, ind, a;
1390 int subdiv_z, sub_z, na_z, ash_z;
1391 int subdiv_y, sub_y, na_y, ash_y;
1392 int subdiv_x, sub_x, na_x, ash_x;
1394 nbnxn_bb_t bb_work_array[2], *bb_work_aligned;
1396 bb_work_aligned = (nbnxn_bb_t *)(((size_t)(bb_work_array+1)) & (~((size_t)15)));
1400 fprintf(debug, "cell0 %d sorting columns %d - %d, atoms %d - %d\n",
1401 grid->cell0, cxy_start, cxy_end, a0, a1);
1404 subdiv_x = grid->na_c;
1405 subdiv_y = GPU_NSUBCELL_X*subdiv_x;
1406 subdiv_z = GPU_NSUBCELL_Y*subdiv_y;
1408 /* Sort the atoms within each x,y column in 3 dimensions */
1409 for (cxy = cxy_start; cxy < cxy_end; cxy++)
1412 cy = cxy - cx*grid->ncy;
1414 na = grid->cxy_na[cxy];
1415 ncz = grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy];
1416 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1418 /* Sort the atoms within each x,y column on z coordinate */
1419 sort_atoms(ZZ, FALSE, dd_zone,
1422 1.0/grid->size[ZZ], ncz*grid->na_sc,
1425 /* This loop goes over the supercells and subcells along z at once */
1426 for (sub_z = 0; sub_z < ncz*GPU_NSUBCELL_Z; sub_z++)
1428 ash_z = ash + sub_z*subdiv_z;
1429 na_z = min(subdiv_z, na-(ash_z-ash));
1431 /* We have already sorted on z */
1433 if (sub_z % GPU_NSUBCELL_Z == 0)
1435 cz = sub_z/GPU_NSUBCELL_Z;
1436 c = grid->cxy_ind[cxy] + cz;
1438 /* The number of atoms in this supercell */
1439 na_c = min(grid->na_sc, na-(ash_z-ash));
1441 grid->nsubc[c] = min(GPU_NSUBCELL, (na_c+grid->na_c-1)/grid->na_c);
1443 /* Store the z-boundaries of the super cell */
1444 grid->bbcz[c*NNBSBB_D ] = x[nbs->a[ash_z]][ZZ];
1445 grid->bbcz[c*NNBSBB_D+1] = x[nbs->a[ash_z+na_c-1]][ZZ];
1448 #if GPU_NSUBCELL_Y > 1
1449 /* Sort the atoms along y */
1450 sort_atoms(YY, (sub_z & 1), dd_zone,
1451 nbs->a+ash_z, na_z, x,
1452 grid->c0[YY]+cy*grid->sy,
1453 grid->inv_sy, subdiv_z,
1457 for (sub_y = 0; sub_y < GPU_NSUBCELL_Y; sub_y++)
1459 ash_y = ash_z + sub_y*subdiv_y;
1460 na_y = min(subdiv_y, na-(ash_y-ash));
1462 #if GPU_NSUBCELL_X > 1
1463 /* Sort the atoms along x */
1464 sort_atoms(XX, ((cz*GPU_NSUBCELL_Y + sub_y) & 1), dd_zone,
1465 nbs->a+ash_y, na_y, x,
1466 grid->c0[XX]+cx*grid->sx,
1467 grid->inv_sx, subdiv_y,
1471 for (sub_x = 0; sub_x < GPU_NSUBCELL_X; sub_x++)
1473 ash_x = ash_y + sub_x*subdiv_x;
1474 na_x = min(subdiv_x, na-(ash_x-ash));
1476 fill_cell(nbs, grid, nbat,
1477 ash_x, ash_x+na_x, atinfo, x,
1478 grid->na_c*(cx*GPU_NSUBCELL_X+sub_x) + (dd_zone >> 2),
1479 grid->na_c*(cy*GPU_NSUBCELL_Y+sub_y) + (dd_zone & 3),
1486 /* Set the unused atom indices to -1 */
1487 for (ind = na; ind < ncz*grid->na_sc; ind++)
1489 nbs->a[ash+ind] = -1;
1494 /* Determine in which grid column atoms should go */
1495 static void calc_column_indices(nbnxn_grid_t *grid,
1498 int dd_zone, const int *move,
1499 int thread, int nthread,
1506 /* We add one extra cell for particles which moved during DD */
1507 for (i = 0; i < grid->ncx*grid->ncy+1; i++)
1512 n0 = a0 + (int)((thread+0)*(a1 - a0))/nthread;
1513 n1 = a0 + (int)((thread+1)*(a1 - a0))/nthread;
1517 for (i = n0; i < n1; i++)
1519 if (move == NULL || move[i] >= 0)
1521 /* We need to be careful with rounding,
1522 * particles might be a few bits outside the local zone.
1523 * The int cast takes care of the lower bound,
1524 * we will explicitly take care of the upper bound.
1526 cx = (int)((x[i][XX] - grid->c0[XX])*grid->inv_sx);
1527 cy = (int)((x[i][YY] - grid->c0[YY])*grid->inv_sy);
1530 if (cx < 0 || cx > grid->ncx ||
1531 cy < 0 || cy > grid->ncy)
1534 "grid cell cx %d cy %d out of range (max %d %d)\n"
1535 "atom %f %f %f, grid->c0 %f %f",
1536 cx, cy, grid->ncx, grid->ncy,
1537 x[i][XX], x[i][YY], x[i][ZZ], grid->c0[XX], grid->c0[YY]);
1540 /* Take care of potential rouding issues */
1541 cx = min(cx, grid->ncx - 1);
1542 cy = min(cy, grid->ncy - 1);
1544 /* For the moment cell will contain only the, grid local,
1545 * x and y indices, not z.
1547 cell[i] = cx*grid->ncy + cy;
1551 /* Put this moved particle after the end of the grid,
1552 * so we can process it later without using conditionals.
1554 cell[i] = grid->ncx*grid->ncy;
1563 for (i = n0; i < n1; i++)
1565 cx = (int)((x[i][XX] - grid->c0[XX])*grid->inv_sx);
1566 cy = (int)((x[i][YY] - grid->c0[YY])*grid->inv_sy);
1568 /* For non-home zones there could be particles outside
1569 * the non-bonded cut-off range, which have been communicated
1570 * for bonded interactions only. For the result it doesn't
1571 * matter where these end up on the grid. For performance
1572 * we put them in an extra row at the border.
1575 cx = min(cx, grid->ncx - 1);
1577 cy = min(cy, grid->ncy - 1);
1579 /* For the moment cell will contain only the, grid local,
1580 * x and y indices, not z.
1582 cell[i] = cx*grid->ncy + cy;
1589 /* Determine in which grid cells the atoms should go */
1590 static void calc_cell_indices(const nbnxn_search_t nbs,
1597 nbnxn_atomdata_t *nbat)
1600 int cx, cy, cxy, ncz_max, ncz;
1601 int nthread, thread;
1602 int *cxy_na, cxy_na_i;
1604 nthread = gmx_omp_nthreads_get(emntPairsearch);
1606 #pragma omp parallel for num_threads(nthread) schedule(static)
1607 for (thread = 0; thread < nthread; thread++)
1609 calc_column_indices(grid, a0, a1, x, dd_zone, move, thread, nthread,
1610 nbs->cell, nbs->work[thread].cxy_na);
1613 /* Make the cell index as a function of x and y */
1616 grid->cxy_ind[0] = 0;
1617 for (i = 0; i < grid->ncx*grid->ncy+1; i++)
1619 /* We set ncz_max at the beginning of the loop iso at the end
1620 * to skip i=grid->ncx*grid->ncy which are moved particles
1621 * that do not need to be ordered on the grid.
1627 cxy_na_i = nbs->work[0].cxy_na[i];
1628 for (thread = 1; thread < nthread; thread++)
1630 cxy_na_i += nbs->work[thread].cxy_na[i];
1632 ncz = (cxy_na_i + grid->na_sc - 1)/grid->na_sc;
1633 if (nbat->XFormat == nbatX8)
1635 /* Make the number of cell a multiple of 2 */
1636 ncz = (ncz + 1) & ~1;
1638 grid->cxy_ind[i+1] = grid->cxy_ind[i] + ncz;
1639 /* Clear cxy_na, so we can reuse the array below */
1640 grid->cxy_na[i] = 0;
1642 grid->nc = grid->cxy_ind[grid->ncx*grid->ncy] - grid->cxy_ind[0];
1644 nbat->natoms = (grid->cell0 + grid->nc)*grid->na_sc;
1648 fprintf(debug, "ns na_sc %d na_c %d super-cells: %d x %d y %d z %.1f maxz %d\n",
1649 grid->na_sc, grid->na_c, grid->nc,
1650 grid->ncx, grid->ncy, grid->nc/((double)(grid->ncx*grid->ncy)),
1655 for (cy = 0; cy < grid->ncy; cy++)
1657 for (cx = 0; cx < grid->ncx; cx++)
1659 fprintf(debug, " %2d", grid->cxy_ind[i+1]-grid->cxy_ind[i]);
1662 fprintf(debug, "\n");
1667 /* Make sure the work array for sorting is large enough */
1668 if (ncz_max*grid->na_sc*SGSF > nbs->work[0].sort_work_nalloc)
1670 for (thread = 0; thread < nbs->nthread_max; thread++)
1672 nbs->work[thread].sort_work_nalloc =
1673 over_alloc_large(ncz_max*grid->na_sc*SGSF);
1674 srenew(nbs->work[thread].sort_work,
1675 nbs->work[thread].sort_work_nalloc);
1676 /* When not in use, all elements should be -1 */
1677 for (i = 0; i < nbs->work[thread].sort_work_nalloc; i++)
1679 nbs->work[thread].sort_work[i] = -1;
1684 /* Now we know the dimensions we can fill the grid.
1685 * This is the first, unsorted fill. We sort the columns after this.
1687 for (i = a0; i < a1; i++)
1689 /* At this point nbs->cell contains the local grid x,y indices */
1691 nbs->a[(grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc + grid->cxy_na[cxy]++] = i;
1696 /* Set the cell indices for the moved particles */
1697 n0 = grid->nc*grid->na_sc;
1698 n1 = grid->nc*grid->na_sc+grid->cxy_na[grid->ncx*grid->ncy];
1701 for (i = n0; i < n1; i++)
1703 nbs->cell[nbs->a[i]] = i;
1708 /* Sort the super-cell columns along z into the sub-cells. */
1709 #pragma omp parallel for num_threads(nthread) schedule(static)
1710 for (thread = 0; thread < nthread; thread++)
1714 sort_columns_simple(nbs, dd_zone, grid, a0, a1, atinfo, x, nbat,
1715 ((thread+0)*grid->ncx*grid->ncy)/nthread,
1716 ((thread+1)*grid->ncx*grid->ncy)/nthread,
1717 nbs->work[thread].sort_work);
1721 sort_columns_supersub(nbs, dd_zone, grid, a0, a1, atinfo, x, nbat,
1722 ((thread+0)*grid->ncx*grid->ncy)/nthread,
1723 ((thread+1)*grid->ncx*grid->ncy)/nthread,
1724 nbs->work[thread].sort_work);
1728 if (grid->bSimple && nbat->XFormat == nbatX8)
1730 combine_bounding_box_pairs(grid, grid->bb);
1735 grid->nsubc_tot = 0;
1736 for (i = 0; i < grid->nc; i++)
1738 grid->nsubc_tot += grid->nsubc[i];
1746 print_bbsizes_simple(debug, grid);
1750 fprintf(debug, "ns non-zero sub-cells: %d average atoms %.2f\n",
1751 grid->nsubc_tot, (a1-a0)/(double)grid->nsubc_tot);
1753 print_bbsizes_supersub(debug, grid);
1758 static void init_buffer_flags(nbnxn_buffer_flags_t *flags,
1763 flags->nflag = (natoms + NBNXN_BUFFERFLAG_SIZE - 1)/NBNXN_BUFFERFLAG_SIZE;
1764 if (flags->nflag > flags->flag_nalloc)
1766 flags->flag_nalloc = over_alloc_large(flags->nflag);
1767 srenew(flags->flag, flags->flag_nalloc);
1769 for (b = 0; b < flags->nflag; b++)
1771 bitmask_clear(&(flags->flag[b]));
1775 /* Sets up a grid and puts the atoms on the grid.
1776 * This function only operates on one domain of the domain decompostion.
1777 * Note that without domain decomposition there is only one domain.
1779 void nbnxn_put_on_grid(nbnxn_search_t nbs,
1780 int ePBC, matrix box,
1782 rvec corner0, rvec corner1,
1787 int nmoved, int *move,
1789 nbnxn_atomdata_t *nbat)
1793 int nc_max_grid, nc_max;
1795 grid = &nbs->grid[dd_zone];
1797 nbs_cycle_start(&nbs->cc[enbsCCgrid]);
1799 grid->bSimple = nbnxn_kernel_pairlist_simple(nb_kernel_type);
1801 grid->na_c = nbnxn_kernel_to_ci_size(nb_kernel_type);
1802 grid->na_cj = nbnxn_kernel_to_cj_size(nb_kernel_type);
1803 grid->na_sc = (grid->bSimple ? 1 : GPU_NSUBCELL)*grid->na_c;
1804 grid->na_c_2log = get_2log(grid->na_c);
1806 nbat->na_c = grid->na_c;
1815 (nbs->grid[dd_zone-1].cell0 + nbs->grid[dd_zone-1].nc)*
1816 nbs->grid[dd_zone-1].na_sc/grid->na_sc;
1824 copy_mat(box, nbs->box);
1826 /* Avoid zero density */
1827 if (atom_density > 0)
1829 grid->atom_density = atom_density;
1833 grid->atom_density = grid_atom_density(n-nmoved, corner0, corner1);
1838 nbs->natoms_local = a1 - nmoved;
1839 /* We assume that nbnxn_put_on_grid is called first
1840 * for the local atoms (dd_zone=0).
1842 nbs->natoms_nonlocal = a1 - nmoved;
1846 fprintf(debug, "natoms_local = %5d atom_density = %5.1f\n",
1847 nbs->natoms_local, grid->atom_density);
1852 nbs->natoms_nonlocal = max(nbs->natoms_nonlocal, a1);
1855 /* We always use the home zone (grid[0]) for setting the cell size,
1856 * since determining densities for non-local zones is difficult.
1858 nc_max_grid = set_grid_size_xy(nbs, grid,
1859 dd_zone, n-nmoved, corner0, corner1,
1860 nbs->grid[0].atom_density);
1862 nc_max = grid->cell0 + nc_max_grid;
1864 if (a1 > nbs->cell_nalloc)
1866 nbs->cell_nalloc = over_alloc_large(a1);
1867 srenew(nbs->cell, nbs->cell_nalloc);
1870 /* To avoid conditionals we store the moved particles at the end of a,
1871 * make sure we have enough space.
1873 if (nc_max*grid->na_sc + nmoved > nbs->a_nalloc)
1875 nbs->a_nalloc = over_alloc_large(nc_max*grid->na_sc + nmoved);
1876 srenew(nbs->a, nbs->a_nalloc);
1879 /* We need padding up to a multiple of the buffer flag size: simply add */
1880 if (nc_max*grid->na_sc + NBNXN_BUFFERFLAG_SIZE > nbat->nalloc)
1882 nbnxn_atomdata_realloc(nbat, nc_max*grid->na_sc+NBNXN_BUFFERFLAG_SIZE);
1885 calc_cell_indices(nbs, dd_zone, grid, a0, a1, atinfo, x, move, nbat);
1889 nbat->natoms_local = nbat->natoms;
1892 nbs_cycle_stop(&nbs->cc[enbsCCgrid]);
1895 /* Calls nbnxn_put_on_grid for all non-local domains */
1896 void nbnxn_put_on_grid_nonlocal(nbnxn_search_t nbs,
1897 const gmx_domdec_zones_t *zones,
1901 nbnxn_atomdata_t *nbat)
1906 for (zone = 1; zone < zones->n; zone++)
1908 for (d = 0; d < DIM; d++)
1910 c0[d] = zones->size[zone].bb_x0[d];
1911 c1[d] = zones->size[zone].bb_x1[d];
1914 nbnxn_put_on_grid(nbs, nbs->ePBC, NULL,
1916 zones->cg_range[zone],
1917 zones->cg_range[zone+1],
1927 /* Add simple grid type information to the local super/sub grid */
1928 void nbnxn_grid_add_simple(nbnxn_search_t nbs,
1929 nbnxn_atomdata_t *nbat)
1935 int nthreads gmx_unused;
1937 grid = &nbs->grid[0];
1941 gmx_incons("nbnxn_grid_simple called with a simple grid");
1944 ncd = grid->na_sc/NBNXN_CPU_CLUSTER_I_SIZE;
1946 if (grid->nc*ncd > grid->nc_nalloc_simple)
1948 grid->nc_nalloc_simple = over_alloc_large(grid->nc*ncd);
1949 srenew(grid->bbcz_simple, grid->nc_nalloc_simple*NNBSBB_D);
1950 srenew(grid->bb_simple, grid->nc_nalloc_simple);
1951 srenew(grid->flags_simple, grid->nc_nalloc_simple);
1954 sfree_aligned(grid->bbj);
1955 snew_aligned(grid->bbj, grid->nc_nalloc_simple/2, 16);
1959 bbcz = grid->bbcz_simple;
1960 bb = grid->bb_simple;
1962 nthreads = gmx_omp_nthreads_get(emntPairsearch);
1963 #pragma omp parallel for num_threads(nthreads) schedule(static)
1964 for (sc = 0; sc < grid->nc; sc++)
1968 for (c = 0; c < ncd; c++)
1972 na = NBNXN_CPU_CLUSTER_I_SIZE;
1974 nbat->type[tx*NBNXN_CPU_CLUSTER_I_SIZE+na-1] == nbat->ntype-1)
1981 switch (nbat->XFormat)
1984 /* PACK_X4==NBNXN_CPU_CLUSTER_I_SIZE, so this is simple */
1985 calc_bounding_box_x_x4(na, nbat->x+tx*STRIDE_P4,
1989 /* PACK_X8>NBNXN_CPU_CLUSTER_I_SIZE, more complicated */
1990 calc_bounding_box_x_x8(na, nbat->x+X8_IND_A(tx*NBNXN_CPU_CLUSTER_I_SIZE),
1994 calc_bounding_box(na, nbat->xstride,
1995 nbat->x+tx*NBNXN_CPU_CLUSTER_I_SIZE*nbat->xstride,
1999 bbcz[tx*NNBSBB_D+0] = bb[tx].lower[BB_Z];
2000 bbcz[tx*NNBSBB_D+1] = bb[tx].upper[BB_Z];
2002 /* No interaction optimization yet here */
2003 grid->flags_simple[tx] = NBNXN_CI_DO_LJ(0) | NBNXN_CI_DO_COUL(0);
2007 grid->flags_simple[tx] = 0;
2012 if (grid->bSimple && nbat->XFormat == nbatX8)
2014 combine_bounding_box_pairs(grid, grid->bb_simple);
2018 void nbnxn_get_ncells(nbnxn_search_t nbs, int *ncx, int *ncy)
2020 *ncx = nbs->grid[0].ncx;
2021 *ncy = nbs->grid[0].ncy;
2024 void nbnxn_get_atomorder(nbnxn_search_t nbs, int **a, int *n)
2026 const nbnxn_grid_t *grid;
2028 grid = &nbs->grid[0];
2030 /* Return the atom order for the home cell (index 0) */
2033 *n = grid->cxy_ind[grid->ncx*grid->ncy]*grid->na_sc;
2036 void nbnxn_set_atomorder(nbnxn_search_t nbs)
2039 int ao, cx, cy, cxy, cz, j;
2041 /* Set the atom order for the home cell (index 0) */
2042 grid = &nbs->grid[0];
2045 for (cx = 0; cx < grid->ncx; cx++)
2047 for (cy = 0; cy < grid->ncy; cy++)
2049 cxy = cx*grid->ncy + cy;
2050 j = grid->cxy_ind[cxy]*grid->na_sc;
2051 for (cz = 0; cz < grid->cxy_na[cxy]; cz++)
2062 /* Determines the cell range along one dimension that
2063 * the bounding box b0 - b1 sees.
2065 static void get_cell_range(real b0, real b1,
2066 int nc, real c0, real s, real invs,
2067 real d2, real r2, int *cf, int *cl)
2069 *cf = max((int)((b0 - c0)*invs), 0);
2071 while (*cf > 0 && d2 + sqr((b0 - c0) - (*cf-1+1)*s) < r2)
2076 *cl = min((int)((b1 - c0)*invs), nc-1);
2077 while (*cl < nc-1 && d2 + sqr((*cl+1)*s - (b1 - c0)) < r2)
2083 /* Reference code calculating the distance^2 between two bounding boxes */
2084 static float box_dist2(float bx0, float bx1, float by0,
2085 float by1, float bz0, float bz1,
2086 const nbnxn_bb_t *bb)
2089 float dl, dh, dm, dm0;
2093 dl = bx0 - bb->upper[BB_X];
2094 dh = bb->lower[BB_X] - bx1;
2099 dl = by0 - bb->upper[BB_Y];
2100 dh = bb->lower[BB_Y] - by1;
2105 dl = bz0 - bb->upper[BB_Z];
2106 dh = bb->lower[BB_Z] - bz1;
2114 /* Plain C code calculating the distance^2 between two bounding boxes */
2115 static float subc_bb_dist2(int si, const nbnxn_bb_t *bb_i_ci,
2116 int csj, const nbnxn_bb_t *bb_j_all)
2118 const nbnxn_bb_t *bb_i, *bb_j;
2120 float dl, dh, dm, dm0;
2122 bb_i = bb_i_ci + si;
2123 bb_j = bb_j_all + csj;
2127 dl = bb_i->lower[BB_X] - bb_j->upper[BB_X];
2128 dh = bb_j->lower[BB_X] - bb_i->upper[BB_X];
2133 dl = bb_i->lower[BB_Y] - bb_j->upper[BB_Y];
2134 dh = bb_j->lower[BB_Y] - bb_i->upper[BB_Y];
2139 dl = bb_i->lower[BB_Z] - bb_j->upper[BB_Z];
2140 dh = bb_j->lower[BB_Z] - bb_i->upper[BB_Z];
2148 #ifdef NBNXN_SEARCH_BB_SIMD4
2150 /* 4-wide SIMD code for bb distance for bb format xyz0 */
2151 static float subc_bb_dist2_simd4(int si, const nbnxn_bb_t *bb_i_ci,
2152 int csj, const nbnxn_bb_t *bb_j_all)
2154 gmx_simd4_float_t bb_i_S0, bb_i_S1;
2155 gmx_simd4_float_t bb_j_S0, bb_j_S1;
2156 gmx_simd4_float_t dl_S;
2157 gmx_simd4_float_t dh_S;
2158 gmx_simd4_float_t dm_S;
2159 gmx_simd4_float_t dm0_S;
2161 bb_i_S0 = gmx_simd4_load_f(&bb_i_ci[si].lower[0]);
2162 bb_i_S1 = gmx_simd4_load_f(&bb_i_ci[si].upper[0]);
2163 bb_j_S0 = gmx_simd4_load_f(&bb_j_all[csj].lower[0]);
2164 bb_j_S1 = gmx_simd4_load_f(&bb_j_all[csj].upper[0]);
2166 dl_S = gmx_simd4_sub_f(bb_i_S0, bb_j_S1);
2167 dh_S = gmx_simd4_sub_f(bb_j_S0, bb_i_S1);
2169 dm_S = gmx_simd4_max_f(dl_S, dh_S);
2170 dm0_S = gmx_simd4_max_f(dm_S, gmx_simd4_setzero_f());
2172 return gmx_simd4_dotproduct3_f(dm0_S, dm0_S);
2175 /* Calculate bb bounding distances of bb_i[si,...,si+3] and store them in d2 */
2176 #define SUBC_BB_DIST2_SIMD4_XXXX_INNER(si, bb_i, d2) \
2180 gmx_simd4_float_t dx_0, dy_0, dz_0; \
2181 gmx_simd4_float_t dx_1, dy_1, dz_1; \
2183 gmx_simd4_float_t mx, my, mz; \
2184 gmx_simd4_float_t m0x, m0y, m0z; \
2186 gmx_simd4_float_t d2x, d2y, d2z; \
2187 gmx_simd4_float_t d2s, d2t; \
2189 shi = si*NNBSBB_D*DIM; \
2191 xi_l = gmx_simd4_load_f(bb_i+shi+0*STRIDE_PBB); \
2192 yi_l = gmx_simd4_load_f(bb_i+shi+1*STRIDE_PBB); \
2193 zi_l = gmx_simd4_load_f(bb_i+shi+2*STRIDE_PBB); \
2194 xi_h = gmx_simd4_load_f(bb_i+shi+3*STRIDE_PBB); \
2195 yi_h = gmx_simd4_load_f(bb_i+shi+4*STRIDE_PBB); \
2196 zi_h = gmx_simd4_load_f(bb_i+shi+5*STRIDE_PBB); \
2198 dx_0 = gmx_simd4_sub_f(xi_l, xj_h); \
2199 dy_0 = gmx_simd4_sub_f(yi_l, yj_h); \
2200 dz_0 = gmx_simd4_sub_f(zi_l, zj_h); \
2202 dx_1 = gmx_simd4_sub_f(xj_l, xi_h); \
2203 dy_1 = gmx_simd4_sub_f(yj_l, yi_h); \
2204 dz_1 = gmx_simd4_sub_f(zj_l, zi_h); \
2206 mx = gmx_simd4_max_f(dx_0, dx_1); \
2207 my = gmx_simd4_max_f(dy_0, dy_1); \
2208 mz = gmx_simd4_max_f(dz_0, dz_1); \
2210 m0x = gmx_simd4_max_f(mx, zero); \
2211 m0y = gmx_simd4_max_f(my, zero); \
2212 m0z = gmx_simd4_max_f(mz, zero); \
2214 d2x = gmx_simd4_mul_f(m0x, m0x); \
2215 d2y = gmx_simd4_mul_f(m0y, m0y); \
2216 d2z = gmx_simd4_mul_f(m0z, m0z); \
2218 d2s = gmx_simd4_add_f(d2x, d2y); \
2219 d2t = gmx_simd4_add_f(d2s, d2z); \
2221 gmx_simd4_store_f(d2+si, d2t); \
2224 /* 4-wide SIMD code for nsi bb distances for bb format xxxxyyyyzzzz */
2225 static void subc_bb_dist2_simd4_xxxx(const float *bb_j,
2226 int nsi, const float *bb_i,
2229 gmx_simd4_float_t xj_l, yj_l, zj_l;
2230 gmx_simd4_float_t xj_h, yj_h, zj_h;
2231 gmx_simd4_float_t xi_l, yi_l, zi_l;
2232 gmx_simd4_float_t xi_h, yi_h, zi_h;
2234 gmx_simd4_float_t zero;
2236 zero = gmx_simd4_setzero_f();
2238 xj_l = gmx_simd4_set1_f(bb_j[0*STRIDE_PBB]);
2239 yj_l = gmx_simd4_set1_f(bb_j[1*STRIDE_PBB]);
2240 zj_l = gmx_simd4_set1_f(bb_j[2*STRIDE_PBB]);
2241 xj_h = gmx_simd4_set1_f(bb_j[3*STRIDE_PBB]);
2242 yj_h = gmx_simd4_set1_f(bb_j[4*STRIDE_PBB]);
2243 zj_h = gmx_simd4_set1_f(bb_j[5*STRIDE_PBB]);
2245 /* Here we "loop" over si (0,STRIDE_PBB) from 0 to nsi with step STRIDE_PBB.
2246 * But as we know the number of iterations is 1 or 2, we unroll manually.
2248 SUBC_BB_DIST2_SIMD4_XXXX_INNER(0, bb_i, d2);
2249 if (STRIDE_PBB < nsi)
2251 SUBC_BB_DIST2_SIMD4_XXXX_INNER(STRIDE_PBB, bb_i, d2);
2255 #endif /* NBNXN_SEARCH_BB_SIMD4 */
2257 /* Plain C function which determines if any atom pair between two cells
2258 * is within distance sqrt(rl2).
2260 static gmx_bool subc_in_range_x(int na_c,
2261 int si, const real *x_i,
2262 int csj, int stride, const real *x_j,
2268 for (i = 0; i < na_c; i++)
2270 i0 = (si*na_c + i)*DIM;
2271 for (j = 0; j < na_c; j++)
2273 j0 = (csj*na_c + j)*stride;
2275 d2 = sqr(x_i[i0 ] - x_j[j0 ]) +
2276 sqr(x_i[i0+1] - x_j[j0+1]) +
2277 sqr(x_i[i0+2] - x_j[j0+2]);
2289 #ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
2291 /* 4-wide SIMD function which determines if any atom pair between two cells,
2292 * both with 8 atoms, is within distance sqrt(rl2).
2293 * Using 8-wide AVX is not faster on Intel Sandy Bridge.
2295 static gmx_bool subc_in_range_simd4(int na_c,
2296 int si, const real *x_i,
2297 int csj, int stride, const real *x_j,
2300 gmx_simd4_real_t ix_S0, iy_S0, iz_S0;
2301 gmx_simd4_real_t ix_S1, iy_S1, iz_S1;
2303 gmx_simd4_real_t rc2_S;
2308 rc2_S = gmx_simd4_set1_r(rl2);
2310 dim_stride = NBNXN_GPU_CLUSTER_SIZE/STRIDE_PBB*DIM;
2311 ix_S0 = gmx_simd4_load_r(x_i+(si*dim_stride+0)*STRIDE_PBB);
2312 iy_S0 = gmx_simd4_load_r(x_i+(si*dim_stride+1)*STRIDE_PBB);
2313 iz_S0 = gmx_simd4_load_r(x_i+(si*dim_stride+2)*STRIDE_PBB);
2314 ix_S1 = gmx_simd4_load_r(x_i+(si*dim_stride+3)*STRIDE_PBB);
2315 iy_S1 = gmx_simd4_load_r(x_i+(si*dim_stride+4)*STRIDE_PBB);
2316 iz_S1 = gmx_simd4_load_r(x_i+(si*dim_stride+5)*STRIDE_PBB);
2318 /* We loop from the outer to the inner particles to maximize
2319 * the chance that we find a pair in range quickly and return.
2325 gmx_simd4_real_t jx0_S, jy0_S, jz0_S;
2326 gmx_simd4_real_t jx1_S, jy1_S, jz1_S;
2328 gmx_simd4_real_t dx_S0, dy_S0, dz_S0;
2329 gmx_simd4_real_t dx_S1, dy_S1, dz_S1;
2330 gmx_simd4_real_t dx_S2, dy_S2, dz_S2;
2331 gmx_simd4_real_t dx_S3, dy_S3, dz_S3;
2333 gmx_simd4_real_t rsq_S0;
2334 gmx_simd4_real_t rsq_S1;
2335 gmx_simd4_real_t rsq_S2;
2336 gmx_simd4_real_t rsq_S3;
2338 gmx_simd4_bool_t wco_S0;
2339 gmx_simd4_bool_t wco_S1;
2340 gmx_simd4_bool_t wco_S2;
2341 gmx_simd4_bool_t wco_S3;
2342 gmx_simd4_bool_t wco_any_S01, wco_any_S23, wco_any_S;
2344 jx0_S = gmx_simd4_set1_r(x_j[j0*stride+0]);
2345 jy0_S = gmx_simd4_set1_r(x_j[j0*stride+1]);
2346 jz0_S = gmx_simd4_set1_r(x_j[j0*stride+2]);
2348 jx1_S = gmx_simd4_set1_r(x_j[j1*stride+0]);
2349 jy1_S = gmx_simd4_set1_r(x_j[j1*stride+1]);
2350 jz1_S = gmx_simd4_set1_r(x_j[j1*stride+2]);
2352 /* Calculate distance */
2353 dx_S0 = gmx_simd4_sub_r(ix_S0, jx0_S);
2354 dy_S0 = gmx_simd4_sub_r(iy_S0, jy0_S);
2355 dz_S0 = gmx_simd4_sub_r(iz_S0, jz0_S);
2356 dx_S1 = gmx_simd4_sub_r(ix_S1, jx0_S);
2357 dy_S1 = gmx_simd4_sub_r(iy_S1, jy0_S);
2358 dz_S1 = gmx_simd4_sub_r(iz_S1, jz0_S);
2359 dx_S2 = gmx_simd4_sub_r(ix_S0, jx1_S);
2360 dy_S2 = gmx_simd4_sub_r(iy_S0, jy1_S);
2361 dz_S2 = gmx_simd4_sub_r(iz_S0, jz1_S);
2362 dx_S3 = gmx_simd4_sub_r(ix_S1, jx1_S);
2363 dy_S3 = gmx_simd4_sub_r(iy_S1, jy1_S);
2364 dz_S3 = gmx_simd4_sub_r(iz_S1, jz1_S);
2366 /* rsq = dx*dx+dy*dy+dz*dz */
2367 rsq_S0 = gmx_simd4_calc_rsq_r(dx_S0, dy_S0, dz_S0);
2368 rsq_S1 = gmx_simd4_calc_rsq_r(dx_S1, dy_S1, dz_S1);
2369 rsq_S2 = gmx_simd4_calc_rsq_r(dx_S2, dy_S2, dz_S2);
2370 rsq_S3 = gmx_simd4_calc_rsq_r(dx_S3, dy_S3, dz_S3);
2372 wco_S0 = gmx_simd4_cmplt_r(rsq_S0, rc2_S);
2373 wco_S1 = gmx_simd4_cmplt_r(rsq_S1, rc2_S);
2374 wco_S2 = gmx_simd4_cmplt_r(rsq_S2, rc2_S);
2375 wco_S3 = gmx_simd4_cmplt_r(rsq_S3, rc2_S);
2377 wco_any_S01 = gmx_simd4_or_b(wco_S0, wco_S1);
2378 wco_any_S23 = gmx_simd4_or_b(wco_S2, wco_S3);
2379 wco_any_S = gmx_simd4_or_b(wco_any_S01, wco_any_S23);
2381 if (gmx_simd4_anytrue_b(wco_any_S))
2395 /* Returns the j sub-cell for index cj_ind */
2396 static int nbl_cj(const nbnxn_pairlist_t *nbl, int cj_ind)
2398 return nbl->cj4[cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG].cj[cj_ind & (NBNXN_GPU_JGROUP_SIZE - 1)];
2401 /* Returns the i-interaction mask of the j sub-cell for index cj_ind */
2402 static unsigned int nbl_imask0(const nbnxn_pairlist_t *nbl, int cj_ind)
2404 return nbl->cj4[cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG].imei[0].imask;
2407 /* Ensures there is enough space for extra extra exclusion masks */
2408 static void check_excl_space(nbnxn_pairlist_t *nbl, int extra)
2410 if (nbl->nexcl+extra > nbl->excl_nalloc)
2412 nbl->excl_nalloc = over_alloc_small(nbl->nexcl+extra);
2413 nbnxn_realloc_void((void **)&nbl->excl,
2414 nbl->nexcl*sizeof(*nbl->excl),
2415 nbl->excl_nalloc*sizeof(*nbl->excl),
2416 nbl->alloc, nbl->free);
2420 /* Ensures there is enough space for ncell extra j-cells in the list */
2421 static void check_subcell_list_space_simple(nbnxn_pairlist_t *nbl,
2426 cj_max = nbl->ncj + ncell;
2428 if (cj_max > nbl->cj_nalloc)
2430 nbl->cj_nalloc = over_alloc_small(cj_max);
2431 nbnxn_realloc_void((void **)&nbl->cj,
2432 nbl->ncj*sizeof(*nbl->cj),
2433 nbl->cj_nalloc*sizeof(*nbl->cj),
2434 nbl->alloc, nbl->free);
2438 /* Ensures there is enough space for ncell extra j-subcells in the list */
2439 static void check_subcell_list_space_supersub(nbnxn_pairlist_t *nbl,
2442 int ncj4_max, j4, j, w, t;
2445 #define WARP_SIZE 32
2447 /* We can have maximally nsupercell*GPU_NSUBCELL sj lists */
2448 /* We can store 4 j-subcell - i-supercell pairs in one struct.
2449 * since we round down, we need one extra entry.
2451 ncj4_max = ((nbl->work->cj_ind + nsupercell*GPU_NSUBCELL + NBNXN_GPU_JGROUP_SIZE - 1) >> NBNXN_GPU_JGROUP_SIZE_2LOG);
2453 if (ncj4_max > nbl->cj4_nalloc)
2455 nbl->cj4_nalloc = over_alloc_small(ncj4_max);
2456 nbnxn_realloc_void((void **)&nbl->cj4,
2457 nbl->work->cj4_init*sizeof(*nbl->cj4),
2458 nbl->cj4_nalloc*sizeof(*nbl->cj4),
2459 nbl->alloc, nbl->free);
2462 if (ncj4_max > nbl->work->cj4_init)
2464 for (j4 = nbl->work->cj4_init; j4 < ncj4_max; j4++)
2466 /* No i-subcells and no excl's in the list initially */
2467 for (w = 0; w < NWARP; w++)
2469 nbl->cj4[j4].imei[w].imask = 0U;
2470 nbl->cj4[j4].imei[w].excl_ind = 0;
2474 nbl->work->cj4_init = ncj4_max;
2478 /* Set all excl masks for one GPU warp no exclusions */
2479 static void set_no_excls(nbnxn_excl_t *excl)
2483 for (t = 0; t < WARP_SIZE; t++)
2485 /* Turn all interaction bits on */
2486 excl->pair[t] = NBNXN_INTERACTION_MASK_ALL;
2490 /* Initializes a single nbnxn_pairlist_t data structure */
2491 static void nbnxn_init_pairlist(nbnxn_pairlist_t *nbl,
2493 nbnxn_alloc_t *alloc,
2498 nbl->alloc = nbnxn_alloc_aligned;
2506 nbl->free = nbnxn_free_aligned;
2513 nbl->bSimple = bSimple;
2524 /* We need one element extra in sj, so alloc initially with 1 */
2525 nbl->cj4_nalloc = 0;
2532 nbl->excl_nalloc = 0;
2534 check_excl_space(nbl, 1);
2536 set_no_excls(&nbl->excl[0]);
2542 snew_aligned(nbl->work->bb_ci, 1, NBNXN_SEARCH_BB_MEM_ALIGN);
2547 snew_aligned(nbl->work->pbb_ci, GPU_NSUBCELL/STRIDE_PBB*NNBSBB_XXXX, NBNXN_SEARCH_BB_MEM_ALIGN);
2549 snew_aligned(nbl->work->bb_ci, GPU_NSUBCELL, NBNXN_SEARCH_BB_MEM_ALIGN);
2552 snew_aligned(nbl->work->x_ci, NBNXN_NA_SC_MAX*DIM, NBNXN_SEARCH_BB_MEM_ALIGN);
2553 #ifdef GMX_NBNXN_SIMD
2554 snew_aligned(nbl->work->x_ci_simd_4xn, 1, NBNXN_MEM_ALIGN);
2555 snew_aligned(nbl->work->x_ci_simd_2xnn, 1, NBNXN_MEM_ALIGN);
2557 snew_aligned(nbl->work->d2, GPU_NSUBCELL, NBNXN_SEARCH_BB_MEM_ALIGN);
2559 nbl->work->sort = NULL;
2560 nbl->work->sort_nalloc = 0;
2561 nbl->work->sci_sort = NULL;
2562 nbl->work->sci_sort_nalloc = 0;
2565 void nbnxn_init_pairlist_set(nbnxn_pairlist_set_t *nbl_list,
2566 gmx_bool bSimple, gmx_bool bCombined,
2567 nbnxn_alloc_t *alloc,
2572 nbl_list->bSimple = bSimple;
2573 nbl_list->bCombined = bCombined;
2575 nbl_list->nnbl = gmx_omp_nthreads_get(emntNonbonded);
2577 if (!nbl_list->bCombined &&
2578 nbl_list->nnbl > NBNXN_BUFFERFLAG_MAX_THREADS)
2580 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.",
2581 nbl_list->nnbl, NBNXN_BUFFERFLAG_MAX_THREADS, NBNXN_BUFFERFLAG_MAX_THREADS);
2584 snew(nbl_list->nbl, nbl_list->nnbl);
2585 snew(nbl_list->nbl_fep, nbl_list->nnbl);
2586 /* Execute in order to avoid memory interleaving between threads */
2587 #pragma omp parallel for num_threads(nbl_list->nnbl) schedule(static)
2588 for (i = 0; i < nbl_list->nnbl; i++)
2590 /* Allocate the nblist data structure locally on each thread
2591 * to optimize memory access for NUMA architectures.
2593 snew(nbl_list->nbl[i], 1);
2595 /* Only list 0 is used on the GPU, use normal allocation for i>0 */
2598 nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, alloc, free);
2602 nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, NULL, NULL);
2605 snew(nbl_list->nbl_fep[i], 1);
2606 nbnxn_init_pairlist_fep(nbl_list->nbl_fep[i]);
2610 /* Print statistics of a pair list, used for debug output */
2611 static void print_nblist_statistics_simple(FILE *fp, const nbnxn_pairlist_t *nbl,
2612 const nbnxn_search_t nbs, real rl)
2614 const nbnxn_grid_t *grid;
2619 /* This code only produces correct statistics with domain decomposition */
2620 grid = &nbs->grid[0];
2622 fprintf(fp, "nbl nci %d ncj %d\n",
2623 nbl->nci, nbl->ncj);
2624 fprintf(fp, "nbl na_sc %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
2625 nbl->na_sc, rl, nbl->ncj, nbl->ncj/(double)grid->nc,
2626 nbl->ncj/(double)grid->nc*grid->na_sc,
2627 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])));
2629 fprintf(fp, "nbl average j cell list length %.1f\n",
2630 0.25*nbl->ncj/(double)nbl->nci);
2632 for (s = 0; s < SHIFTS; s++)
2637 for (i = 0; i < nbl->nci; i++)
2639 cs[nbl->ci[i].shift & NBNXN_CI_SHIFT] +=
2640 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start;
2642 j = nbl->ci[i].cj_ind_start;
2643 while (j < nbl->ci[i].cj_ind_end &&
2644 nbl->cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
2650 fprintf(fp, "nbl cell pairs, total: %d excl: %d %.1f%%\n",
2651 nbl->ncj, npexcl, 100*npexcl/(double)nbl->ncj);
2652 for (s = 0; s < SHIFTS; s++)
2656 fprintf(fp, "nbl shift %2d ncj %3d\n", s, cs[s]);
2661 /* Print statistics of a pair lists, used for debug output */
2662 static void print_nblist_statistics_supersub(FILE *fp, const nbnxn_pairlist_t *nbl,
2663 const nbnxn_search_t nbs, real rl)
2665 const nbnxn_grid_t *grid;
2666 int i, j4, j, si, b;
2667 int c[GPU_NSUBCELL+1];
2669 /* This code only produces correct statistics with domain decomposition */
2670 grid = &nbs->grid[0];
2672 fprintf(fp, "nbl nsci %d ncj4 %d nsi %d excl4 %d\n",
2673 nbl->nsci, nbl->ncj4, nbl->nci_tot, nbl->nexcl);
2674 fprintf(fp, "nbl na_c %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
2675 nbl->na_ci, rl, nbl->nci_tot, nbl->nci_tot/(double)grid->nsubc_tot,
2676 nbl->nci_tot/(double)grid->nsubc_tot*grid->na_c,
2677 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])));
2679 fprintf(fp, "nbl average j super cell list length %.1f\n",
2680 0.25*nbl->ncj4/(double)nbl->nsci);
2681 fprintf(fp, "nbl average i sub cell list length %.1f\n",
2682 nbl->nci_tot/((double)nbl->ncj4));
2684 for (si = 0; si <= GPU_NSUBCELL; si++)
2688 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)))
2706 for (b = 0; b <= GPU_NSUBCELL; b++)
2708 fprintf(fp, "nbl j-list #i-subcell %d %7d %4.1f\n",
2709 b, c[b], 100.0*c[b]/(double)(nbl->ncj4*NBNXN_GPU_JGROUP_SIZE));
2713 /* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp */
2714 static void low_get_nbl_exclusions(nbnxn_pairlist_t *nbl, int cj4,
2715 int warp, nbnxn_excl_t **excl)
2717 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
2719 /* No exclusions set, make a new list entry */
2720 nbl->cj4[cj4].imei[warp].excl_ind = nbl->nexcl;
2722 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
2723 set_no_excls(*excl);
2727 /* We already have some exclusions, new ones can be added to the list */
2728 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
2732 /* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp,
2733 * generates a new element and allocates extra memory, if necessary.
2735 static void get_nbl_exclusions_1(nbnxn_pairlist_t *nbl, int cj4,
2736 int warp, nbnxn_excl_t **excl)
2738 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
2740 /* We need to make a new list entry, check if we have space */
2741 check_excl_space(nbl, 1);
2743 low_get_nbl_exclusions(nbl, cj4, warp, excl);
2746 /* Returns pointers to the exclusion mask for cj4-unit cj4 for both warps,
2747 * generates a new element and allocates extra memory, if necessary.
2749 static void get_nbl_exclusions_2(nbnxn_pairlist_t *nbl, int cj4,
2750 nbnxn_excl_t **excl_w0,
2751 nbnxn_excl_t **excl_w1)
2753 /* Check for space we might need */
2754 check_excl_space(nbl, 2);
2756 low_get_nbl_exclusions(nbl, cj4, 0, excl_w0);
2757 low_get_nbl_exclusions(nbl, cj4, 1, excl_w1);
2760 /* Sets the self exclusions i=j and pair exclusions i>j */
2761 static void set_self_and_newton_excls_supersub(nbnxn_pairlist_t *nbl,
2762 int cj4_ind, int sj_offset,
2765 nbnxn_excl_t *excl[2];
2768 /* Here we only set the set self and double pair exclusions */
2770 get_nbl_exclusions_2(nbl, cj4_ind, &excl[0], &excl[1]);
2772 /* Only minor < major bits set */
2773 for (ej = 0; ej < nbl->na_ci; ej++)
2776 for (ei = ej; ei < nbl->na_ci; ei++)
2778 excl[w]->pair[(ej & (NBNXN_GPU_JGROUP_SIZE-1))*nbl->na_ci + ei] &=
2779 ~(1U << (sj_offset*GPU_NSUBCELL + si));
2784 /* Returns a diagonal or off-diagonal interaction mask for plain C lists */
2785 static unsigned int get_imask(gmx_bool rdiag, int ci, int cj)
2787 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
2790 /* Returns a diagonal or off-diagonal interaction mask for cj-size=2 */
2791 static unsigned int get_imask_simd_j2(gmx_bool rdiag, int ci, int cj)
2793 return (rdiag && ci*2 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_0 :
2794 (rdiag && ci*2+1 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_1 :
2795 NBNXN_INTERACTION_MASK_ALL));
2798 /* Returns a diagonal or off-diagonal interaction mask for cj-size=4 */
2799 static unsigned int get_imask_simd_j4(gmx_bool rdiag, int ci, int cj)
2801 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
2804 /* Returns a diagonal or off-diagonal interaction mask for cj-size=8 */
2805 static unsigned int get_imask_simd_j8(gmx_bool rdiag, int ci, int cj)
2807 return (rdiag && ci == cj*2 ? NBNXN_INTERACTION_MASK_DIAG_J8_0 :
2808 (rdiag && ci == cj*2+1 ? NBNXN_INTERACTION_MASK_DIAG_J8_1 :
2809 NBNXN_INTERACTION_MASK_ALL));
2812 #ifdef GMX_NBNXN_SIMD
2813 #if GMX_SIMD_REAL_WIDTH == 2
2814 #define get_imask_simd_4xn get_imask_simd_j2
2816 #if GMX_SIMD_REAL_WIDTH == 4
2817 #define get_imask_simd_4xn get_imask_simd_j4
2819 #if GMX_SIMD_REAL_WIDTH == 8
2820 #define get_imask_simd_4xn get_imask_simd_j8
2821 #define get_imask_simd_2xnn get_imask_simd_j4
2823 #if GMX_SIMD_REAL_WIDTH == 16
2824 #define get_imask_simd_2xnn get_imask_simd_j8
2828 /* Plain C code for making a pair list of cell ci vs cell cjf-cjl.
2829 * Checks bounding box distances and possibly atom pair distances.
2831 static void make_cluster_list_simple(const nbnxn_grid_t *gridj,
2832 nbnxn_pairlist_t *nbl,
2833 int ci, int cjf, int cjl,
2834 gmx_bool remove_sub_diag,
2836 real rl2, float rbb2,
2839 const nbnxn_list_work_t *work;
2841 const nbnxn_bb_t *bb_ci;
2846 int cjf_gl, cjl_gl, cj;
2850 bb_ci = nbl->work->bb_ci;
2851 x_ci = nbl->work->x_ci;
2854 while (!InRange && cjf <= cjl)
2856 d2 = subc_bb_dist2(0, bb_ci, cjf, gridj->bb);
2859 /* Check if the distance is within the distance where
2860 * we use only the bounding box distance rbb,
2861 * or within the cut-off and there is at least one atom pair
2862 * within the cut-off.
2872 cjf_gl = gridj->cell0 + cjf;
2873 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
2875 for (j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE; j++)
2877 InRange = InRange ||
2878 (sqr(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
2879 sqr(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
2880 sqr(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rl2);
2883 *ndistc += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
2896 while (!InRange && cjl > cjf)
2898 d2 = subc_bb_dist2(0, bb_ci, cjl, gridj->bb);
2901 /* Check if the distance is within the distance where
2902 * we use only the bounding box distance rbb,
2903 * or within the cut-off and there is at least one atom pair
2904 * within the cut-off.
2914 cjl_gl = gridj->cell0 + cjl;
2915 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
2917 for (j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE; j++)
2919 InRange = InRange ||
2920 (sqr(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
2921 sqr(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
2922 sqr(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rl2);
2925 *ndistc += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
2935 for (cj = cjf; cj <= cjl; cj++)
2937 /* Store cj and the interaction mask */
2938 nbl->cj[nbl->ncj].cj = gridj->cell0 + cj;
2939 nbl->cj[nbl->ncj].excl = get_imask(remove_sub_diag, ci, cj);
2942 /* Increase the closing index in i super-cell list */
2943 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
2947 #ifdef GMX_NBNXN_SIMD_4XN
2948 #include "gromacs/mdlib/nbnxn_search_simd_4xn.h"
2950 #ifdef GMX_NBNXN_SIMD_2XNN
2951 #include "gromacs/mdlib/nbnxn_search_simd_2xnn.h"
2954 /* Plain C or SIMD4 code for making a pair list of super-cell sci vs scj.
2955 * Checks bounding box distances and possibly atom pair distances.
2957 static void make_cluster_list_supersub(const nbnxn_grid_t *gridi,
2958 const nbnxn_grid_t *gridj,
2959 nbnxn_pairlist_t *nbl,
2961 gmx_bool sci_equals_scj,
2962 int stride, const real *x,
2963 real rl2, float rbb2,
2968 int cjo, ci1, ci, cj, cj_gl;
2969 int cj4_ind, cj_offset;
2973 const float *pbb_ci;
2975 const nbnxn_bb_t *bb_ci;
2980 #define PRUNE_LIST_CPU_ONE
2981 #ifdef PRUNE_LIST_CPU_ONE
2985 d2l = nbl->work->d2;
2988 pbb_ci = nbl->work->pbb_ci;
2990 bb_ci = nbl->work->bb_ci;
2992 x_ci = nbl->work->x_ci;
2996 for (cjo = 0; cjo < gridj->nsubc[scj]; cjo++)
2998 cj4_ind = (nbl->work->cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG);
2999 cj_offset = nbl->work->cj_ind - cj4_ind*NBNXN_GPU_JGROUP_SIZE;
3000 cj4 = &nbl->cj4[cj4_ind];
3002 cj = scj*GPU_NSUBCELL + cjo;
3004 cj_gl = gridj->cell0*GPU_NSUBCELL + cj;
3006 /* Initialize this j-subcell i-subcell list */
3007 cj4->cj[cj_offset] = cj_gl;
3016 ci1 = gridi->nsubc[sci];
3020 /* Determine all ci1 bb distances in one call with SIMD4 */
3021 subc_bb_dist2_simd4_xxxx(gridj->pbb+(cj>>STRIDE_PBB_2LOG)*NNBSBB_XXXX+(cj & (STRIDE_PBB-1)),
3027 /* We use a fixed upper-bound instead of ci1 to help optimization */
3028 for (ci = 0; ci < GPU_NSUBCELL; ci++)
3035 #ifndef NBNXN_BBXXXX
3036 /* Determine the bb distance between ci and cj */
3037 d2l[ci] = subc_bb_dist2(ci, bb_ci, cj, gridj->bb);
3042 #ifdef PRUNE_LIST_CPU_ALL
3043 /* Check if the distance is within the distance where
3044 * we use only the bounding box distance rbb,
3045 * or within the cut-off and there is at least one atom pair
3046 * within the cut-off. This check is very costly.
3048 *ndistc += na_c*na_c;
3051 #ifdef NBNXN_PBB_SIMD4
3056 (na_c, ci, x_ci, cj_gl, stride, x, rl2)))
3058 /* Check if the distance between the two bounding boxes
3059 * in within the pair-list cut-off.
3064 /* Flag this i-subcell to be taken into account */
3065 imask |= (1U << (cj_offset*GPU_NSUBCELL+ci));
3067 #ifdef PRUNE_LIST_CPU_ONE
3075 #ifdef PRUNE_LIST_CPU_ONE
3076 /* If we only found 1 pair, check if any atoms are actually
3077 * within the cut-off, so we could get rid of it.
3079 if (npair == 1 && d2l[ci_last] >= rbb2)
3081 /* Avoid using function pointers here, as it's slower */
3083 #ifdef NBNXN_PBB_SIMD4
3084 !subc_in_range_simd4
3088 (na_c, ci_last, x_ci, cj_gl, stride, x, rl2))
3090 imask &= ~(1U << (cj_offset*GPU_NSUBCELL+ci_last));
3098 /* We have a useful sj entry, close it now */
3100 /* Set the exclucions for the ci== sj entry.
3101 * Here we don't bother to check if this entry is actually flagged,
3102 * as it will nearly always be in the list.
3106 set_self_and_newton_excls_supersub(nbl, cj4_ind, cj_offset, cjo);
3109 /* Copy the cluster interaction mask to the list */
3110 for (w = 0; w < NWARP; w++)
3112 cj4->imei[w].imask |= imask;
3115 nbl->work->cj_ind++;
3117 /* Keep the count */
3118 nbl->nci_tot += npair;
3120 /* Increase the closing index in i super-cell list */
3121 nbl->sci[nbl->nsci].cj4_ind_end =
3122 ((nbl->work->cj_ind+NBNXN_GPU_JGROUP_SIZE-1) >> NBNXN_GPU_JGROUP_SIZE_2LOG);
3127 /* Set all atom-pair exclusions from the topology stored in excl
3128 * as masks in the pair-list for simple list i-entry nbl_ci
3130 static void set_ci_top_excls(const nbnxn_search_t nbs,
3131 nbnxn_pairlist_t *nbl,
3132 gmx_bool diagRemoved,
3135 const nbnxn_ci_t *nbl_ci,
3136 const t_blocka *excl)
3140 int cj_ind_first, cj_ind_last;
3141 int cj_first, cj_last;
3143 int i, ai, aj, si, eind, ge, se;
3144 int found, cj_ind_0, cj_ind_1, cj_ind_m;
3148 nbnxn_excl_t *nbl_excl;
3149 int inner_i, inner_e;
3153 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
3161 cj_ind_first = nbl_ci->cj_ind_start;
3162 cj_ind_last = nbl->ncj - 1;
3164 cj_first = nbl->cj[cj_ind_first].cj;
3165 cj_last = nbl->cj[cj_ind_last].cj;
3167 /* Determine how many contiguous j-cells we have starting
3168 * from the first i-cell. This number can be used to directly
3169 * calculate j-cell indices for excluded atoms.
3172 if (na_ci_2log == na_cj_2log)
3174 while (cj_ind_first + ndirect <= cj_ind_last &&
3175 nbl->cj[cj_ind_first+ndirect].cj == ci + ndirect)
3180 #ifdef NBNXN_SEARCH_BB_SIMD4
3183 while (cj_ind_first + ndirect <= cj_ind_last &&
3184 nbl->cj[cj_ind_first+ndirect].cj == ci_to_cj(na_cj_2log, ci) + ndirect)
3191 /* Loop over the atoms in the i super-cell */
3192 for (i = 0; i < nbl->na_sc; i++)
3194 ai = nbs->a[ci*nbl->na_sc+i];
3197 si = (i>>na_ci_2log);
3199 /* Loop over the topology-based exclusions for this i-atom */
3200 for (eind = excl->index[ai]; eind < excl->index[ai+1]; eind++)
3206 /* The self exclusion are already set, save some time */
3212 /* Without shifts we only calculate interactions j>i
3213 * for one-way pair-lists.
3215 if (diagRemoved && ge <= ci*nbl->na_sc + i)
3220 se = (ge >> na_cj_2log);
3222 /* Could the cluster se be in our list? */
3223 if (se >= cj_first && se <= cj_last)
3225 if (se < cj_first + ndirect)
3227 /* We can calculate cj_ind directly from se */
3228 found = cj_ind_first + se - cj_first;
3232 /* Search for se using bisection */
3234 cj_ind_0 = cj_ind_first + ndirect;
3235 cj_ind_1 = cj_ind_last + 1;
3236 while (found == -1 && cj_ind_0 < cj_ind_1)
3238 cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
3240 cj_m = nbl->cj[cj_ind_m].cj;
3248 cj_ind_1 = cj_ind_m;
3252 cj_ind_0 = cj_ind_m + 1;
3259 inner_i = i - (si << na_ci_2log);
3260 inner_e = ge - (se << na_cj_2log);
3262 nbl->cj[found].excl &= ~(1U<<((inner_i<<na_cj_2log) + inner_e));
3270 /* Add a new i-entry to the FEP list and copy the i-properties */
3271 static gmx_inline void fep_list_new_nri_copy(t_nblist *nlist)
3273 /* Add a new i-entry */
3276 assert(nlist->nri < nlist->maxnri);
3278 /* Duplicate the last i-entry, except for jindex, which continues */
3279 nlist->iinr[nlist->nri] = nlist->iinr[nlist->nri-1];
3280 nlist->shift[nlist->nri] = nlist->shift[nlist->nri-1];
3281 nlist->gid[nlist->nri] = nlist->gid[nlist->nri-1];
3282 nlist->jindex[nlist->nri] = nlist->nrj;
3285 /* For load balancing of the free-energy lists over threads, we set
3286 * the maximum nrj size of an i-entry to 40. This leads to good
3287 * load balancing in the worst case scenario of a single perturbed
3288 * particle on 16 threads, while not introducing significant overhead.
3289 * Note that half of the perturbed pairs will anyhow end up in very small lists,
3290 * since non perturbed i-particles will see few perturbed j-particles).
3292 const int max_nrj_fep = 40;
3294 /* Exclude the perturbed pairs from the Verlet list. This is only done to avoid
3295 * singularities for overlapping particles (0/0), since the charges and
3296 * LJ parameters have been zeroed in the nbnxn data structure.
3297 * Simultaneously make a group pair list for the perturbed pairs.
3299 static void make_fep_list(const nbnxn_search_t nbs,
3300 const nbnxn_atomdata_t *nbat,
3301 nbnxn_pairlist_t *nbl,
3302 gmx_bool bDiagRemoved,
3304 const nbnxn_grid_t *gridi,
3305 const nbnxn_grid_t *gridj,
3308 int ci, cj_ind_start, cj_ind_end, cj_ind, cja, cjr;
3310 int ngid, gid_i = 0, gid_j, gid;
3311 int egp_shift, egp_mask;
3313 int i, j, ind_i, ind_j, ai, aj;
3315 gmx_bool bFEP_i, bFEP_i_all;
3317 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
3325 cj_ind_start = nbl_ci->cj_ind_start;
3326 cj_ind_end = nbl_ci->cj_ind_end;
3328 /* In worst case we have alternating energy groups
3329 * and create #atom-pair lists, which means we need the size
3330 * of a cluster pair (na_ci*na_cj) times the number of cj's.
3332 nri_max = nbl->na_ci*nbl->na_cj*(cj_ind_end - cj_ind_start);
3333 if (nlist->nri + nri_max > nlist->maxnri)
3335 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
3336 reallocate_nblist(nlist);
3339 ngid = nbat->nenergrp;
3341 if (ngid*gridj->na_cj > sizeof(gid_cj)*8)
3343 gmx_fatal(FARGS, "The Verlet scheme with %dx%d kernels and free-energy only supports up to %d energy groups",
3344 gridi->na_c, gridj->na_cj, (sizeof(gid_cj)*8)/gridj->na_cj);
3347 egp_shift = nbat->neg_2log;
3348 egp_mask = (1<<nbat->neg_2log) - 1;
3350 /* Loop over the atoms in the i sub-cell */
3352 for (i = 0; i < nbl->na_ci; i++)
3354 ind_i = ci*nbl->na_ci + i;
3359 nlist->jindex[nri+1] = nlist->jindex[nri];
3360 nlist->iinr[nri] = ai;
3361 /* The actual energy group pair index is set later */
3362 nlist->gid[nri] = 0;
3363 nlist->shift[nri] = nbl_ci->shift & NBNXN_CI_SHIFT;
3365 bFEP_i = gridi->fep[ci - gridi->cell0] & (1 << i);
3367 bFEP_i_all = bFEP_i_all && bFEP_i;
3369 if ((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj > nlist->maxnrj)
3371 nlist->maxnrj = over_alloc_small((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj);
3372 srenew(nlist->jjnr, nlist->maxnrj);
3373 srenew(nlist->excl_fep, nlist->maxnrj);
3378 gid_i = (nbat->energrp[ci] >> (egp_shift*i)) & egp_mask;
3381 for (cj_ind = cj_ind_start; cj_ind < cj_ind_end; cj_ind++)
3383 unsigned int fep_cj;
3385 cja = nbl->cj[cj_ind].cj;
3387 if (gridj->na_cj == gridj->na_c)
3389 cjr = cja - gridj->cell0;
3390 fep_cj = gridj->fep[cjr];
3393 gid_cj = nbat->energrp[cja];
3396 else if (2*gridj->na_cj == gridj->na_c)
3398 cjr = cja - gridj->cell0*2;
3399 /* Extract half of the ci fep/energrp mask */
3400 fep_cj = (gridj->fep[cjr>>1] >> ((cjr&1)*gridj->na_cj)) & ((1<<gridj->na_cj) - 1);
3403 gid_cj = nbat->energrp[cja>>1] >> ((cja&1)*gridj->na_cj*egp_shift) & ((1<<(gridj->na_cj*egp_shift)) - 1);
3408 cjr = cja - (gridj->cell0>>1);
3409 /* Combine two ci fep masks/energrp */
3410 fep_cj = gridj->fep[cjr*2] + (gridj->fep[cjr*2+1] << gridj->na_c);
3413 gid_cj = nbat->energrp[cja*2] + (nbat->energrp[cja*2+1] << (gridj->na_c*egp_shift));
3417 if (bFEP_i || fep_cj != 0)
3419 for (j = 0; j < nbl->na_cj; j++)
3421 /* Is this interaction perturbed and not excluded? */
3422 ind_j = cja*nbl->na_cj + j;
3425 (bFEP_i || (fep_cj & (1 << j))) &&
3426 (!bDiagRemoved || ind_j >= ind_i))
3430 gid_j = (gid_cj >> (j*egp_shift)) & egp_mask;
3431 gid = GID(gid_i, gid_j, ngid);
3433 if (nlist->nrj > nlist->jindex[nri] &&
3434 nlist->gid[nri] != gid)
3436 /* Energy group pair changed: new list */
3437 fep_list_new_nri_copy(nlist);
3440 nlist->gid[nri] = gid;
3443 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
3445 fep_list_new_nri_copy(nlist);
3449 /* Add it to the FEP list */
3450 nlist->jjnr[nlist->nrj] = aj;
3451 nlist->excl_fep[nlist->nrj] = (nbl->cj[cj_ind].excl >> (i*nbl->na_cj + j)) & 1;
3454 /* Exclude it from the normal list.
3455 * Note that the charge has been set to zero,
3456 * but we need to avoid 0/0, as perturbed atoms
3457 * can be on top of each other.
3459 nbl->cj[cj_ind].excl &= ~(1U << (i*nbl->na_cj + j));
3465 if (nlist->nrj > nlist->jindex[nri])
3467 /* Actually add this new, non-empty, list */
3469 nlist->jindex[nlist->nri] = nlist->nrj;
3476 /* All interactions are perturbed, we can skip this entry */
3477 nbl_ci->cj_ind_end = cj_ind_start;
3481 /* Return the index of atom a within a cluster */
3482 static gmx_inline int cj_mod_cj4(int cj)
3484 return cj & (NBNXN_GPU_JGROUP_SIZE - 1);
3487 /* Convert a j-cluster to a cj4 group */
3488 static gmx_inline int cj_to_cj4(int cj)
3490 return cj >> NBNXN_GPU_JGROUP_SIZE_2LOG;
3493 /* Return the index of an j-atom within a warp */
3494 static gmx_inline int a_mod_wj(int a)
3496 return a & (NBNXN_GPU_CLUSTER_SIZE/2 - 1);
3499 /* As make_fep_list above, but for super/sub lists. */
3500 static void make_fep_list_supersub(const nbnxn_search_t nbs,
3501 const nbnxn_atomdata_t *nbat,
3502 nbnxn_pairlist_t *nbl,
3503 gmx_bool bDiagRemoved,
3504 const nbnxn_sci_t *nbl_sci,
3509 const nbnxn_grid_t *gridi,
3510 const nbnxn_grid_t *gridj,
3513 int sci, cj4_ind_start, cj4_ind_end, cj4_ind, gcj, cjr;
3516 int i, j, ind_i, ind_j, ai, aj;
3520 const nbnxn_cj4_t *cj4;
3522 if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
3530 cj4_ind_start = nbl_sci->cj4_ind_start;
3531 cj4_ind_end = nbl_sci->cj4_ind_end;
3533 /* Here we process one super-cell, max #atoms na_sc, versus a list
3534 * cj4 entries, each with max NBNXN_GPU_JGROUP_SIZE cj's, each
3535 * of size na_cj atoms.
3536 * On the GPU we don't support energy groups (yet).
3537 * So for each of the na_sc i-atoms, we need max one FEP list
3538 * for each max_nrj_fep j-atoms.
3540 nri_max = nbl->na_sc*nbl->na_cj*(1 + ((cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE)/max_nrj_fep);
3541 if (nlist->nri + nri_max > nlist->maxnri)
3543 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
3544 reallocate_nblist(nlist);
3547 /* Loop over the atoms in the i super-cluster */
3548 for (c = 0; c < GPU_NSUBCELL; c++)
3550 c_abs = sci*GPU_NSUBCELL + c;
3552 for (i = 0; i < nbl->na_ci; i++)
3554 ind_i = c_abs*nbl->na_ci + i;
3559 nlist->jindex[nri+1] = nlist->jindex[nri];
3560 nlist->iinr[nri] = ai;
3561 /* With GPUs, energy groups are not supported */
3562 nlist->gid[nri] = 0;
3563 nlist->shift[nri] = nbl_sci->shift & NBNXN_CI_SHIFT;
3565 bFEP_i = (gridi->fep[c_abs - gridi->cell0*GPU_NSUBCELL] & (1 << i));
3567 xi = nbat->x[ind_i*nbat->xstride+XX] + shx;
3568 yi = nbat->x[ind_i*nbat->xstride+YY] + shy;
3569 zi = nbat->x[ind_i*nbat->xstride+ZZ] + shz;
3571 if ((nlist->nrj + cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE*nbl->na_cj > nlist->maxnrj)
3573 nlist->maxnrj = over_alloc_small((nlist->nrj + cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE*nbl->na_cj);
3574 srenew(nlist->jjnr, nlist->maxnrj);
3575 srenew(nlist->excl_fep, nlist->maxnrj);
3578 for (cj4_ind = cj4_ind_start; cj4_ind < cj4_ind_end; cj4_ind++)
3580 cj4 = &nbl->cj4[cj4_ind];
3582 for (gcj = 0; gcj < NBNXN_GPU_JGROUP_SIZE; gcj++)
3584 unsigned int fep_cj;
3586 if ((cj4->imei[0].imask & (1U << (gcj*GPU_NSUBCELL + c))) == 0)
3588 /* Skip this ci for this cj */
3592 cjr = cj4->cj[gcj] - gridj->cell0*GPU_NSUBCELL;
3594 fep_cj = gridj->fep[cjr];
3596 if (bFEP_i || fep_cj != 0)
3598 for (j = 0; j < nbl->na_cj; j++)
3600 /* Is this interaction perturbed and not excluded? */
3601 ind_j = (gridj->cell0*GPU_NSUBCELL + cjr)*nbl->na_cj + j;
3604 (bFEP_i || (fep_cj & (1 << j))) &&
3605 (!bDiagRemoved || ind_j >= ind_i))
3609 unsigned int excl_bit;
3612 get_nbl_exclusions_1(nbl, cj4_ind, j>>2, &excl);
3614 excl_pair = a_mod_wj(j)*nbl->na_ci + i;
3615 excl_bit = (1U << (gcj*GPU_NSUBCELL + c));
3617 dx = nbat->x[ind_j*nbat->xstride+XX] - xi;
3618 dy = nbat->x[ind_j*nbat->xstride+YY] - yi;
3619 dz = nbat->x[ind_j*nbat->xstride+ZZ] - zi;
3621 /* The unpruned GPU list has more than 2/3
3622 * of the atom pairs beyond rlist. Using
3623 * this list will cause a lot of overhead
3624 * in the CPU FEP kernels, especially
3625 * relative to the fast GPU kernels.
3626 * So we prune the FEP list here.
3628 if (dx*dx + dy*dy + dz*dz < rlist_fep2)
3630 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
3632 fep_list_new_nri_copy(nlist);
3636 /* Add it to the FEP list */
3637 nlist->jjnr[nlist->nrj] = aj;
3638 nlist->excl_fep[nlist->nrj] = (excl->pair[excl_pair] & excl_bit) ? 1 : 0;
3642 /* Exclude it from the normal list.
3643 * Note that the charge and LJ parameters have
3644 * been set to zero, but we need to avoid 0/0,
3645 * as perturbed atoms can be on top of each other.
3647 excl->pair[excl_pair] &= ~excl_bit;
3651 /* Note that we could mask out this pair in imask
3652 * if all i- and/or all j-particles are perturbed.
3653 * But since the perturbed pairs on the CPU will
3654 * take an order of magnitude more time, the GPU
3655 * will finish before the CPU and there is no gain.
3661 if (nlist->nrj > nlist->jindex[nri])
3663 /* Actually add this new, non-empty, list */
3665 nlist->jindex[nlist->nri] = nlist->nrj;
3672 /* Set all atom-pair exclusions from the topology stored in excl
3673 * as masks in the pair-list for i-super-cell entry nbl_sci
3675 static void set_sci_top_excls(const nbnxn_search_t nbs,
3676 nbnxn_pairlist_t *nbl,
3677 gmx_bool diagRemoved,
3679 const nbnxn_sci_t *nbl_sci,
3680 const t_blocka *excl)
3685 int cj_ind_first, cj_ind_last;
3686 int cj_first, cj_last;
3688 int i, ai, aj, si, eind, ge, se;
3689 int found, cj_ind_0, cj_ind_1, cj_ind_m;
3693 nbnxn_excl_t *nbl_excl;
3694 int inner_i, inner_e, w;
3700 if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
3708 cj_ind_first = nbl_sci->cj4_ind_start*NBNXN_GPU_JGROUP_SIZE;
3709 cj_ind_last = nbl->work->cj_ind - 1;
3711 cj_first = nbl->cj4[nbl_sci->cj4_ind_start].cj[0];
3712 cj_last = nbl_cj(nbl, cj_ind_last);
3714 /* Determine how many contiguous j-clusters we have starting
3715 * from the first i-cluster. This number can be used to directly
3716 * calculate j-cluster indices for excluded atoms.
3719 while (cj_ind_first + ndirect <= cj_ind_last &&
3720 nbl_cj(nbl, cj_ind_first+ndirect) == sci*GPU_NSUBCELL + ndirect)
3725 /* Loop over the atoms in the i super-cell */
3726 for (i = 0; i < nbl->na_sc; i++)
3728 ai = nbs->a[sci*nbl->na_sc+i];
3731 si = (i>>na_c_2log);
3733 /* Loop over the topology-based exclusions for this i-atom */
3734 for (eind = excl->index[ai]; eind < excl->index[ai+1]; eind++)
3740 /* The self exclusion are already set, save some time */
3746 /* Without shifts we only calculate interactions j>i
3747 * for one-way pair-lists.
3749 if (diagRemoved && ge <= sci*nbl->na_sc + i)
3755 /* Could the cluster se be in our list? */
3756 if (se >= cj_first && se <= cj_last)
3758 if (se < cj_first + ndirect)
3760 /* We can calculate cj_ind directly from se */
3761 found = cj_ind_first + se - cj_first;
3765 /* Search for se using bisection */
3767 cj_ind_0 = cj_ind_first + ndirect;
3768 cj_ind_1 = cj_ind_last + 1;
3769 while (found == -1 && cj_ind_0 < cj_ind_1)
3771 cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
3773 cj_m = nbl_cj(nbl, cj_ind_m);
3781 cj_ind_1 = cj_ind_m;
3785 cj_ind_0 = cj_ind_m + 1;
3792 inner_i = i - si*na_c;
3793 inner_e = ge - se*na_c;
3795 if (nbl_imask0(nbl, found) & (1U << (cj_mod_cj4(found)*GPU_NSUBCELL + si)))
3799 get_nbl_exclusions_1(nbl, cj_to_cj4(found), w, &nbl_excl);
3801 nbl_excl->pair[a_mod_wj(inner_e)*nbl->na_ci+inner_i] &=
3802 ~(1U << (cj_mod_cj4(found)*GPU_NSUBCELL + si));
3811 /* Reallocate the simple ci list for at least n entries */
3812 static void nb_realloc_ci(nbnxn_pairlist_t *nbl, int n)
3814 nbl->ci_nalloc = over_alloc_small(n);
3815 nbnxn_realloc_void((void **)&nbl->ci,
3816 nbl->nci*sizeof(*nbl->ci),
3817 nbl->ci_nalloc*sizeof(*nbl->ci),
3818 nbl->alloc, nbl->free);
3821 /* Reallocate the super-cell sci list for at least n entries */
3822 static void nb_realloc_sci(nbnxn_pairlist_t *nbl, int n)
3824 nbl->sci_nalloc = over_alloc_small(n);
3825 nbnxn_realloc_void((void **)&nbl->sci,
3826 nbl->nsci*sizeof(*nbl->sci),
3827 nbl->sci_nalloc*sizeof(*nbl->sci),
3828 nbl->alloc, nbl->free);
3831 /* Make a new ci entry at index nbl->nci */
3832 static void new_ci_entry(nbnxn_pairlist_t *nbl, int ci, int shift, int flags)
3834 if (nbl->nci + 1 > nbl->ci_nalloc)
3836 nb_realloc_ci(nbl, nbl->nci+1);
3838 nbl->ci[nbl->nci].ci = ci;
3839 nbl->ci[nbl->nci].shift = shift;
3840 /* Store the interaction flags along with the shift */
3841 nbl->ci[nbl->nci].shift |= flags;
3842 nbl->ci[nbl->nci].cj_ind_start = nbl->ncj;
3843 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
3846 /* Make a new sci entry at index nbl->nsci */
3847 static void new_sci_entry(nbnxn_pairlist_t *nbl, int sci, int shift)
3849 if (nbl->nsci + 1 > nbl->sci_nalloc)
3851 nb_realloc_sci(nbl, nbl->nsci+1);
3853 nbl->sci[nbl->nsci].sci = sci;
3854 nbl->sci[nbl->nsci].shift = shift;
3855 nbl->sci[nbl->nsci].cj4_ind_start = nbl->ncj4;
3856 nbl->sci[nbl->nsci].cj4_ind_end = nbl->ncj4;
3859 /* Sort the simple j-list cj on exclusions.
3860 * Entries with exclusions will all be sorted to the beginning of the list.
3862 static void sort_cj_excl(nbnxn_cj_t *cj, int ncj,
3863 nbnxn_list_work_t *work)
3867 if (ncj > work->cj_nalloc)
3869 work->cj_nalloc = over_alloc_large(ncj);
3870 srenew(work->cj, work->cj_nalloc);
3873 /* Make a list of the j-cells involving exclusions */
3875 for (j = 0; j < ncj; j++)
3877 if (cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
3879 work->cj[jnew++] = cj[j];
3882 /* Check if there are exclusions at all or not just the first entry */
3883 if (!((jnew == 0) ||
3884 (jnew == 1 && cj[0].excl != NBNXN_INTERACTION_MASK_ALL)))
3886 for (j = 0; j < ncj; j++)
3888 if (cj[j].excl == NBNXN_INTERACTION_MASK_ALL)
3890 work->cj[jnew++] = cj[j];
3893 for (j = 0; j < ncj; j++)
3895 cj[j] = work->cj[j];
3900 /* Close this simple list i entry */
3901 static void close_ci_entry_simple(nbnxn_pairlist_t *nbl)
3905 /* All content of the new ci entry have already been filled correctly,
3906 * we only need to increase the count here (for non empty lists).
3908 jlen = nbl->ci[nbl->nci].cj_ind_end - nbl->ci[nbl->nci].cj_ind_start;
3911 sort_cj_excl(nbl->cj+nbl->ci[nbl->nci].cj_ind_start, jlen, nbl->work);
3913 /* The counts below are used for non-bonded pair/flop counts
3914 * and should therefore match the available kernel setups.
3916 if (!(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_COUL(0)))
3918 nbl->work->ncj_noq += jlen;
3920 else if ((nbl->ci[nbl->nci].shift & NBNXN_CI_HALF_LJ(0)) ||
3921 !(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_LJ(0)))
3923 nbl->work->ncj_hlj += jlen;
3930 /* Split sci entry for load balancing on the GPU.
3931 * Splitting ensures we have enough lists to fully utilize the whole GPU.
3932 * With progBal we generate progressively smaller lists, which improves
3933 * load balancing. As we only know the current count on our own thread,
3934 * we will need to estimate the current total amount of i-entries.
3935 * As the lists get concatenated later, this estimate depends
3936 * both on nthread and our own thread index.
3938 static void split_sci_entry(nbnxn_pairlist_t *nbl,
3939 int nsp_max_av, gmx_bool progBal, int nc_bal,
3940 int thread, int nthread)
3944 int cj4_start, cj4_end, j4len, cj4;
3946 int nsp, nsp_sci, nsp_cj4, nsp_cj4_e, nsp_cj4_p;
3951 /* Estimate the total numbers of ci's of the nblist combined
3952 * over all threads using the target number of ci's.
3954 nsci_est = nc_bal*thread/nthread + nbl->nsci;
3956 /* The first ci blocks should be larger, to avoid overhead.
3957 * The last ci blocks should be smaller, to improve load balancing.
3960 nsp_max_av*nc_bal*3/(2*(nsci_est - 1 + nc_bal)));
3964 nsp_max = nsp_max_av;
3967 cj4_start = nbl->sci[nbl->nsci-1].cj4_ind_start;
3968 cj4_end = nbl->sci[nbl->nsci-1].cj4_ind_end;
3969 j4len = cj4_end - cj4_start;
3971 if (j4len > 1 && j4len*GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE > nsp_max)
3973 /* Remove the last ci entry and process the cj4's again */
3981 for (cj4 = cj4_start; cj4 < cj4_end; cj4++)
3983 nsp_cj4_p = nsp_cj4;
3984 /* Count the number of cluster pairs in this cj4 group */
3986 for (p = 0; p < GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE; p++)
3988 nsp_cj4 += (nbl->cj4[cj4].imei[0].imask >> p) & 1;
3991 if (nsp_cj4 > 0 && nsp + nsp_cj4 > nsp_max)
3993 /* Split the list at cj4 */
3994 nbl->sci[sci].cj4_ind_end = cj4;
3995 /* Create a new sci entry */
3998 if (nbl->nsci+1 > nbl->sci_nalloc)
4000 nb_realloc_sci(nbl, nbl->nsci+1);
4002 nbl->sci[sci].sci = nbl->sci[nbl->nsci-1].sci;
4003 nbl->sci[sci].shift = nbl->sci[nbl->nsci-1].shift;
4004 nbl->sci[sci].cj4_ind_start = cj4;
4006 nsp_cj4_e = nsp_cj4_p;
4012 /* Put the remaining cj4's in the last sci entry */
4013 nbl->sci[sci].cj4_ind_end = cj4_end;
4015 /* Possibly balance out the last two sci's
4016 * by moving the last cj4 of the second last sci.
4018 if (nsp_sci - nsp_cj4_e >= nsp + nsp_cj4_e)
4020 nbl->sci[sci-1].cj4_ind_end--;
4021 nbl->sci[sci].cj4_ind_start--;
4028 /* Clost this super/sub list i entry */
4029 static void close_ci_entry_supersub(nbnxn_pairlist_t *nbl,
4031 gmx_bool progBal, int nc_bal,
4032 int thread, int nthread)
4037 /* All content of the new ci entry have already been filled correctly,
4038 * we only need to increase the count here (for non empty lists).
4040 j4len = nbl->sci[nbl->nsci].cj4_ind_end - nbl->sci[nbl->nsci].cj4_ind_start;
4043 /* We can only have complete blocks of 4 j-entries in a list,
4044 * so round the count up before closing.
4046 nbl->ncj4 = ((nbl->work->cj_ind + NBNXN_GPU_JGROUP_SIZE - 1) >> NBNXN_GPU_JGROUP_SIZE_2LOG);
4047 nbl->work->cj_ind = nbl->ncj4*NBNXN_GPU_JGROUP_SIZE;
4053 /* Measure the size of the new entry and potentially split it */
4054 split_sci_entry(nbl, nsp_max_av, progBal, nc_bal, thread, nthread);
4059 /* Syncs the working array before adding another grid pair to the list */
4060 static void sync_work(nbnxn_pairlist_t *nbl)
4064 nbl->work->cj_ind = nbl->ncj4*NBNXN_GPU_JGROUP_SIZE;
4065 nbl->work->cj4_init = nbl->ncj4;
4069 /* Clears an nbnxn_pairlist_t data structure */
4070 static void clear_pairlist(nbnxn_pairlist_t *nbl)
4079 nbl->work->ncj_noq = 0;
4080 nbl->work->ncj_hlj = 0;
4083 /* Clears a group scheme pair list */
4084 static void clear_pairlist_fep(t_nblist *nl)
4088 if (nl->jindex == NULL)
4090 snew(nl->jindex, 1);
4095 /* Sets a simple list i-cell bounding box, including PBC shift */
4096 static gmx_inline void set_icell_bb_simple(const nbnxn_bb_t *bb, int ci,
4097 real shx, real shy, real shz,
4100 bb_ci->lower[BB_X] = bb[ci].lower[BB_X] + shx;
4101 bb_ci->lower[BB_Y] = bb[ci].lower[BB_Y] + shy;
4102 bb_ci->lower[BB_Z] = bb[ci].lower[BB_Z] + shz;
4103 bb_ci->upper[BB_X] = bb[ci].upper[BB_X] + shx;
4104 bb_ci->upper[BB_Y] = bb[ci].upper[BB_Y] + shy;
4105 bb_ci->upper[BB_Z] = bb[ci].upper[BB_Z] + shz;
4109 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
4110 static void set_icell_bbxxxx_supersub(const float *bb, int ci,
4111 real shx, real shy, real shz,
4116 ia = ci*(GPU_NSUBCELL>>STRIDE_PBB_2LOG)*NNBSBB_XXXX;
4117 for (m = 0; m < (GPU_NSUBCELL>>STRIDE_PBB_2LOG)*NNBSBB_XXXX; m += NNBSBB_XXXX)
4119 for (i = 0; i < STRIDE_PBB; i++)
4121 bb_ci[m+0*STRIDE_PBB+i] = bb[ia+m+0*STRIDE_PBB+i] + shx;
4122 bb_ci[m+1*STRIDE_PBB+i] = bb[ia+m+1*STRIDE_PBB+i] + shy;
4123 bb_ci[m+2*STRIDE_PBB+i] = bb[ia+m+2*STRIDE_PBB+i] + shz;
4124 bb_ci[m+3*STRIDE_PBB+i] = bb[ia+m+3*STRIDE_PBB+i] + shx;
4125 bb_ci[m+4*STRIDE_PBB+i] = bb[ia+m+4*STRIDE_PBB+i] + shy;
4126 bb_ci[m+5*STRIDE_PBB+i] = bb[ia+m+5*STRIDE_PBB+i] + shz;
4132 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
4133 static void set_icell_bb_supersub(const nbnxn_bb_t *bb, int ci,
4134 real shx, real shy, real shz,
4139 for (i = 0; i < GPU_NSUBCELL; i++)
4141 set_icell_bb_simple(bb, ci*GPU_NSUBCELL+i,
4147 /* Copies PBC shifted i-cell atom coordinates x,y,z to working array */
4148 static void icell_set_x_simple(int ci,
4149 real shx, real shy, real shz,
4150 int gmx_unused na_c,
4151 int stride, const real *x,
4152 nbnxn_list_work_t *work)
4156 ia = ci*NBNXN_CPU_CLUSTER_I_SIZE;
4158 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE; i++)
4160 work->x_ci[i*STRIDE_XYZ+XX] = x[(ia+i)*stride+XX] + shx;
4161 work->x_ci[i*STRIDE_XYZ+YY] = x[(ia+i)*stride+YY] + shy;
4162 work->x_ci[i*STRIDE_XYZ+ZZ] = x[(ia+i)*stride+ZZ] + shz;
4166 /* Copies PBC shifted super-cell atom coordinates x,y,z to working array */
4167 static void icell_set_x_supersub(int ci,
4168 real shx, real shy, real shz,
4170 int stride, const real *x,
4171 nbnxn_list_work_t *work)
4178 ia = ci*GPU_NSUBCELL*na_c;
4179 for (i = 0; i < GPU_NSUBCELL*na_c; i++)
4181 x_ci[i*DIM + XX] = x[(ia+i)*stride + XX] + shx;
4182 x_ci[i*DIM + YY] = x[(ia+i)*stride + YY] + shy;
4183 x_ci[i*DIM + ZZ] = x[(ia+i)*stride + ZZ] + shz;
4187 #ifdef NBNXN_SEARCH_BB_SIMD4
4188 /* Copies PBC shifted super-cell packed atom coordinates to working array */
4189 static void icell_set_x_supersub_simd4(int ci,
4190 real shx, real shy, real shz,
4192 int stride, const real *x,
4193 nbnxn_list_work_t *work)
4195 int si, io, ia, i, j;
4200 for (si = 0; si < GPU_NSUBCELL; si++)
4202 for (i = 0; i < na_c; i += STRIDE_PBB)
4205 ia = ci*GPU_NSUBCELL*na_c + io;
4206 for (j = 0; j < STRIDE_PBB; j++)
4208 x_ci[io*DIM + j + XX*STRIDE_PBB] = x[(ia+j)*stride+XX] + shx;
4209 x_ci[io*DIM + j + YY*STRIDE_PBB] = x[(ia+j)*stride+YY] + shy;
4210 x_ci[io*DIM + j + ZZ*STRIDE_PBB] = x[(ia+j)*stride+ZZ] + shz;
4217 static real minimum_subgrid_size_xy(const nbnxn_grid_t *grid)
4221 return min(grid->sx, grid->sy);
4225 return min(grid->sx/GPU_NSUBCELL_X, grid->sy/GPU_NSUBCELL_Y);
4229 static real effective_buffer_1x1_vs_MxN(const nbnxn_grid_t *gridi,
4230 const nbnxn_grid_t *gridj)
4232 const real eff_1x1_buffer_fac_overest = 0.1;
4234 /* Determine an atom-pair list cut-off buffer size for atom pairs,
4235 * to be added to rlist (including buffer) used for MxN.
4236 * This is for converting an MxN list to a 1x1 list. This means we can't
4237 * use the normal buffer estimate, as we have an MxN list in which
4238 * some atom pairs beyond rlist are missing. We want to capture
4239 * the beneficial effect of buffering by extra pairs just outside rlist,
4240 * while removing the useless pairs that are further away from rlist.
4241 * (Also the buffer could have been set manually not using the estimate.)
4242 * This buffer size is an overestimate.
4243 * We add 10% of the smallest grid sub-cell dimensions.
4244 * Note that the z-size differs per cell and we don't use this,
4245 * so we overestimate.
4246 * With PME, the 10% value gives a buffer that is somewhat larger
4247 * than the effective buffer with a tolerance of 0.005 kJ/mol/ps.
4248 * Smaller tolerances or using RF lead to a smaller effective buffer,
4249 * so 10% gives a safe overestimate.
4251 return eff_1x1_buffer_fac_overest*(minimum_subgrid_size_xy(gridi) +
4252 minimum_subgrid_size_xy(gridj));
4255 /* Clusters at the cut-off only increase rlist by 60% of their size */
4256 static real nbnxn_rlist_inc_outside_fac = 0.6;
4258 /* Due to the cluster size the effective pair-list is longer than
4259 * that of a simple atom pair-list. This function gives the extra distance.
4261 real nbnxn_get_rlist_effective_inc(int cluster_size_j, real atom_density)
4264 real vol_inc_i, vol_inc_j;
4266 /* We should get this from the setup, but currently it's the same for
4267 * all setups, including GPUs.
4269 cluster_size_i = NBNXN_CPU_CLUSTER_I_SIZE;
4271 vol_inc_i = (cluster_size_i - 1)/atom_density;
4272 vol_inc_j = (cluster_size_j - 1)/atom_density;
4274 return nbnxn_rlist_inc_outside_fac*pow(vol_inc_i + vol_inc_j, 1.0/3.0);
4277 /* Estimates the interaction volume^2 for non-local interactions */
4278 static real nonlocal_vol2(const gmx_domdec_zones_t *zones, rvec ls, real r)
4287 /* Here we simply add up the volumes of 1, 2 or 3 1D decomposition
4288 * not home interaction volume^2. As these volumes are not additive,
4289 * this is an overestimate, but it would only be significant in the limit
4290 * of small cells, where we anyhow need to split the lists into
4291 * as small parts as possible.
4294 for (z = 0; z < zones->n; z++)
4296 if (zones->shift[z][XX] + zones->shift[z][YY] + zones->shift[z][ZZ] == 1)
4301 for (d = 0; d < DIM; d++)
4303 if (zones->shift[z][d] == 0)
4307 za *= zones->size[z].x1[d] - zones->size[z].x0[d];
4311 /* 4 octants of a sphere */
4312 vold_est = 0.25*M_PI*r*r*r*r;
4313 /* 4 quarter pie slices on the edges */
4314 vold_est += 4*cl*M_PI/6.0*r*r*r;
4315 /* One rectangular volume on a face */
4316 vold_est += ca*0.5*r*r;
4318 vol2_est_tot += vold_est*za;
4322 return vol2_est_tot;
4325 /* Estimates the average size of a full j-list for super/sub setup */
4326 static int get_nsubpair_max(const nbnxn_search_t nbs,
4329 int min_ci_balanced)
4331 const nbnxn_grid_t *grid;
4333 real xy_diag2, r_eff_sup, vol_est, nsp_est, nsp_est_nl;
4336 grid = &nbs->grid[0];
4338 if (min_ci_balanced <= 0 || grid->nc >= min_ci_balanced || grid->nc == 0)
4340 /* We don't need to worry */
4344 ls[XX] = (grid->c1[XX] - grid->c0[XX])/(grid->ncx*GPU_NSUBCELL_X);
4345 ls[YY] = (grid->c1[YY] - grid->c0[YY])/(grid->ncy*GPU_NSUBCELL_Y);
4346 ls[ZZ] = (grid->c1[ZZ] - grid->c0[ZZ])*grid->ncx*grid->ncy/(grid->nc*GPU_NSUBCELL_Z);
4348 /* The average squared length of the diagonal of a sub cell */
4349 xy_diag2 = ls[XX]*ls[XX] + ls[YY]*ls[YY] + ls[ZZ]*ls[ZZ];
4351 /* The formulas below are a heuristic estimate of the average nsj per si*/
4352 r_eff_sup = rlist + nbnxn_rlist_inc_outside_fac*sqr((grid->na_c - 1.0)/grid->na_c)*sqrt(xy_diag2/3);
4354 if (!nbs->DomDec || nbs->zones->n == 1)
4361 sqr(grid->atom_density/grid->na_c)*
4362 nonlocal_vol2(nbs->zones, ls, r_eff_sup);
4367 /* Sub-cell interacts with itself */
4368 vol_est = ls[XX]*ls[YY]*ls[ZZ];
4369 /* 6/2 rectangular volume on the faces */
4370 vol_est += (ls[XX]*ls[YY] + ls[XX]*ls[ZZ] + ls[YY]*ls[ZZ])*r_eff_sup;
4371 /* 12/2 quarter pie slices on the edges */
4372 vol_est += 2*(ls[XX] + ls[YY] + ls[ZZ])*0.25*M_PI*sqr(r_eff_sup);
4373 /* 4 octants of a sphere */
4374 vol_est += 0.5*4.0/3.0*M_PI*pow(r_eff_sup, 3);
4376 nsp_est = grid->nsubc_tot*vol_est*grid->atom_density/grid->na_c;
4378 /* Subtract the non-local pair count */
4379 nsp_est -= nsp_est_nl;
4383 fprintf(debug, "nsp_est local %5.1f non-local %5.1f\n",
4384 nsp_est, nsp_est_nl);
4389 nsp_est = nsp_est_nl;
4392 /* Thus the (average) maximum j-list size should be as follows */
4393 nsubpair_max = max(1, (int)(nsp_est/min_ci_balanced+0.5));
4395 /* Since the target value is a maximum (this avoids high outliers,
4396 * which lead to load imbalance), not average, we add half the
4397 * number of pairs in a cj4 block to get the average about right.
4399 nsubpair_max += GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE/2;
4403 fprintf(debug, "nbl nsp estimate %.1f, nsubpair_max %d\n",
4404 nsp_est, nsubpair_max);
4407 return nsubpair_max;
4410 /* Debug list print function */
4411 static void print_nblist_ci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
4415 for (i = 0; i < nbl->nci; i++)
4417 fprintf(fp, "ci %4d shift %2d ncj %3d\n",
4418 nbl->ci[i].ci, nbl->ci[i].shift,
4419 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start);
4421 for (j = nbl->ci[i].cj_ind_start; j < nbl->ci[i].cj_ind_end; j++)
4423 fprintf(fp, " cj %5d imask %x\n",
4430 /* Debug list print function */
4431 static void print_nblist_sci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
4433 int i, j4, j, ncp, si;
4435 for (i = 0; i < nbl->nsci; i++)
4437 fprintf(fp, "ci %4d shift %2d ncj4 %2d\n",
4438 nbl->sci[i].sci, nbl->sci[i].shift,
4439 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start);
4442 for (j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
4444 for (j = 0; j < NBNXN_GPU_JGROUP_SIZE; j++)
4446 fprintf(fp, " sj %5d imask %x\n",
4448 nbl->cj4[j4].imei[0].imask);
4449 for (si = 0; si < GPU_NSUBCELL; si++)
4451 if (nbl->cj4[j4].imei[0].imask & (1U << (j*GPU_NSUBCELL + si)))
4458 fprintf(fp, "ci %4d shift %2d ncj4 %2d ncp %3d\n",
4459 nbl->sci[i].sci, nbl->sci[i].shift,
4460 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start,
4465 /* Combine pair lists *nbl generated on multiple threads nblc */
4466 static void combine_nblists(int nnbl, nbnxn_pairlist_t **nbl,
4467 nbnxn_pairlist_t *nblc)
4469 int nsci, ncj4, nexcl;
4471 int nthreads gmx_unused;
4475 gmx_incons("combine_nblists does not support simple lists");
4480 nexcl = nblc->nexcl;
4481 for (i = 0; i < nnbl; i++)
4483 nsci += nbl[i]->nsci;
4484 ncj4 += nbl[i]->ncj4;
4485 nexcl += nbl[i]->nexcl;
4488 if (nsci > nblc->sci_nalloc)
4490 nb_realloc_sci(nblc, nsci);
4492 if (ncj4 > nblc->cj4_nalloc)
4494 nblc->cj4_nalloc = over_alloc_small(ncj4);
4495 nbnxn_realloc_void((void **)&nblc->cj4,
4496 nblc->ncj4*sizeof(*nblc->cj4),
4497 nblc->cj4_nalloc*sizeof(*nblc->cj4),
4498 nblc->alloc, nblc->free);
4500 if (nexcl > nblc->excl_nalloc)
4502 nblc->excl_nalloc = over_alloc_small(nexcl);
4503 nbnxn_realloc_void((void **)&nblc->excl,
4504 nblc->nexcl*sizeof(*nblc->excl),
4505 nblc->excl_nalloc*sizeof(*nblc->excl),
4506 nblc->alloc, nblc->free);
4509 /* Each thread should copy its own data to the combined arrays,
4510 * as otherwise data will go back and forth between different caches.
4512 nthreads = gmx_omp_nthreads_get(emntPairsearch);
4513 #pragma omp parallel for num_threads(nthreads) schedule(static)
4514 for (n = 0; n < nnbl; n++)
4521 const nbnxn_pairlist_t *nbli;
4523 /* Determine the offset in the combined data for our thread */
4524 sci_offset = nblc->nsci;
4525 cj4_offset = nblc->ncj4;
4526 ci_offset = nblc->nci_tot;
4527 excl_offset = nblc->nexcl;
4529 for (i = 0; i < n; i++)
4531 sci_offset += nbl[i]->nsci;
4532 cj4_offset += nbl[i]->ncj4;
4533 ci_offset += nbl[i]->nci_tot;
4534 excl_offset += nbl[i]->nexcl;
4539 for (i = 0; i < nbli->nsci; i++)
4541 nblc->sci[sci_offset+i] = nbli->sci[i];
4542 nblc->sci[sci_offset+i].cj4_ind_start += cj4_offset;
4543 nblc->sci[sci_offset+i].cj4_ind_end += cj4_offset;
4546 for (j4 = 0; j4 < nbli->ncj4; j4++)
4548 nblc->cj4[cj4_offset+j4] = nbli->cj4[j4];
4549 nblc->cj4[cj4_offset+j4].imei[0].excl_ind += excl_offset;
4550 nblc->cj4[cj4_offset+j4].imei[1].excl_ind += excl_offset;
4553 for (j4 = 0; j4 < nbli->nexcl; j4++)
4555 nblc->excl[excl_offset+j4] = nbli->excl[j4];
4559 for (n = 0; n < nnbl; n++)
4561 nblc->nsci += nbl[n]->nsci;
4562 nblc->ncj4 += nbl[n]->ncj4;
4563 nblc->nci_tot += nbl[n]->nci_tot;
4564 nblc->nexcl += nbl[n]->nexcl;
4568 static void balance_fep_lists(const nbnxn_search_t nbs,
4569 nbnxn_pairlist_set_t *nbl_lists)
4572 int nri_tot, nrj_tot, nrj_target;
4576 nnbl = nbl_lists->nnbl;
4580 /* Nothing to balance */
4584 /* Count the total i-lists and pairs */
4587 for (th = 0; th < nnbl; th++)
4589 nri_tot += nbl_lists->nbl_fep[th]->nri;
4590 nrj_tot += nbl_lists->nbl_fep[th]->nrj;
4593 nrj_target = (nrj_tot + nnbl - 1)/nnbl;
4595 assert(gmx_omp_nthreads_get(emntNonbonded) == nnbl);
4597 #pragma omp parallel for schedule(static) num_threads(nnbl)
4598 for (th = 0; th < nnbl; th++)
4602 nbl = nbs->work[th].nbl_fep;
4604 /* Note that here we allocate for the total size, instead of
4605 * a per-thread esimate (which is hard to obtain).
4607 if (nri_tot > nbl->maxnri)
4609 nbl->maxnri = over_alloc_large(nri_tot);
4610 reallocate_nblist(nbl);
4612 if (nri_tot > nbl->maxnri || nrj_tot > nbl->maxnrj)
4614 nbl->maxnrj = over_alloc_small(nrj_tot);
4615 srenew(nbl->jjnr, nbl->maxnrj);
4616 srenew(nbl->excl_fep, nbl->maxnrj);
4619 clear_pairlist_fep(nbl);
4622 /* Loop over the source lists and assign and copy i-entries */
4624 nbld = nbs->work[th_dest].nbl_fep;
4625 for (th = 0; th < nnbl; th++)
4630 nbls = nbl_lists->nbl_fep[th];
4632 for (i = 0; i < nbls->nri; i++)
4636 /* The number of pairs in this i-entry */
4637 nrj = nbls->jindex[i+1] - nbls->jindex[i];
4639 /* Decide if list th_dest is too large and we should procede
4640 * to the next destination list.
4642 if (th_dest+1 < nnbl && nbld->nrj > 0 &&
4643 nbld->nrj + nrj - nrj_target > nrj_target - nbld->nrj)
4646 nbld = nbs->work[th_dest].nbl_fep;
4649 nbld->iinr[nbld->nri] = nbls->iinr[i];
4650 nbld->gid[nbld->nri] = nbls->gid[i];
4651 nbld->shift[nbld->nri] = nbls->shift[i];
4653 for (j = nbls->jindex[i]; j < nbls->jindex[i+1]; j++)
4655 nbld->jjnr[nbld->nrj] = nbls->jjnr[j];
4656 nbld->excl_fep[nbld->nrj] = nbls->excl_fep[j];
4660 nbld->jindex[nbld->nri] = nbld->nrj;
4664 /* Swap the list pointers */
4665 for (th = 0; th < nnbl; th++)
4669 nbl_tmp = nbl_lists->nbl_fep[th];
4670 nbl_lists->nbl_fep[th] = nbs->work[th].nbl_fep;
4671 nbs->work[th].nbl_fep = nbl_tmp;
4675 fprintf(debug, "nbl_fep[%d] nri %4d nrj %4d\n",
4677 nbl_lists->nbl_fep[th]->nri,
4678 nbl_lists->nbl_fep[th]->nrj);
4683 /* Returns the next ci to be processes by our thread */
4684 static gmx_bool next_ci(const nbnxn_grid_t *grid,
4686 int nth, int ci_block,
4687 int *ci_x, int *ci_y,
4693 if (*ci_b == ci_block)
4695 /* Jump to the next block assigned to this task */
4696 *ci += (nth - 1)*ci_block;
4700 if (*ci >= grid->nc*conv)
4705 while (*ci >= grid->cxy_ind[*ci_x*grid->ncy + *ci_y + 1]*conv)
4708 if (*ci_y == grid->ncy)
4718 /* Returns the distance^2 for which we put cell pairs in the list
4719 * without checking atom pair distances. This is usually < rlist^2.
4721 static float boundingbox_only_distance2(const nbnxn_grid_t *gridi,
4722 const nbnxn_grid_t *gridj,
4726 /* If the distance between two sub-cell bounding boxes is less
4727 * than this distance, do not check the distance between
4728 * all particle pairs in the sub-cell, since then it is likely
4729 * that the box pair has atom pairs within the cut-off.
4730 * We use the nblist cut-off minus 0.5 times the average x/y diagonal
4731 * spacing of the sub-cells. Around 40% of the checked pairs are pruned.
4732 * Using more than 0.5 gains at most 0.5%.
4733 * If forces are calculated more than twice, the performance gain
4734 * in the force calculation outweighs the cost of checking.
4735 * Note that with subcell lists, the atom-pair distance check
4736 * is only performed when only 1 out of 8 sub-cells in within range,
4737 * this is because the GPU is much faster than the cpu.
4742 bbx = 0.5*(gridi->sx + gridj->sx);
4743 bby = 0.5*(gridi->sy + gridj->sy);
4746 bbx /= GPU_NSUBCELL_X;
4747 bby /= GPU_NSUBCELL_Y;
4750 rbb2 = sqr(max(0, rlist - 0.5*sqrt(bbx*bbx + bby*bby)));
4755 return (float)((1+GMX_FLOAT_EPS)*rbb2);
4759 static int get_ci_block_size(const nbnxn_grid_t *gridi,
4760 gmx_bool bDomDec, int nth)
4762 const int ci_block_enum = 5;
4763 const int ci_block_denom = 11;
4764 const int ci_block_min_atoms = 16;
4767 /* Here we decide how to distribute the blocks over the threads.
4768 * We use prime numbers to try to avoid that the grid size becomes
4769 * a multiple of the number of threads, which would lead to some
4770 * threads getting "inner" pairs and others getting boundary pairs,
4771 * which in turns will lead to load imbalance between threads.
4772 * Set the block size as 5/11/ntask times the average number of cells
4773 * in a y,z slab. This should ensure a quite uniform distribution
4774 * of the grid parts of the different thread along all three grid
4775 * zone boundaries with 3D domain decomposition. At the same time
4776 * the blocks will not become too small.
4778 ci_block = (gridi->nc*ci_block_enum)/(ci_block_denom*gridi->ncx*nth);
4780 /* Ensure the blocks are not too small: avoids cache invalidation */
4781 if (ci_block*gridi->na_sc < ci_block_min_atoms)
4783 ci_block = (ci_block_min_atoms + gridi->na_sc - 1)/gridi->na_sc;
4786 /* Without domain decomposition
4787 * or with less than 3 blocks per task, divide in nth blocks.
4789 if (!bDomDec || ci_block*3*nth > gridi->nc)
4791 ci_block = (gridi->nc + nth - 1)/nth;
4797 /* Generates the part of pair-list nbl assigned to our thread */
4798 static void nbnxn_make_pairlist_part(const nbnxn_search_t nbs,
4799 const nbnxn_grid_t *gridi,
4800 const nbnxn_grid_t *gridj,
4801 nbnxn_search_work_t *work,
4802 const nbnxn_atomdata_t *nbat,
4803 const t_blocka *excl,
4807 gmx_bool bFBufferFlag,
4810 int min_ci_balanced,
4812 nbnxn_pairlist_t *nbl,
4817 real rl2, rl_fep2 = 0;
4820 int ci_b, ci, ci_x, ci_y, ci_xy, cj;
4826 int conv_i, cell0_i;
4827 const nbnxn_bb_t *bb_i = NULL;
4829 const float *pbb_i = NULL;
4831 const float *bbcz_i, *bbcz_j;
4833 real bx0, bx1, by0, by1, bz0, bz1;
4835 real d2cx, d2z, d2z_cx, d2z_cy, d2zx, d2zxy, d2xy;
4836 int cxf, cxl, cyf, cyf_x, cyl;
4838 int c0, c1, cs, cf, cl;
4841 int gridi_flag_shift = 0, gridj_flag_shift = 0;
4842 gmx_bitmask_t *gridj_flag = NULL;
4843 int ncj_old_i, ncj_old_j;
4845 nbs_cycle_start(&work->cc[enbsCCsearch]);
4847 if (gridj->bSimple != nbl->bSimple)
4849 gmx_incons("Grid incompatible with pair-list");
4853 nbl->na_sc = gridj->na_sc;
4854 nbl->na_ci = gridj->na_c;
4855 nbl->na_cj = nbnxn_kernel_to_cj_size(nb_kernel_type);
4856 na_cj_2log = get_2log(nbl->na_cj);
4862 /* Determine conversion of clusters to flag blocks */
4863 gridi_flag_shift = 0;
4864 while ((nbl->na_ci<<gridi_flag_shift) < NBNXN_BUFFERFLAG_SIZE)
4868 gridj_flag_shift = 0;
4869 while ((nbl->na_cj<<gridj_flag_shift) < NBNXN_BUFFERFLAG_SIZE)
4874 gridj_flag = work->buffer_flags.flag;
4877 copy_mat(nbs->box, box);
4879 rl2 = nbl->rlist*nbl->rlist;
4881 if (nbs->bFEP && !nbl->bSimple)
4883 /* Determine an atom-pair list cut-off distance for FEP atom pairs.
4884 * We should not simply use rlist, since then we would not have
4885 * the small, effective buffering of the NxN lists.
4886 * The buffer is on overestimate, but the resulting cost for pairs
4887 * beyond rlist is neglible compared to the FEP pairs within rlist.
4889 rl_fep2 = nbl->rlist + effective_buffer_1x1_vs_MxN(gridi, gridj);
4893 fprintf(debug, "nbl_fep atom-pair rlist %f\n", rl_fep2);
4895 rl_fep2 = rl_fep2*rl_fep2;
4898 rbb2 = boundingbox_only_distance2(gridi, gridj, nbl->rlist, nbl->bSimple);
4902 fprintf(debug, "nbl bounding box only distance %f\n", sqrt(rbb2));
4905 /* Set the shift range */
4906 for (d = 0; d < DIM; d++)
4908 /* Check if we need periodicity shifts.
4909 * Without PBC or with domain decomposition we don't need them.
4911 if (d >= ePBC2npbcdim(nbs->ePBC) || nbs->dd_dim[d])
4918 box[XX][XX] - fabs(box[YY][XX]) - fabs(box[ZZ][XX]) < sqrt(rl2))
4929 if (nbl->bSimple && !gridi->bSimple)
4931 conv_i = gridi->na_sc/gridj->na_sc;
4932 bb_i = gridi->bb_simple;
4933 bbcz_i = gridi->bbcz_simple;
4934 flags_i = gridi->flags_simple;
4949 /* We use the normal bounding box format for both grid types */
4952 bbcz_i = gridi->bbcz;
4953 flags_i = gridi->flags;
4955 cell0_i = gridi->cell0*conv_i;
4957 bbcz_j = gridj->bbcz;
4961 /* Blocks of the conversion factor - 1 give a large repeat count
4962 * combined with a small block size. This should result in good
4963 * load balancing for both small and large domains.
4965 ci_block = conv_i - 1;
4969 fprintf(debug, "nbl nc_i %d col.av. %.1f ci_block %d\n",
4970 gridi->nc, gridi->nc/(double)(gridi->ncx*gridi->ncy), ci_block);
4976 /* Initially ci_b and ci to 1 before where we want them to start,
4977 * as they will both be incremented in next_ci.
4980 ci = th*ci_block - 1;
4983 while (next_ci(gridi, conv_i, nth, ci_block, &ci_x, &ci_y, &ci_b, &ci))
4985 if (nbl->bSimple && flags_i[ci] == 0)
4990 ncj_old_i = nbl->ncj;
4993 if (gridj != gridi && shp[XX] == 0)
4997 bx1 = bb_i[ci].upper[BB_X];
5001 bx1 = gridi->c0[XX] + (ci_x+1)*gridi->sx;
5003 if (bx1 < gridj->c0[XX])
5005 d2cx = sqr(gridj->c0[XX] - bx1);
5014 ci_xy = ci_x*gridi->ncy + ci_y;
5016 /* Loop over shift vectors in three dimensions */
5017 for (tz = -shp[ZZ]; tz <= shp[ZZ]; tz++)
5019 shz = tz*box[ZZ][ZZ];
5021 bz0 = bbcz_i[ci*NNBSBB_D ] + shz;
5022 bz1 = bbcz_i[ci*NNBSBB_D+1] + shz;
5034 d2z = sqr(bz0 - box[ZZ][ZZ]);
5037 d2z_cx = d2z + d2cx;
5045 bz1/((real)(gridi->cxy_ind[ci_xy+1] - gridi->cxy_ind[ci_xy]));
5050 /* The check with bz1_frac close to or larger than 1 comes later */
5052 for (ty = -shp[YY]; ty <= shp[YY]; ty++)
5054 shy = ty*box[YY][YY] + tz*box[ZZ][YY];
5058 by0 = bb_i[ci].lower[BB_Y] + shy;
5059 by1 = bb_i[ci].upper[BB_Y] + shy;
5063 by0 = gridi->c0[YY] + (ci_y )*gridi->sy + shy;
5064 by1 = gridi->c0[YY] + (ci_y+1)*gridi->sy + shy;
5067 get_cell_range(by0, by1,
5068 gridj->ncy, gridj->c0[YY], gridj->sy, gridj->inv_sy,
5078 if (by1 < gridj->c0[YY])
5080 d2z_cy += sqr(gridj->c0[YY] - by1);
5082 else if (by0 > gridj->c1[YY])
5084 d2z_cy += sqr(by0 - gridj->c1[YY]);
5087 for (tx = -shp[XX]; tx <= shp[XX]; tx++)
5089 shift = XYZ2IS(tx, ty, tz);
5091 #ifdef NBNXN_SHIFT_BACKWARD
5092 if (gridi == gridj && shift > CENTRAL)
5098 shx = tx*box[XX][XX] + ty*box[YY][XX] + tz*box[ZZ][XX];
5102 bx0 = bb_i[ci].lower[BB_X] + shx;
5103 bx1 = bb_i[ci].upper[BB_X] + shx;
5107 bx0 = gridi->c0[XX] + (ci_x )*gridi->sx + shx;
5108 bx1 = gridi->c0[XX] + (ci_x+1)*gridi->sx + shx;
5111 get_cell_range(bx0, bx1,
5112 gridj->ncx, gridj->c0[XX], gridj->sx, gridj->inv_sx,
5123 new_ci_entry(nbl, cell0_i+ci, shift, flags_i[ci]);
5127 new_sci_entry(nbl, cell0_i+ci, shift);
5130 #ifndef NBNXN_SHIFT_BACKWARD
5133 if (shift == CENTRAL && gridi == gridj &&
5137 /* Leave the pairs with i > j.
5138 * x is the major index, so skip half of it.
5145 set_icell_bb_simple(bb_i, ci, shx, shy, shz,
5151 set_icell_bbxxxx_supersub(pbb_i, ci, shx, shy, shz,
5154 set_icell_bb_supersub(bb_i, ci, shx, shy, shz,
5159 nbs->icell_set_x(cell0_i+ci, shx, shy, shz,
5160 gridi->na_c, nbat->xstride, nbat->x,
5163 for (cx = cxf; cx <= cxl; cx++)
5166 if (gridj->c0[XX] + cx*gridj->sx > bx1)
5168 d2zx += sqr(gridj->c0[XX] + cx*gridj->sx - bx1);
5170 else if (gridj->c0[XX] + (cx+1)*gridj->sx < bx0)
5172 d2zx += sqr(gridj->c0[XX] + (cx+1)*gridj->sx - bx0);
5175 #ifndef NBNXN_SHIFT_BACKWARD
5176 if (gridi == gridj &&
5177 cx == 0 && cyf < ci_y)
5179 if (gridi == gridj &&
5180 cx == 0 && shift == CENTRAL && cyf < ci_y)
5183 /* Leave the pairs with i > j.
5184 * Skip half of y when i and j have the same x.
5193 for (cy = cyf_x; cy <= cyl; cy++)
5195 c0 = gridj->cxy_ind[cx*gridj->ncy+cy];
5196 c1 = gridj->cxy_ind[cx*gridj->ncy+cy+1];
5197 #ifdef NBNXN_SHIFT_BACKWARD
5198 if (gridi == gridj &&
5199 shift == CENTRAL && c0 < ci)
5206 if (gridj->c0[YY] + cy*gridj->sy > by1)
5208 d2zxy += sqr(gridj->c0[YY] + cy*gridj->sy - by1);
5210 else if (gridj->c0[YY] + (cy+1)*gridj->sy < by0)
5212 d2zxy += sqr(gridj->c0[YY] + (cy+1)*gridj->sy - by0);
5214 if (c1 > c0 && d2zxy < rl2)
5216 cs = c0 + (int)(bz1_frac*(c1 - c0));
5224 /* Find the lowest cell that can possibly
5229 (bbcz_j[cf*NNBSBB_D+1] >= bz0 ||
5230 d2xy + sqr(bbcz_j[cf*NNBSBB_D+1] - bz0) < rl2))
5235 /* Find the highest cell that can possibly
5240 (bbcz_j[cl*NNBSBB_D] <= bz1 ||
5241 d2xy + sqr(bbcz_j[cl*NNBSBB_D] - bz1) < rl2))
5246 #ifdef NBNXN_REFCODE
5248 /* Simple reference code, for debugging,
5249 * overrides the more complex code above.
5254 for (k = c0; k < c1; k++)
5256 if (box_dist2(bx0, bx1, by0, by1, bz0, bz1, bb+k) < rl2 &&
5261 if (box_dist2(bx0, bx1, by0, by1, bz0, bz1, bb+k) < rl2 &&
5272 /* We want each atom/cell pair only once,
5273 * only use cj >= ci.
5275 #ifndef NBNXN_SHIFT_BACKWARD
5278 if (shift == CENTRAL)
5287 /* For f buffer flags with simple lists */
5288 ncj_old_j = nbl->ncj;
5290 switch (nb_kernel_type)
5292 case nbnxnk4x4_PlainC:
5293 check_subcell_list_space_simple(nbl, cl-cf+1);
5295 make_cluster_list_simple(gridj,
5297 (gridi == gridj && shift == CENTRAL),
5302 #ifdef GMX_NBNXN_SIMD_4XN
5303 case nbnxnk4xN_SIMD_4xN:
5304 check_subcell_list_space_simple(nbl, ci_to_cj(na_cj_2log, cl-cf)+2);
5305 make_cluster_list_simd_4xn(gridj,
5307 (gridi == gridj && shift == CENTRAL),
5313 #ifdef GMX_NBNXN_SIMD_2XNN
5314 case nbnxnk4xN_SIMD_2xNN:
5315 check_subcell_list_space_simple(nbl, ci_to_cj(na_cj_2log, cl-cf)+2);
5316 make_cluster_list_simd_2xnn(gridj,
5318 (gridi == gridj && shift == CENTRAL),
5324 case nbnxnk8x8x8_PlainC:
5325 case nbnxnk8x8x8_CUDA:
5326 check_subcell_list_space_supersub(nbl, cl-cf+1);
5327 for (cj = cf; cj <= cl; cj++)
5329 make_cluster_list_supersub(gridi, gridj,
5331 (gridi == gridj && shift == CENTRAL && ci == cj),
5332 nbat->xstride, nbat->x,
5338 ncpcheck += cl - cf + 1;
5340 if (bFBufferFlag && nbl->ncj > ncj_old_j)
5344 cbf = nbl->cj[ncj_old_j].cj >> gridj_flag_shift;
5345 cbl = nbl->cj[nbl->ncj-1].cj >> gridj_flag_shift;
5346 for (cb = cbf; cb <= cbl; cb++)
5348 bitmask_init_bit(&gridj_flag[cb], th);
5356 /* Set the exclusions for this ci list */
5359 set_ci_top_excls(nbs,
5361 shift == CENTRAL && gridi == gridj,
5364 &(nbl->ci[nbl->nci]),
5369 make_fep_list(nbs, nbat, nbl,
5370 shift == CENTRAL && gridi == gridj,
5371 &(nbl->ci[nbl->nci]),
5372 gridi, gridj, nbl_fep);
5377 set_sci_top_excls(nbs,
5379 shift == CENTRAL && gridi == gridj,
5381 &(nbl->sci[nbl->nsci]),
5386 make_fep_list_supersub(nbs, nbat, nbl,
5387 shift == CENTRAL && gridi == gridj,
5388 &(nbl->sci[nbl->nsci]),
5391 gridi, gridj, nbl_fep);
5395 /* Close this ci list */
5398 close_ci_entry_simple(nbl);
5402 close_ci_entry_supersub(nbl,
5404 progBal, min_ci_balanced,
5411 if (bFBufferFlag && nbl->ncj > ncj_old_i)
5413 bitmask_init_bit(&(work->buffer_flags.flag[(gridi->cell0+ci)>>gridi_flag_shift]), th);
5417 work->ndistc = ndistc;
5419 nbs_cycle_stop(&work->cc[enbsCCsearch]);
5423 fprintf(debug, "number of distance checks %d\n", ndistc);
5424 fprintf(debug, "ncpcheck %s %d\n", gridi == gridj ? "local" : "non-local",
5429 print_nblist_statistics_simple(debug, nbl, nbs, rlist);
5433 print_nblist_statistics_supersub(debug, nbl, nbs, rlist);
5438 fprintf(debug, "nbl FEP list pairs: %d\n", nbl_fep->nrj);
5443 static void reduce_buffer_flags(const nbnxn_search_t nbs,
5445 const nbnxn_buffer_flags_t *dest)
5448 gmx_bitmask_t *flag;
5450 for (s = 0; s < nsrc; s++)
5452 flag = nbs->work[s].buffer_flags.flag;
5454 for (b = 0; b < dest->nflag; b++)
5456 bitmask_union(&(dest->flag[b]), flag[b]);
5461 static void print_reduction_cost(const nbnxn_buffer_flags_t *flags, int nout)
5463 int nelem, nkeep, ncopy, nred, b, c, out;
5464 gmx_bitmask_t mask_0;
5470 bitmask_init_bit(&mask_0, 0);
5471 for (b = 0; b < flags->nflag; b++)
5473 if (bitmask_is_equal(flags->flag[b], mask_0))
5475 /* Only flag 0 is set, no copy of reduction required */
5479 else if (!bitmask_is_zero(flags->flag[b]))
5482 for (out = 0; out < nout; out++)
5484 if (bitmask_is_set(flags->flag[b], out))
5501 fprintf(debug, "nbnxn reduction: #flag %d #list %d elem %4.2f, keep %4.2f copy %4.2f red %4.2f\n",
5503 nelem/(double)(flags->nflag),
5504 nkeep/(double)(flags->nflag),
5505 ncopy/(double)(flags->nflag),
5506 nred/(double)(flags->nflag));
5509 /* Perform a count (linear) sort to sort the smaller lists to the end.
5510 * This avoids load imbalance on the GPU, as large lists will be
5511 * scheduled and executed first and the smaller lists later.
5512 * Load balancing between multi-processors only happens at the end
5513 * and there smaller lists lead to more effective load balancing.
5514 * The sorting is done on the cj4 count, not on the actual pair counts.
5515 * Not only does this make the sort faster, but it also results in
5516 * better load balancing than using a list sorted on exact load.
5517 * This function swaps the pointer in the pair list to avoid a copy operation.
5519 static void sort_sci(nbnxn_pairlist_t *nbl)
5521 nbnxn_list_work_t *work;
5522 int m, i, s, s0, s1;
5523 nbnxn_sci_t *sci_sort;
5525 if (nbl->ncj4 <= nbl->nsci)
5527 /* nsci = 0 or all sci have size 1, sorting won't change the order */
5533 /* We will distinguish differences up to double the average */
5534 m = (2*nbl->ncj4)/nbl->nsci;
5536 if (m + 1 > work->sort_nalloc)
5538 work->sort_nalloc = over_alloc_large(m + 1);
5539 srenew(work->sort, work->sort_nalloc);
5542 if (work->sci_sort_nalloc != nbl->sci_nalloc)
5544 work->sci_sort_nalloc = nbl->sci_nalloc;
5545 nbnxn_realloc_void((void **)&work->sci_sort,
5547 work->sci_sort_nalloc*sizeof(*work->sci_sort),
5548 nbl->alloc, nbl->free);
5551 /* Count the entries of each size */
5552 for (i = 0; i <= m; i++)
5556 for (s = 0; s < nbl->nsci; s++)
5558 i = min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start);
5561 /* Calculate the offset for each count */
5564 for (i = m - 1; i >= 0; i--)
5567 work->sort[i] = work->sort[i + 1] + s0;
5571 /* Sort entries directly into place */
5572 sci_sort = work->sci_sort;
5573 for (s = 0; s < nbl->nsci; s++)
5575 i = min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start);
5576 sci_sort[work->sort[i]++] = nbl->sci[s];
5579 /* Swap the sci pointers so we use the new, sorted list */
5580 work->sci_sort = nbl->sci;
5581 nbl->sci = sci_sort;
5584 /* Make a local or non-local pair-list, depending on iloc */
5585 void nbnxn_make_pairlist(const nbnxn_search_t nbs,
5586 nbnxn_atomdata_t *nbat,
5587 const t_blocka *excl,
5589 int min_ci_balanced,
5590 nbnxn_pairlist_set_t *nbl_list,
5595 nbnxn_grid_t *gridi, *gridj;
5597 int nzi, zi, zj0, zj1, zj;
5601 nbnxn_pairlist_t **nbl;
5603 gmx_bool CombineNBLists;
5605 int np_tot, np_noq, np_hlj, nap;
5607 /* Check if we are running hybrid GPU + CPU nbnxn mode */
5608 bGPUCPU = (!nbs->grid[0].bSimple && nbl_list->bSimple);
5610 nnbl = nbl_list->nnbl;
5611 nbl = nbl_list->nbl;
5612 CombineNBLists = nbl_list->bCombined;
5616 fprintf(debug, "ns making %d nblists\n", nnbl);
5619 nbat->bUseBufferFlags = (nbat->nout > 1);
5620 /* We should re-init the flags before making the first list */
5621 if (nbat->bUseBufferFlags && (LOCAL_I(iloc) || bGPUCPU))
5623 init_buffer_flags(&nbat->buffer_flags, nbat->natoms);
5626 if (nbl_list->bSimple)
5628 switch (nb_kernel_type)
5630 #ifdef GMX_NBNXN_SIMD_4XN
5631 case nbnxnk4xN_SIMD_4xN:
5632 nbs->icell_set_x = icell_set_x_simd_4xn;
5635 #ifdef GMX_NBNXN_SIMD_2XNN
5636 case nbnxnk4xN_SIMD_2xNN:
5637 nbs->icell_set_x = icell_set_x_simd_2xnn;
5641 nbs->icell_set_x = icell_set_x_simple;
5647 #ifdef NBNXN_SEARCH_BB_SIMD4
5648 nbs->icell_set_x = icell_set_x_supersub_simd4;
5650 nbs->icell_set_x = icell_set_x_supersub;
5656 /* Only zone (grid) 0 vs 0 */
5663 nzi = nbs->zones->nizone;
5666 if (!nbl_list->bSimple && min_ci_balanced > 0)
5668 nsubpair_max = get_nsubpair_max(nbs, iloc, rlist, min_ci_balanced);
5675 /* Clear all pair-lists */
5676 for (th = 0; th < nnbl; th++)
5678 clear_pairlist(nbl[th]);
5682 clear_pairlist_fep(nbl_list->nbl_fep[th]);
5686 for (zi = 0; zi < nzi; zi++)
5688 gridi = &nbs->grid[zi];
5690 if (NONLOCAL_I(iloc))
5692 zj0 = nbs->zones->izone[zi].j0;
5693 zj1 = nbs->zones->izone[zi].j1;
5699 for (zj = zj0; zj < zj1; zj++)
5701 gridj = &nbs->grid[zj];
5705 fprintf(debug, "ns search grid %d vs %d\n", zi, zj);
5708 nbs_cycle_start(&nbs->cc[enbsCCsearch]);
5710 if (nbl[0]->bSimple && !gridi->bSimple)
5712 /* Hybrid list, determine blocking later */
5717 ci_block = get_ci_block_size(gridi, nbs->DomDec, nnbl);
5720 /* With GPU: generate progressively smaller lists for
5721 * load balancing for local only or non-local with 2 zones.
5723 progBal = (LOCAL_I(iloc) || nbs->zones->n <= 2);
5725 #pragma omp parallel for num_threads(nnbl) schedule(static)
5726 for (th = 0; th < nnbl; th++)
5728 /* Re-init the thread-local work flag data before making
5729 * the first list (not an elegant conditional).
5731 if (nbat->bUseBufferFlags && ((zi == 0 && zj == 0) ||
5732 (bGPUCPU && zi == 0 && zj == 1)))
5734 init_buffer_flags(&nbs->work[th].buffer_flags, nbat->natoms);
5737 if (CombineNBLists && th > 0)
5739 clear_pairlist(nbl[th]);
5742 /* Divide the i super cell equally over the nblists */
5743 nbnxn_make_pairlist_part(nbs, gridi, gridj,
5744 &nbs->work[th], nbat, excl,
5748 nbat->bUseBufferFlags,
5750 progBal, min_ci_balanced,
5753 nbl_list->nbl_fep[th]);
5755 nbs_cycle_stop(&nbs->cc[enbsCCsearch]);
5760 for (th = 0; th < nnbl; th++)
5762 inc_nrnb(nrnb, eNR_NBNXN_DIST2, nbs->work[th].ndistc);
5764 if (nbl_list->bSimple)
5766 np_tot += nbl[th]->ncj;
5767 np_noq += nbl[th]->work->ncj_noq;
5768 np_hlj += nbl[th]->work->ncj_hlj;
5772 /* This count ignores potential subsequent pair pruning */
5773 np_tot += nbl[th]->nci_tot;
5776 nap = nbl[0]->na_ci*nbl[0]->na_cj;
5777 nbl_list->natpair_ljq = (np_tot - np_noq)*nap - np_hlj*nap/2;
5778 nbl_list->natpair_lj = np_noq*nap;
5779 nbl_list->natpair_q = np_hlj*nap/2;
5781 if (CombineNBLists && nnbl > 1)
5783 nbs_cycle_start(&nbs->cc[enbsCCcombine]);
5785 combine_nblists(nnbl-1, nbl+1, nbl[0]);
5787 nbs_cycle_stop(&nbs->cc[enbsCCcombine]);
5792 if (!nbl_list->bSimple)
5794 /* Sort the entries on size, large ones first */
5795 if (CombineNBLists || nnbl == 1)
5801 #pragma omp parallel for num_threads(nnbl) schedule(static)
5802 for (th = 0; th < nnbl; th++)
5809 if (nbat->bUseBufferFlags)
5811 reduce_buffer_flags(nbs, nnbl, &nbat->buffer_flags);
5816 /* Balance the free-energy lists over all the threads */
5817 balance_fep_lists(nbs, nbl_list);
5820 /* Special performance logging stuff (env.var. GMX_NBNXN_CYCLE) */
5823 nbs->search_count++;
5825 if (nbs->print_cycles &&
5826 (!nbs->DomDec || (nbs->DomDec && !LOCAL_I(iloc))) &&
5827 nbs->search_count % 100 == 0)
5829 nbs_cycle_print(stderr, nbs);
5832 if (debug && (CombineNBLists && nnbl > 1))
5834 if (nbl[0]->bSimple)
5836 print_nblist_statistics_simple(debug, nbl[0], nbs, rlist);
5840 print_nblist_statistics_supersub(debug, nbl[0], nbs, rlist);
5848 if (nbl[0]->bSimple)
5850 print_nblist_ci_cj(debug, nbl[0]);
5854 print_nblist_sci_cj(debug, nbl[0]);
5858 if (nbat->bUseBufferFlags)
5860 print_reduction_cost(&nbat->buffer_flags, nnbl);