2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
44 #include "gromacs/utility/smalloc.h"
45 #include "types/commrec.h"
47 #include "gromacs/math/utilities.h"
48 #include "gromacs/math/vec.h"
49 #include "gromacs/pbcutil/pbc.h"
50 #include "nbnxn_consts.h"
51 /* nbnxn_internal.h included gromacs/simd/macros.h */
52 #include "nbnxn_internal.h"
54 #include "gromacs/simd/vector_operations.h"
56 #include "nbnxn_atomdata.h"
57 #include "nbnxn_search.h"
58 #include "gmx_omp_nthreads.h"
62 #include "gromacs/fileio/gmxfio.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)
421 rvec_sub(corner1, corner0, size);
423 return n/(size[XX]*size[YY]*size[ZZ]);
426 static int set_grid_size_xy(const nbnxn_search_t nbs,
429 int n, rvec corner0, rvec corner1,
434 real adens, tlen, tlen_x, tlen_y, nc_max;
437 rvec_sub(corner1, corner0, size);
441 /* target cell length */
444 /* To minimize the zero interactions, we should make
445 * the largest of the i/j cell cubic.
447 na_c = max(grid->na_c, grid->na_cj);
449 /* Approximately cubic cells */
450 tlen = pow(na_c/atom_density, 1.0/3.0);
456 /* Approximately cubic sub cells */
457 tlen = pow(grid->na_c/atom_density, 1.0/3.0);
458 tlen_x = tlen*GPU_NSUBCELL_X;
459 tlen_y = tlen*GPU_NSUBCELL_Y;
461 /* We round ncx and ncy down, because we get less cell pairs
462 * in the nbsist when the fixed cell dimensions (x,y) are
463 * larger than the variable one (z) than the other way around.
465 grid->ncx = max(1, (int)(size[XX]/tlen_x));
466 grid->ncy = max(1, (int)(size[YY]/tlen_y));
474 grid->sx = size[XX]/grid->ncx;
475 grid->sy = size[YY]/grid->ncy;
476 grid->inv_sx = 1/grid->sx;
477 grid->inv_sy = 1/grid->sy;
481 /* This is a non-home zone, add an extra row of cells
482 * for particles communicated for bonded interactions.
483 * These can be beyond the cut-off. It doesn't matter where
484 * they end up on the grid, but for performance it's better
485 * if they don't end up in cells that can be within cut-off range.
491 /* We need one additional cell entry for particles moved by DD */
492 if (grid->ncx*grid->ncy+1 > grid->cxy_nalloc)
494 grid->cxy_nalloc = over_alloc_large(grid->ncx*grid->ncy+1);
495 srenew(grid->cxy_na, grid->cxy_nalloc);
496 srenew(grid->cxy_ind, grid->cxy_nalloc+1);
498 for (t = 0; t < nbs->nthread_max; t++)
500 if (grid->ncx*grid->ncy+1 > nbs->work[t].cxy_na_nalloc)
502 nbs->work[t].cxy_na_nalloc = over_alloc_large(grid->ncx*grid->ncy+1);
503 srenew(nbs->work[t].cxy_na, nbs->work[t].cxy_na_nalloc);
507 /* Worst case scenario of 1 atom in each last cell */
508 if (grid->na_cj <= grid->na_c)
510 nc_max = n/grid->na_sc + grid->ncx*grid->ncy;
514 nc_max = n/grid->na_sc + grid->ncx*grid->ncy*grid->na_cj/grid->na_c;
517 if (nc_max > grid->nc_nalloc)
519 grid->nc_nalloc = over_alloc_large(nc_max);
520 srenew(grid->nsubc, grid->nc_nalloc);
521 srenew(grid->bbcz, grid->nc_nalloc*NNBSBB_D);
523 sfree_aligned(grid->bb);
524 /* This snew also zeros the contents, this avoid possible
525 * floating exceptions in SIMD with the unused bb elements.
529 snew_aligned(grid->bb, grid->nc_nalloc, 16);
536 pbb_nalloc = grid->nc_nalloc*GPU_NSUBCELL/STRIDE_PBB*NNBSBB_XXXX;
537 snew_aligned(grid->pbb, pbb_nalloc, 16);
539 snew_aligned(grid->bb, grid->nc_nalloc*GPU_NSUBCELL, 16);
545 if (grid->na_cj == grid->na_c)
547 grid->bbj = grid->bb;
551 sfree_aligned(grid->bbj);
552 snew_aligned(grid->bbj, grid->nc_nalloc*grid->na_c/grid->na_cj, 16);
556 srenew(grid->flags, grid->nc_nalloc);
559 srenew(grid->fep, grid->nc_nalloc*grid->na_sc/grid->na_c);
563 copy_rvec(corner0, grid->c0);
564 copy_rvec(corner1, grid->c1);
569 /* We need to sort paricles in grid columns on z-coordinate.
570 * As particle are very often distributed homogeneously, we a sorting
571 * algorithm similar to pigeonhole sort. We multiply the z-coordinate
572 * by a factor, cast to an int and try to store in that hole. If the hole
573 * is full, we move this or another particle. A second pass is needed to make
574 * contiguous elements. SORT_GRID_OVERSIZE is the ratio of holes to particles.
575 * 4 is the optimal value for homogeneous particle distribution and allows
576 * for an O(#particles) sort up till distributions were all particles are
577 * concentrated in 1/4 of the space. No NlogN fallback is implemented,
578 * as it can be expensive to detect imhomogeneous particle distributions.
579 * SGSF is the maximum ratio of holes used, in the worst case all particles
580 * end up in the last hole and we need #particles extra holes at the end.
582 #define SORT_GRID_OVERSIZE 4
583 #define SGSF (SORT_GRID_OVERSIZE + 1)
585 /* Sort particle index a on coordinates x along dim.
586 * Backwards tells if we want decreasing iso increasing coordinates.
587 * h0 is the minimum of the coordinate range.
588 * invh is the 1/length of the sorting range.
589 * n_per_h (>=n) is the expected average number of particles per 1/invh
590 * sort is the sorting work array.
591 * sort should have a size of at least n_per_h*SORT_GRID_OVERSIZE + n,
592 * or easier, allocate at least n*SGSF elements.
594 static void sort_atoms(int dim, gmx_bool Backwards,
595 int gmx_unused dd_zone,
596 int *a, int n, rvec *x,
597 real h0, real invh, int n_per_h,
601 int zi, zim, zi_min, zi_max;
613 gmx_incons("n > n_per_h");
617 /* Transform the inverse range height into the inverse hole height */
618 invh *= n_per_h*SORT_GRID_OVERSIZE;
620 /* Set nsort to the maximum possible number of holes used.
621 * In worst case all n elements end up in the last bin.
623 nsort = n_per_h*SORT_GRID_OVERSIZE + n;
625 /* Determine the index range used, so we can limit it for the second pass */
629 /* Sort the particles using a simple index sort */
630 for (i = 0; i < n; i++)
632 /* The cast takes care of float-point rounding effects below zero.
633 * This code assumes particles are less than 1/SORT_GRID_OVERSIZE
634 * times the box height out of the box.
636 zi = (int)((x[a[i]][dim] - h0)*invh);
639 /* As we can have rounding effect, we use > iso >= here */
640 if (zi < 0 || (dd_zone == 0 && zi > n_per_h*SORT_GRID_OVERSIZE))
642 gmx_fatal(FARGS, "(int)((x[%d][%c]=%f - %f)*%f) = %d, not in 0 - %d*%d\n",
643 a[i], 'x'+dim, x[a[i]][dim], h0, invh, zi,
644 n_per_h, SORT_GRID_OVERSIZE);
648 /* In a non-local domain, particles communcated for bonded interactions
649 * can be far beyond the grid size, which is set by the non-bonded
650 * cut-off distance. We sort such particles into the last cell.
652 if (zi > n_per_h*SORT_GRID_OVERSIZE)
654 zi = n_per_h*SORT_GRID_OVERSIZE;
657 /* Ideally this particle should go in sort cell zi,
658 * but that might already be in use,
659 * in that case find the first empty cell higher up
664 zi_min = min(zi_min, zi);
665 zi_max = max(zi_max, zi);
669 /* We have multiple atoms in the same sorting slot.
670 * Sort on real z for minimal bounding box size.
671 * There is an extra check for identical z to ensure
672 * well-defined output order, independent of input order
673 * to ensure binary reproducibility after restarts.
675 while (sort[zi] >= 0 && ( x[a[i]][dim] > x[sort[zi]][dim] ||
676 (x[a[i]][dim] == x[sort[zi]][dim] &&
684 /* Shift all elements by one slot until we find an empty slot */
687 while (sort[zim] >= 0)
695 zi_max = max(zi_max, zim);
698 zi_max = max(zi_max, zi);
705 for (zi = 0; zi < nsort; zi++)
716 for (zi = zi_max; zi >= zi_min; zi--)
727 gmx_incons("Lost particles while sorting");
732 #define R2F_D(x) ((float)((x) >= 0 ? ((1-GMX_FLOAT_EPS)*(x)) : ((1+GMX_FLOAT_EPS)*(x))))
733 #define R2F_U(x) ((float)((x) >= 0 ? ((1+GMX_FLOAT_EPS)*(x)) : ((1-GMX_FLOAT_EPS)*(x))))
739 /* Coordinate order x,y,z, bb order xyz0 */
740 static void calc_bounding_box(int na, int stride, const real *x, nbnxn_bb_t *bb)
743 real xl, xh, yl, yh, zl, zh;
753 for (j = 1; j < na; j++)
755 xl = min(xl, x[i+XX]);
756 xh = max(xh, x[i+XX]);
757 yl = min(yl, x[i+YY]);
758 yh = max(yh, x[i+YY]);
759 zl = min(zl, x[i+ZZ]);
760 zh = max(zh, x[i+ZZ]);
763 /* Note: possible double to float conversion here */
764 bb->lower[BB_X] = R2F_D(xl);
765 bb->lower[BB_Y] = R2F_D(yl);
766 bb->lower[BB_Z] = R2F_D(zl);
767 bb->upper[BB_X] = R2F_U(xh);
768 bb->upper[BB_Y] = R2F_U(yh);
769 bb->upper[BB_Z] = R2F_U(zh);
772 /* Packed coordinates, bb order xyz0 */
773 static void calc_bounding_box_x_x4(int na, const real *x, nbnxn_bb_t *bb)
776 real xl, xh, yl, yh, zl, zh;
784 for (j = 1; j < na; j++)
786 xl = min(xl, x[j+XX*PACK_X4]);
787 xh = max(xh, x[j+XX*PACK_X4]);
788 yl = min(yl, x[j+YY*PACK_X4]);
789 yh = max(yh, x[j+YY*PACK_X4]);
790 zl = min(zl, x[j+ZZ*PACK_X4]);
791 zh = max(zh, x[j+ZZ*PACK_X4]);
793 /* Note: possible double to float conversion here */
794 bb->lower[BB_X] = R2F_D(xl);
795 bb->lower[BB_Y] = R2F_D(yl);
796 bb->lower[BB_Z] = R2F_D(zl);
797 bb->upper[BB_X] = R2F_U(xh);
798 bb->upper[BB_Y] = R2F_U(yh);
799 bb->upper[BB_Z] = R2F_U(zh);
802 /* Packed coordinates, bb order xyz0 */
803 static void calc_bounding_box_x_x8(int na, const real *x, nbnxn_bb_t *bb)
806 real xl, xh, yl, yh, zl, zh;
814 for (j = 1; j < na; j++)
816 xl = min(xl, x[j+XX*PACK_X8]);
817 xh = max(xh, x[j+XX*PACK_X8]);
818 yl = min(yl, x[j+YY*PACK_X8]);
819 yh = max(yh, x[j+YY*PACK_X8]);
820 zl = min(zl, x[j+ZZ*PACK_X8]);
821 zh = max(zh, x[j+ZZ*PACK_X8]);
823 /* Note: possible double to float conversion here */
824 bb->lower[BB_X] = R2F_D(xl);
825 bb->lower[BB_Y] = R2F_D(yl);
826 bb->lower[BB_Z] = R2F_D(zl);
827 bb->upper[BB_X] = R2F_U(xh);
828 bb->upper[BB_Y] = R2F_U(yh);
829 bb->upper[BB_Z] = R2F_U(zh);
832 /* Packed coordinates, bb order xyz0 */
833 static void calc_bounding_box_x_x4_halves(int na, const real *x,
834 nbnxn_bb_t *bb, nbnxn_bb_t *bbj)
836 calc_bounding_box_x_x4(min(na, 2), x, bbj);
840 calc_bounding_box_x_x4(min(na-2, 2), x+(PACK_X4>>1), bbj+1);
844 /* Set the "empty" bounding box to the same as the first one,
845 * so we don't need to treat special cases in the rest of the code.
847 #ifdef NBNXN_SEARCH_BB_SIMD4
848 gmx_simd4_store_f(&bbj[1].lower[0], gmx_simd4_load_f(&bbj[0].lower[0]));
849 gmx_simd4_store_f(&bbj[1].upper[0], gmx_simd4_load_f(&bbj[0].upper[0]));
855 #ifdef NBNXN_SEARCH_BB_SIMD4
856 gmx_simd4_store_f(&bb->lower[0],
857 gmx_simd4_min_f(gmx_simd4_load_f(&bbj[0].lower[0]),
858 gmx_simd4_load_f(&bbj[1].lower[0])));
859 gmx_simd4_store_f(&bb->upper[0],
860 gmx_simd4_max_f(gmx_simd4_load_f(&bbj[0].upper[0]),
861 gmx_simd4_load_f(&bbj[1].upper[0])));
866 for (i = 0; i < NNBSBB_C; i++)
868 bb->lower[i] = min(bbj[0].lower[i], bbj[1].lower[i]);
869 bb->upper[i] = max(bbj[0].upper[i], bbj[1].upper[i]);
875 #ifdef NBNXN_SEARCH_BB_SIMD4
877 /* Coordinate order xyz, bb order xxxxyyyyzzzz */
878 static void calc_bounding_box_xxxx(int na, int stride, const real *x, float *bb)
881 real xl, xh, yl, yh, zl, zh;
891 for (j = 1; j < na; j++)
893 xl = min(xl, x[i+XX]);
894 xh = max(xh, x[i+XX]);
895 yl = min(yl, x[i+YY]);
896 yh = max(yh, x[i+YY]);
897 zl = min(zl, x[i+ZZ]);
898 zh = max(zh, x[i+ZZ]);
901 /* Note: possible double to float conversion here */
902 bb[0*STRIDE_PBB] = R2F_D(xl);
903 bb[1*STRIDE_PBB] = R2F_D(yl);
904 bb[2*STRIDE_PBB] = R2F_D(zl);
905 bb[3*STRIDE_PBB] = R2F_U(xh);
906 bb[4*STRIDE_PBB] = R2F_U(yh);
907 bb[5*STRIDE_PBB] = R2F_U(zh);
910 #endif /* NBNXN_SEARCH_BB_SIMD4 */
912 #ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
914 /* Coordinate order xyz?, bb order xyz0 */
915 static void calc_bounding_box_simd4(int na, const float *x, nbnxn_bb_t *bb)
917 gmx_simd4_float_t bb_0_S, bb_1_S;
918 gmx_simd4_float_t x_S;
922 bb_0_S = gmx_simd4_load_f(x);
925 for (i = 1; i < na; i++)
927 x_S = gmx_simd4_load_f(x+i*NNBSBB_C);
928 bb_0_S = gmx_simd4_min_f(bb_0_S, x_S);
929 bb_1_S = gmx_simd4_max_f(bb_1_S, x_S);
932 gmx_simd4_store_f(&bb->lower[0], bb_0_S);
933 gmx_simd4_store_f(&bb->upper[0], bb_1_S);
936 /* Coordinate order xyz?, bb order xxxxyyyyzzzz */
937 static void calc_bounding_box_xxxx_simd4(int na, const float *x,
938 nbnxn_bb_t *bb_work_aligned,
941 calc_bounding_box_simd4(na, x, bb_work_aligned);
943 bb[0*STRIDE_PBB] = bb_work_aligned->lower[BB_X];
944 bb[1*STRIDE_PBB] = bb_work_aligned->lower[BB_Y];
945 bb[2*STRIDE_PBB] = bb_work_aligned->lower[BB_Z];
946 bb[3*STRIDE_PBB] = bb_work_aligned->upper[BB_X];
947 bb[4*STRIDE_PBB] = bb_work_aligned->upper[BB_Y];
948 bb[5*STRIDE_PBB] = bb_work_aligned->upper[BB_Z];
951 #endif /* NBNXN_SEARCH_SIMD4_FLOAT_X_BB */
954 /* Combines pairs of consecutive bounding boxes */
955 static void combine_bounding_box_pairs(nbnxn_grid_t *grid, const nbnxn_bb_t *bb)
957 int i, j, sc2, nc2, c2;
959 for (i = 0; i < grid->ncx*grid->ncy; i++)
961 /* Starting bb in a column is expected to be 2-aligned */
962 sc2 = grid->cxy_ind[i]>>1;
963 /* For odd numbers skip the last bb here */
964 nc2 = (grid->cxy_na[i]+3)>>(2+1);
965 for (c2 = sc2; c2 < sc2+nc2; c2++)
967 #ifdef NBNXN_SEARCH_BB_SIMD4
968 gmx_simd4_float_t min_S, max_S;
970 min_S = gmx_simd4_min_f(gmx_simd4_load_f(&bb[c2*2+0].lower[0]),
971 gmx_simd4_load_f(&bb[c2*2+1].lower[0]));
972 max_S = gmx_simd4_max_f(gmx_simd4_load_f(&bb[c2*2+0].upper[0]),
973 gmx_simd4_load_f(&bb[c2*2+1].upper[0]));
974 gmx_simd4_store_f(&grid->bbj[c2].lower[0], min_S);
975 gmx_simd4_store_f(&grid->bbj[c2].upper[0], max_S);
977 for (j = 0; j < NNBSBB_C; j++)
979 grid->bbj[c2].lower[j] = min(bb[c2*2+0].lower[j],
980 bb[c2*2+1].lower[j]);
981 grid->bbj[c2].upper[j] = max(bb[c2*2+0].upper[j],
982 bb[c2*2+1].upper[j]);
986 if (((grid->cxy_na[i]+3)>>2) & 1)
988 /* The bb count in this column is odd: duplicate the last bb */
989 for (j = 0; j < NNBSBB_C; j++)
991 grid->bbj[c2].lower[j] = bb[c2*2].lower[j];
992 grid->bbj[c2].upper[j] = bb[c2*2].upper[j];
999 /* Prints the average bb size, used for debug output */
1000 static void print_bbsizes_simple(FILE *fp,
1001 const nbnxn_search_t nbs,
1002 const nbnxn_grid_t *grid)
1008 for (c = 0; c < grid->nc; c++)
1010 for (d = 0; d < DIM; d++)
1012 ba[d] += grid->bb[c].upper[d] - grid->bb[c].lower[d];
1015 dsvmul(1.0/grid->nc, ba, ba);
1017 fprintf(fp, "ns bb: %4.2f %4.2f %4.2f %4.2f %4.2f %4.2f rel %4.2f %4.2f %4.2f\n",
1018 nbs->box[XX][XX]/grid->ncx,
1019 nbs->box[YY][YY]/grid->ncy,
1020 nbs->box[ZZ][ZZ]*grid->ncx*grid->ncy/grid->nc,
1021 ba[XX], ba[YY], ba[ZZ],
1022 ba[XX]*grid->ncx/nbs->box[XX][XX],
1023 ba[YY]*grid->ncy/nbs->box[YY][YY],
1024 ba[ZZ]*grid->nc/(grid->ncx*grid->ncy*nbs->box[ZZ][ZZ]));
1027 /* Prints the average bb size, used for debug output */
1028 static void print_bbsizes_supersub(FILE *fp,
1029 const nbnxn_search_t nbs,
1030 const nbnxn_grid_t *grid)
1037 for (c = 0; c < grid->nc; c++)
1040 for (s = 0; s < grid->nsubc[c]; s += STRIDE_PBB)
1044 cs_w = (c*GPU_NSUBCELL + s)/STRIDE_PBB;
1045 for (i = 0; i < STRIDE_PBB; i++)
1047 for (d = 0; d < DIM; d++)
1050 grid->pbb[cs_w*NNBSBB_XXXX+(DIM+d)*STRIDE_PBB+i] -
1051 grid->pbb[cs_w*NNBSBB_XXXX+ d *STRIDE_PBB+i];
1056 for (s = 0; s < grid->nsubc[c]; s++)
1060 cs = c*GPU_NSUBCELL + s;
1061 for (d = 0; d < DIM; d++)
1063 ba[d] += grid->bb[cs].upper[d] - grid->bb[cs].lower[d];
1067 ns += grid->nsubc[c];
1069 dsvmul(1.0/ns, ba, ba);
1071 fprintf(fp, "ns bb: %4.2f %4.2f %4.2f %4.2f %4.2f %4.2f rel %4.2f %4.2f %4.2f\n",
1072 nbs->box[XX][XX]/(grid->ncx*GPU_NSUBCELL_X),
1073 nbs->box[YY][YY]/(grid->ncy*GPU_NSUBCELL_Y),
1074 nbs->box[ZZ][ZZ]*grid->ncx*grid->ncy/(grid->nc*GPU_NSUBCELL_Z),
1075 ba[XX], ba[YY], ba[ZZ],
1076 ba[XX]*grid->ncx*GPU_NSUBCELL_X/nbs->box[XX][XX],
1077 ba[YY]*grid->ncy*GPU_NSUBCELL_Y/nbs->box[YY][YY],
1078 ba[ZZ]*grid->nc*GPU_NSUBCELL_Z/(grid->ncx*grid->ncy*nbs->box[ZZ][ZZ]));
1081 /* Potentially sorts atoms on LJ coefficients !=0 and ==0.
1082 * Also sets interaction flags.
1084 void sort_on_lj(int na_c,
1085 int a0, int a1, const int *atinfo,
1089 int subc, s, a, n1, n2, a_lj_max, i, j;
1090 int sort1[NBNXN_NA_SC_MAX/GPU_NSUBCELL];
1091 int sort2[NBNXN_NA_SC_MAX/GPU_NSUBCELL];
1092 gmx_bool haveQ, bFEP;
1097 for (s = a0; s < a1; s += na_c)
1099 /* Make lists for this (sub-)cell on atoms with and without LJ */
1104 for (a = s; a < min(s+na_c, a1); a++)
1106 haveQ = haveQ || GET_CGINFO_HAS_Q(atinfo[order[a]]);
1108 if (GET_CGINFO_HAS_VDW(atinfo[order[a]]))
1110 sort1[n1++] = order[a];
1115 sort2[n2++] = order[a];
1119 /* If we don't have atoms with LJ, there's nothing to sort */
1122 *flags |= NBNXN_CI_DO_LJ(subc);
1126 /* Only sort when strictly necessary. Ordering particles
1127 * Ordering particles can lead to less accurate summation
1128 * due to rounding, both for LJ and Coulomb interactions.
1130 if (2*(a_lj_max - s) >= na_c)
1132 for (i = 0; i < n1; i++)
1134 order[a0+i] = sort1[i];
1136 for (j = 0; j < n2; j++)
1138 order[a0+n1+j] = sort2[j];
1142 *flags |= NBNXN_CI_HALF_LJ(subc);
1147 *flags |= NBNXN_CI_DO_COUL(subc);
1153 /* Fill a pair search cell with atoms.
1154 * Potentially sorts atoms and sets the interaction flags.
1156 void fill_cell(const nbnxn_search_t nbs,
1158 nbnxn_atomdata_t *nbat,
1162 int sx, int sy, int sz,
1163 nbnxn_bb_t gmx_unused *bb_work_aligned)
1176 sort_on_lj(grid->na_c, a0, a1, atinfo, nbs->a,
1177 grid->flags+(a0>>grid->na_c_2log)-grid->cell0);
1182 /* Set the fep flag for perturbed atoms in this (sub-)cell */
1185 /* The grid-local cluster/(sub-)cell index */
1186 c = (a0 >> grid->na_c_2log) - grid->cell0*(grid->bSimple ? 1 : GPU_NSUBCELL);
1188 for (at = a0; at < a1; at++)
1190 if (nbs->a[at] >= 0 && GET_CGINFO_FEP(atinfo[nbs->a[at]]))
1192 grid->fep[c] |= (1 << (at - a0));
1197 /* Now we have sorted the atoms, set the cell indices */
1198 for (a = a0; a < a1; a++)
1200 nbs->cell[nbs->a[a]] = a;
1203 copy_rvec_to_nbat_real(nbs->a+a0, a1-a0, grid->na_c, x,
1204 nbat->XFormat, nbat->x, a0,
1207 if (nbat->XFormat == nbatX4)
1209 /* Store the bounding boxes as xyz.xyz. */
1210 offset = (a0 - grid->cell0*grid->na_sc) >> grid->na_c_2log;
1211 bb_ptr = grid->bb + offset;
1213 #if defined GMX_NBNXN_SIMD && GMX_SIMD_REAL_WIDTH == 2
1214 if (2*grid->na_cj == grid->na_c)
1216 calc_bounding_box_x_x4_halves(na, nbat->x+X4_IND_A(a0), bb_ptr,
1217 grid->bbj+offset*2);
1222 calc_bounding_box_x_x4(na, nbat->x+X4_IND_A(a0), bb_ptr);
1225 else if (nbat->XFormat == nbatX8)
1227 /* Store the bounding boxes as xyz.xyz. */
1228 offset = (a0 - grid->cell0*grid->na_sc) >> grid->na_c_2log;
1229 bb_ptr = grid->bb + offset;
1231 calc_bounding_box_x_x8(na, nbat->x+X8_IND_A(a0), bb_ptr);
1234 else if (!grid->bSimple)
1236 /* Store the bounding boxes in a format convenient
1237 * for SIMD4 calculations: xxxxyyyyzzzz...
1241 ((a0-grid->cell0*grid->na_sc)>>(grid->na_c_2log+STRIDE_PBB_2LOG))*NNBSBB_XXXX +
1242 (((a0-grid->cell0*grid->na_sc)>>grid->na_c_2log) & (STRIDE_PBB-1));
1244 #ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
1245 if (nbat->XFormat == nbatXYZQ)
1247 calc_bounding_box_xxxx_simd4(na, nbat->x+a0*nbat->xstride,
1248 bb_work_aligned, pbb_ptr);
1253 calc_bounding_box_xxxx(na, nbat->xstride, nbat->x+a0*nbat->xstride,
1258 fprintf(debug, "%2d %2d %2d bb %5.2f %5.2f %5.2f %5.2f %5.2f %5.2f\n",
1260 pbb_ptr[0*STRIDE_PBB], pbb_ptr[3*STRIDE_PBB],
1261 pbb_ptr[1*STRIDE_PBB], pbb_ptr[4*STRIDE_PBB],
1262 pbb_ptr[2*STRIDE_PBB], pbb_ptr[5*STRIDE_PBB]);
1268 /* Store the bounding boxes as xyz.xyz. */
1269 bb_ptr = grid->bb+((a0-grid->cell0*grid->na_sc)>>grid->na_c_2log);
1271 calc_bounding_box(na, nbat->xstride, nbat->x+a0*nbat->xstride,
1277 bbo = (a0 - grid->cell0*grid->na_sc)/grid->na_c;
1278 fprintf(debug, "%2d %2d %2d bb %5.2f %5.2f %5.2f %5.2f %5.2f %5.2f\n",
1280 grid->bb[bbo].lower[BB_X],
1281 grid->bb[bbo].lower[BB_Y],
1282 grid->bb[bbo].lower[BB_Z],
1283 grid->bb[bbo].upper[BB_X],
1284 grid->bb[bbo].upper[BB_Y],
1285 grid->bb[bbo].upper[BB_Z]);
1290 /* Spatially sort the atoms within one grid column */
1291 static void sort_columns_simple(const nbnxn_search_t nbs,
1297 nbnxn_atomdata_t *nbat,
1298 int cxy_start, int cxy_end,
1302 int cx, cy, cz, ncz, cfilled, c;
1303 int na, ash, ind, a;
1308 fprintf(debug, "cell0 %d sorting columns %d - %d, atoms %d - %d\n",
1309 grid->cell0, cxy_start, cxy_end, a0, a1);
1312 /* Sort the atoms within each x,y column in 3 dimensions */
1313 for (cxy = cxy_start; cxy < cxy_end; cxy++)
1316 cy = cxy - cx*grid->ncy;
1318 na = grid->cxy_na[cxy];
1319 ncz = grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy];
1320 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1322 /* Sort the atoms within each x,y column on z coordinate */
1323 sort_atoms(ZZ, FALSE, dd_zone,
1326 1.0/nbs->box[ZZ][ZZ], ncz*grid->na_sc,
1329 /* Fill the ncz cells in this column */
1330 cfilled = grid->cxy_ind[cxy];
1331 for (cz = 0; cz < ncz; cz++)
1333 c = grid->cxy_ind[cxy] + cz;
1335 ash_c = ash + cz*grid->na_sc;
1336 na_c = min(grid->na_sc, na-(ash_c-ash));
1338 fill_cell(nbs, grid, nbat,
1339 ash_c, ash_c+na_c, atinfo, x,
1340 grid->na_sc*cx + (dd_zone >> 2),
1341 grid->na_sc*cy + (dd_zone & 3),
1345 /* This copy to bbcz is not really necessary.
1346 * But it allows to use the same grid search code
1347 * for the simple and supersub cell setups.
1353 grid->bbcz[c*NNBSBB_D ] = grid->bb[cfilled].lower[BB_Z];
1354 grid->bbcz[c*NNBSBB_D+1] = grid->bb[cfilled].upper[BB_Z];
1357 /* Set the unused atom indices to -1 */
1358 for (ind = na; ind < ncz*grid->na_sc; ind++)
1360 nbs->a[ash+ind] = -1;
1365 /* Spatially sort the atoms within one grid column */
1366 static void sort_columns_supersub(const nbnxn_search_t nbs,
1372 nbnxn_atomdata_t *nbat,
1373 int cxy_start, int cxy_end,
1377 int cx, cy, cz = -1, c = -1, ncz;
1378 int na, ash, na_c, ind, a;
1379 int subdiv_z, sub_z, na_z, ash_z;
1380 int subdiv_y, sub_y, na_y, ash_y;
1381 int subdiv_x, sub_x, na_x, ash_x;
1383 nbnxn_bb_t bb_work_array[2], *bb_work_aligned;
1385 bb_work_aligned = (nbnxn_bb_t *)(((size_t)(bb_work_array+1)) & (~((size_t)15)));
1389 fprintf(debug, "cell0 %d sorting columns %d - %d, atoms %d - %d\n",
1390 grid->cell0, cxy_start, cxy_end, a0, a1);
1393 subdiv_x = grid->na_c;
1394 subdiv_y = GPU_NSUBCELL_X*subdiv_x;
1395 subdiv_z = GPU_NSUBCELL_Y*subdiv_y;
1397 /* Sort the atoms within each x,y column in 3 dimensions */
1398 for (cxy = cxy_start; cxy < cxy_end; cxy++)
1401 cy = cxy - cx*grid->ncy;
1403 na = grid->cxy_na[cxy];
1404 ncz = grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy];
1405 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1407 /* Sort the atoms within each x,y column on z coordinate */
1408 sort_atoms(ZZ, FALSE, dd_zone,
1411 1.0/nbs->box[ZZ][ZZ], ncz*grid->na_sc,
1414 /* This loop goes over the supercells and subcells along z at once */
1415 for (sub_z = 0; sub_z < ncz*GPU_NSUBCELL_Z; sub_z++)
1417 ash_z = ash + sub_z*subdiv_z;
1418 na_z = min(subdiv_z, na-(ash_z-ash));
1420 /* We have already sorted on z */
1422 if (sub_z % GPU_NSUBCELL_Z == 0)
1424 cz = sub_z/GPU_NSUBCELL_Z;
1425 c = grid->cxy_ind[cxy] + cz;
1427 /* The number of atoms in this supercell */
1428 na_c = min(grid->na_sc, na-(ash_z-ash));
1430 grid->nsubc[c] = min(GPU_NSUBCELL, (na_c+grid->na_c-1)/grid->na_c);
1432 /* Store the z-boundaries of the super cell */
1433 grid->bbcz[c*NNBSBB_D ] = x[nbs->a[ash_z]][ZZ];
1434 grid->bbcz[c*NNBSBB_D+1] = x[nbs->a[ash_z+na_c-1]][ZZ];
1437 #if GPU_NSUBCELL_Y > 1
1438 /* Sort the atoms along y */
1439 sort_atoms(YY, (sub_z & 1), dd_zone,
1440 nbs->a+ash_z, na_z, x,
1441 grid->c0[YY]+cy*grid->sy,
1442 grid->inv_sy, subdiv_z,
1446 for (sub_y = 0; sub_y < GPU_NSUBCELL_Y; sub_y++)
1448 ash_y = ash_z + sub_y*subdiv_y;
1449 na_y = min(subdiv_y, na-(ash_y-ash));
1451 #if GPU_NSUBCELL_X > 1
1452 /* Sort the atoms along x */
1453 sort_atoms(XX, ((cz*GPU_NSUBCELL_Y + sub_y) & 1), dd_zone,
1454 nbs->a+ash_y, na_y, x,
1455 grid->c0[XX]+cx*grid->sx,
1456 grid->inv_sx, subdiv_y,
1460 for (sub_x = 0; sub_x < GPU_NSUBCELL_X; sub_x++)
1462 ash_x = ash_y + sub_x*subdiv_x;
1463 na_x = min(subdiv_x, na-(ash_x-ash));
1465 fill_cell(nbs, grid, nbat,
1466 ash_x, ash_x+na_x, atinfo, x,
1467 grid->na_c*(cx*GPU_NSUBCELL_X+sub_x) + (dd_zone >> 2),
1468 grid->na_c*(cy*GPU_NSUBCELL_Y+sub_y) + (dd_zone & 3),
1475 /* Set the unused atom indices to -1 */
1476 for (ind = na; ind < ncz*grid->na_sc; ind++)
1478 nbs->a[ash+ind] = -1;
1483 /* Determine in which grid column atoms should go */
1484 static void calc_column_indices(nbnxn_grid_t *grid,
1487 int dd_zone, const int *move,
1488 int thread, int nthread,
1495 /* We add one extra cell for particles which moved during DD */
1496 for (i = 0; i < grid->ncx*grid->ncy+1; i++)
1501 n0 = a0 + (int)((thread+0)*(a1 - a0))/nthread;
1502 n1 = a0 + (int)((thread+1)*(a1 - a0))/nthread;
1506 for (i = n0; i < n1; i++)
1508 if (move == NULL || move[i] >= 0)
1510 /* We need to be careful with rounding,
1511 * particles might be a few bits outside the local zone.
1512 * The int cast takes care of the lower bound,
1513 * we will explicitly take care of the upper bound.
1515 cx = (int)((x[i][XX] - grid->c0[XX])*grid->inv_sx);
1516 cy = (int)((x[i][YY] - grid->c0[YY])*grid->inv_sy);
1519 if (cx < 0 || cx > grid->ncx ||
1520 cy < 0 || cy > grid->ncy)
1523 "grid cell cx %d cy %d out of range (max %d %d)\n"
1524 "atom %f %f %f, grid->c0 %f %f",
1525 cx, cy, grid->ncx, grid->ncy,
1526 x[i][XX], x[i][YY], x[i][ZZ], grid->c0[XX], grid->c0[YY]);
1529 /* Take care of potential rouding issues */
1530 cx = min(cx, grid->ncx - 1);
1531 cy = min(cy, grid->ncy - 1);
1533 /* For the moment cell will contain only the, grid local,
1534 * x and y indices, not z.
1536 cell[i] = cx*grid->ncy + cy;
1540 /* Put this moved particle after the end of the grid,
1541 * so we can process it later without using conditionals.
1543 cell[i] = grid->ncx*grid->ncy;
1552 for (i = n0; i < n1; i++)
1554 cx = (int)((x[i][XX] - grid->c0[XX])*grid->inv_sx);
1555 cy = (int)((x[i][YY] - grid->c0[YY])*grid->inv_sy);
1557 /* For non-home zones there could be particles outside
1558 * the non-bonded cut-off range, which have been communicated
1559 * for bonded interactions only. For the result it doesn't
1560 * matter where these end up on the grid. For performance
1561 * we put them in an extra row at the border.
1564 cx = min(cx, grid->ncx - 1);
1566 cy = min(cy, grid->ncy - 1);
1568 /* For the moment cell will contain only the, grid local,
1569 * x and y indices, not z.
1571 cell[i] = cx*grid->ncy + cy;
1578 /* Determine in which grid cells the atoms should go */
1579 static void calc_cell_indices(const nbnxn_search_t nbs,
1586 nbnxn_atomdata_t *nbat)
1589 int cx, cy, cxy, ncz_max, ncz;
1590 int nthread, thread;
1591 int *cxy_na, cxy_na_i;
1593 nthread = gmx_omp_nthreads_get(emntPairsearch);
1595 #pragma omp parallel for num_threads(nthread) schedule(static)
1596 for (thread = 0; thread < nthread; thread++)
1598 calc_column_indices(grid, a0, a1, x, dd_zone, move, thread, nthread,
1599 nbs->cell, nbs->work[thread].cxy_na);
1602 /* Make the cell index as a function of x and y */
1605 grid->cxy_ind[0] = 0;
1606 for (i = 0; i < grid->ncx*grid->ncy+1; i++)
1608 /* We set ncz_max at the beginning of the loop iso at the end
1609 * to skip i=grid->ncx*grid->ncy which are moved particles
1610 * that do not need to be ordered on the grid.
1616 cxy_na_i = nbs->work[0].cxy_na[i];
1617 for (thread = 1; thread < nthread; thread++)
1619 cxy_na_i += nbs->work[thread].cxy_na[i];
1621 ncz = (cxy_na_i + grid->na_sc - 1)/grid->na_sc;
1622 if (nbat->XFormat == nbatX8)
1624 /* Make the number of cell a multiple of 2 */
1625 ncz = (ncz + 1) & ~1;
1627 grid->cxy_ind[i+1] = grid->cxy_ind[i] + ncz;
1628 /* Clear cxy_na, so we can reuse the array below */
1629 grid->cxy_na[i] = 0;
1631 grid->nc = grid->cxy_ind[grid->ncx*grid->ncy] - grid->cxy_ind[0];
1633 nbat->natoms = (grid->cell0 + grid->nc)*grid->na_sc;
1637 fprintf(debug, "ns na_sc %d na_c %d super-cells: %d x %d y %d z %.1f maxz %d\n",
1638 grid->na_sc, grid->na_c, grid->nc,
1639 grid->ncx, grid->ncy, grid->nc/((double)(grid->ncx*grid->ncy)),
1644 for (cy = 0; cy < grid->ncy; cy++)
1646 for (cx = 0; cx < grid->ncx; cx++)
1648 fprintf(debug, " %2d", grid->cxy_ind[i+1]-grid->cxy_ind[i]);
1651 fprintf(debug, "\n");
1656 /* Make sure the work array for sorting is large enough */
1657 if (ncz_max*grid->na_sc*SGSF > nbs->work[0].sort_work_nalloc)
1659 for (thread = 0; thread < nbs->nthread_max; thread++)
1661 nbs->work[thread].sort_work_nalloc =
1662 over_alloc_large(ncz_max*grid->na_sc*SGSF);
1663 srenew(nbs->work[thread].sort_work,
1664 nbs->work[thread].sort_work_nalloc);
1665 /* When not in use, all elements should be -1 */
1666 for (i = 0; i < nbs->work[thread].sort_work_nalloc; i++)
1668 nbs->work[thread].sort_work[i] = -1;
1673 /* Now we know the dimensions we can fill the grid.
1674 * This is the first, unsorted fill. We sort the columns after this.
1676 for (i = a0; i < a1; i++)
1678 /* At this point nbs->cell contains the local grid x,y indices */
1680 nbs->a[(grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc + grid->cxy_na[cxy]++] = i;
1685 /* Set the cell indices for the moved particles */
1686 n0 = grid->nc*grid->na_sc;
1687 n1 = grid->nc*grid->na_sc+grid->cxy_na[grid->ncx*grid->ncy];
1690 for (i = n0; i < n1; i++)
1692 nbs->cell[nbs->a[i]] = i;
1697 /* Sort the super-cell columns along z into the sub-cells. */
1698 #pragma omp parallel for num_threads(nbs->nthread_max) schedule(static)
1699 for (thread = 0; thread < nbs->nthread_max; thread++)
1703 sort_columns_simple(nbs, dd_zone, grid, a0, a1, atinfo, x, nbat,
1704 ((thread+0)*grid->ncx*grid->ncy)/nthread,
1705 ((thread+1)*grid->ncx*grid->ncy)/nthread,
1706 nbs->work[thread].sort_work);
1710 sort_columns_supersub(nbs, dd_zone, grid, a0, a1, atinfo, x, nbat,
1711 ((thread+0)*grid->ncx*grid->ncy)/nthread,
1712 ((thread+1)*grid->ncx*grid->ncy)/nthread,
1713 nbs->work[thread].sort_work);
1717 if (grid->bSimple && nbat->XFormat == nbatX8)
1719 combine_bounding_box_pairs(grid, grid->bb);
1724 grid->nsubc_tot = 0;
1725 for (i = 0; i < grid->nc; i++)
1727 grid->nsubc_tot += grid->nsubc[i];
1735 print_bbsizes_simple(debug, nbs, grid);
1739 fprintf(debug, "ns non-zero sub-cells: %d average atoms %.2f\n",
1740 grid->nsubc_tot, (a1-a0)/(double)grid->nsubc_tot);
1742 print_bbsizes_supersub(debug, nbs, grid);
1747 static void init_buffer_flags(nbnxn_buffer_flags_t *flags,
1752 flags->nflag = (natoms + NBNXN_BUFFERFLAG_SIZE - 1)/NBNXN_BUFFERFLAG_SIZE;
1753 if (flags->nflag > flags->flag_nalloc)
1755 flags->flag_nalloc = over_alloc_large(flags->nflag);
1756 srenew(flags->flag, flags->flag_nalloc);
1758 for (b = 0; b < flags->nflag; b++)
1764 /* Sets up a grid and puts the atoms on the grid.
1765 * This function only operates on one domain of the domain decompostion.
1766 * Note that without domain decomposition there is only one domain.
1768 void nbnxn_put_on_grid(nbnxn_search_t nbs,
1769 int ePBC, matrix box,
1771 rvec corner0, rvec corner1,
1776 int nmoved, int *move,
1778 nbnxn_atomdata_t *nbat)
1782 int nc_max_grid, nc_max;
1784 grid = &nbs->grid[dd_zone];
1786 nbs_cycle_start(&nbs->cc[enbsCCgrid]);
1788 grid->bSimple = nbnxn_kernel_pairlist_simple(nb_kernel_type);
1790 grid->na_c = nbnxn_kernel_to_ci_size(nb_kernel_type);
1791 grid->na_cj = nbnxn_kernel_to_cj_size(nb_kernel_type);
1792 grid->na_sc = (grid->bSimple ? 1 : GPU_NSUBCELL)*grid->na_c;
1793 grid->na_c_2log = get_2log(grid->na_c);
1795 nbat->na_c = grid->na_c;
1804 (nbs->grid[dd_zone-1].cell0 + nbs->grid[dd_zone-1].nc)*
1805 nbs->grid[dd_zone-1].na_sc/grid->na_sc;
1813 copy_mat(box, nbs->box);
1815 if (atom_density >= 0)
1817 grid->atom_density = atom_density;
1821 grid->atom_density = grid_atom_density(n-nmoved, corner0, corner1);
1826 nbs->natoms_local = a1 - nmoved;
1827 /* We assume that nbnxn_put_on_grid is called first
1828 * for the local atoms (dd_zone=0).
1830 nbs->natoms_nonlocal = a1 - nmoved;
1834 nbs->natoms_nonlocal = max(nbs->natoms_nonlocal, a1);
1837 nc_max_grid = set_grid_size_xy(nbs, grid,
1838 dd_zone, n-nmoved, corner0, corner1,
1839 nbs->grid[0].atom_density);
1841 nc_max = grid->cell0 + nc_max_grid;
1843 if (a1 > nbs->cell_nalloc)
1845 nbs->cell_nalloc = over_alloc_large(a1);
1846 srenew(nbs->cell, nbs->cell_nalloc);
1849 /* To avoid conditionals we store the moved particles at the end of a,
1850 * make sure we have enough space.
1852 if (nc_max*grid->na_sc + nmoved > nbs->a_nalloc)
1854 nbs->a_nalloc = over_alloc_large(nc_max*grid->na_sc + nmoved);
1855 srenew(nbs->a, nbs->a_nalloc);
1858 /* We need padding up to a multiple of the buffer flag size: simply add */
1859 if (nc_max*grid->na_sc + NBNXN_BUFFERFLAG_SIZE > nbat->nalloc)
1861 nbnxn_atomdata_realloc(nbat, nc_max*grid->na_sc+NBNXN_BUFFERFLAG_SIZE);
1864 calc_cell_indices(nbs, dd_zone, grid, a0, a1, atinfo, x, move, nbat);
1868 nbat->natoms_local = nbat->natoms;
1871 nbs_cycle_stop(&nbs->cc[enbsCCgrid]);
1874 /* Calls nbnxn_put_on_grid for all non-local domains */
1875 void nbnxn_put_on_grid_nonlocal(nbnxn_search_t nbs,
1876 const gmx_domdec_zones_t *zones,
1880 nbnxn_atomdata_t *nbat)
1885 for (zone = 1; zone < zones->n; zone++)
1887 for (d = 0; d < DIM; d++)
1889 c0[d] = zones->size[zone].bb_x0[d];
1890 c1[d] = zones->size[zone].bb_x1[d];
1893 nbnxn_put_on_grid(nbs, nbs->ePBC, NULL,
1895 zones->cg_range[zone],
1896 zones->cg_range[zone+1],
1906 /* Add simple grid type information to the local super/sub grid */
1907 void nbnxn_grid_add_simple(nbnxn_search_t nbs,
1908 nbnxn_atomdata_t *nbat)
1915 grid = &nbs->grid[0];
1919 gmx_incons("nbnxn_grid_simple called with a simple grid");
1922 ncd = grid->na_sc/NBNXN_CPU_CLUSTER_I_SIZE;
1924 if (grid->nc*ncd > grid->nc_nalloc_simple)
1926 grid->nc_nalloc_simple = over_alloc_large(grid->nc*ncd);
1927 srenew(grid->bbcz_simple, grid->nc_nalloc_simple*NNBSBB_D);
1928 srenew(grid->bb_simple, grid->nc_nalloc_simple);
1929 srenew(grid->flags_simple, grid->nc_nalloc_simple);
1932 sfree_aligned(grid->bbj);
1933 snew_aligned(grid->bbj, grid->nc_nalloc_simple/2, 16);
1937 bbcz = grid->bbcz_simple;
1938 bb = grid->bb_simple;
1940 #pragma omp parallel for num_threads(gmx_omp_nthreads_get(emntPairsearch)) schedule(static)
1941 for (sc = 0; sc < grid->nc; sc++)
1945 for (c = 0; c < ncd; c++)
1949 na = NBNXN_CPU_CLUSTER_I_SIZE;
1951 nbat->type[tx*NBNXN_CPU_CLUSTER_I_SIZE+na-1] == nbat->ntype-1)
1958 switch (nbat->XFormat)
1961 /* PACK_X4==NBNXN_CPU_CLUSTER_I_SIZE, so this is simple */
1962 calc_bounding_box_x_x4(na, nbat->x+tx*STRIDE_P4,
1966 /* PACK_X8>NBNXN_CPU_CLUSTER_I_SIZE, more complicated */
1967 calc_bounding_box_x_x8(na, nbat->x+X8_IND_A(tx*NBNXN_CPU_CLUSTER_I_SIZE),
1971 calc_bounding_box(na, nbat->xstride,
1972 nbat->x+tx*NBNXN_CPU_CLUSTER_I_SIZE*nbat->xstride,
1976 bbcz[tx*NNBSBB_D+0] = bb[tx].lower[BB_Z];
1977 bbcz[tx*NNBSBB_D+1] = bb[tx].upper[BB_Z];
1979 /* No interaction optimization yet here */
1980 grid->flags_simple[tx] = NBNXN_CI_DO_LJ(0) | NBNXN_CI_DO_COUL(0);
1984 grid->flags_simple[tx] = 0;
1989 if (grid->bSimple && nbat->XFormat == nbatX8)
1991 combine_bounding_box_pairs(grid, grid->bb_simple);
1995 void nbnxn_get_ncells(nbnxn_search_t nbs, int *ncx, int *ncy)
1997 *ncx = nbs->grid[0].ncx;
1998 *ncy = nbs->grid[0].ncy;
2001 void nbnxn_get_atomorder(nbnxn_search_t nbs, int **a, int *n)
2003 const nbnxn_grid_t *grid;
2005 grid = &nbs->grid[0];
2007 /* Return the atom order for the home cell (index 0) */
2010 *n = grid->cxy_ind[grid->ncx*grid->ncy]*grid->na_sc;
2013 void nbnxn_set_atomorder(nbnxn_search_t nbs)
2016 int ao, cx, cy, cxy, cz, j;
2018 /* Set the atom order for the home cell (index 0) */
2019 grid = &nbs->grid[0];
2022 for (cx = 0; cx < grid->ncx; cx++)
2024 for (cy = 0; cy < grid->ncy; cy++)
2026 cxy = cx*grid->ncy + cy;
2027 j = grid->cxy_ind[cxy]*grid->na_sc;
2028 for (cz = 0; cz < grid->cxy_na[cxy]; cz++)
2039 /* Determines the cell range along one dimension that
2040 * the bounding box b0 - b1 sees.
2042 static void get_cell_range(real b0, real b1,
2043 int nc, real c0, real s, real invs,
2044 real d2, real r2, int *cf, int *cl)
2046 *cf = max((int)((b0 - c0)*invs), 0);
2048 while (*cf > 0 && d2 + sqr((b0 - c0) - (*cf-1+1)*s) < r2)
2053 *cl = min((int)((b1 - c0)*invs), nc-1);
2054 while (*cl < nc-1 && d2 + sqr((*cl+1)*s - (b1 - c0)) < r2)
2060 /* Reference code calculating the distance^2 between two bounding boxes */
2061 static float box_dist2(float bx0, float bx1, float by0,
2062 float by1, float bz0, float bz1,
2063 const nbnxn_bb_t *bb)
2066 float dl, dh, dm, dm0;
2070 dl = bx0 - bb->upper[BB_X];
2071 dh = bb->lower[BB_X] - bx1;
2076 dl = by0 - bb->upper[BB_Y];
2077 dh = bb->lower[BB_Y] - by1;
2082 dl = bz0 - bb->upper[BB_Z];
2083 dh = bb->lower[BB_Z] - bz1;
2091 /* Plain C code calculating the distance^2 between two bounding boxes */
2092 static float subc_bb_dist2(int si, const nbnxn_bb_t *bb_i_ci,
2093 int csj, const nbnxn_bb_t *bb_j_all)
2095 const nbnxn_bb_t *bb_i, *bb_j;
2097 float dl, dh, dm, dm0;
2099 bb_i = bb_i_ci + si;
2100 bb_j = bb_j_all + csj;
2104 dl = bb_i->lower[BB_X] - bb_j->upper[BB_X];
2105 dh = bb_j->lower[BB_X] - bb_i->upper[BB_X];
2110 dl = bb_i->lower[BB_Y] - bb_j->upper[BB_Y];
2111 dh = bb_j->lower[BB_Y] - bb_i->upper[BB_Y];
2116 dl = bb_i->lower[BB_Z] - bb_j->upper[BB_Z];
2117 dh = bb_j->lower[BB_Z] - bb_i->upper[BB_Z];
2125 #ifdef NBNXN_SEARCH_BB_SIMD4
2127 /* 4-wide SIMD code for bb distance for bb format xyz0 */
2128 static float subc_bb_dist2_simd4(int si, const nbnxn_bb_t *bb_i_ci,
2129 int csj, const nbnxn_bb_t *bb_j_all)
2131 gmx_simd4_float_t bb_i_S0, bb_i_S1;
2132 gmx_simd4_float_t bb_j_S0, bb_j_S1;
2133 gmx_simd4_float_t dl_S;
2134 gmx_simd4_float_t dh_S;
2135 gmx_simd4_float_t dm_S;
2136 gmx_simd4_float_t dm0_S;
2138 bb_i_S0 = gmx_simd4_load_f(&bb_i_ci[si].lower[0]);
2139 bb_i_S1 = gmx_simd4_load_f(&bb_i_ci[si].upper[0]);
2140 bb_j_S0 = gmx_simd4_load_f(&bb_j_all[csj].lower[0]);
2141 bb_j_S1 = gmx_simd4_load_f(&bb_j_all[csj].upper[0]);
2143 dl_S = gmx_simd4_sub_f(bb_i_S0, bb_j_S1);
2144 dh_S = gmx_simd4_sub_f(bb_j_S0, bb_i_S1);
2146 dm_S = gmx_simd4_max_f(dl_S, dh_S);
2147 dm0_S = gmx_simd4_max_f(dm_S, gmx_simd4_setzero_f());
2149 return gmx_simd4_dotproduct3_f(dm0_S, dm0_S);
2152 /* Calculate bb bounding distances of bb_i[si,...,si+3] and store them in d2 */
2153 #define SUBC_BB_DIST2_SIMD4_XXXX_INNER(si, bb_i, d2) \
2157 gmx_simd4_float_t dx_0, dy_0, dz_0; \
2158 gmx_simd4_float_t dx_1, dy_1, dz_1; \
2160 gmx_simd4_float_t mx, my, mz; \
2161 gmx_simd4_float_t m0x, m0y, m0z; \
2163 gmx_simd4_float_t d2x, d2y, d2z; \
2164 gmx_simd4_float_t d2s, d2t; \
2166 shi = si*NNBSBB_D*DIM; \
2168 xi_l = gmx_simd4_load_f(bb_i+shi+0*STRIDE_PBB); \
2169 yi_l = gmx_simd4_load_f(bb_i+shi+1*STRIDE_PBB); \
2170 zi_l = gmx_simd4_load_f(bb_i+shi+2*STRIDE_PBB); \
2171 xi_h = gmx_simd4_load_f(bb_i+shi+3*STRIDE_PBB); \
2172 yi_h = gmx_simd4_load_f(bb_i+shi+4*STRIDE_PBB); \
2173 zi_h = gmx_simd4_load_f(bb_i+shi+5*STRIDE_PBB); \
2175 dx_0 = gmx_simd4_sub_f(xi_l, xj_h); \
2176 dy_0 = gmx_simd4_sub_f(yi_l, yj_h); \
2177 dz_0 = gmx_simd4_sub_f(zi_l, zj_h); \
2179 dx_1 = gmx_simd4_sub_f(xj_l, xi_h); \
2180 dy_1 = gmx_simd4_sub_f(yj_l, yi_h); \
2181 dz_1 = gmx_simd4_sub_f(zj_l, zi_h); \
2183 mx = gmx_simd4_max_f(dx_0, dx_1); \
2184 my = gmx_simd4_max_f(dy_0, dy_1); \
2185 mz = gmx_simd4_max_f(dz_0, dz_1); \
2187 m0x = gmx_simd4_max_f(mx, zero); \
2188 m0y = gmx_simd4_max_f(my, zero); \
2189 m0z = gmx_simd4_max_f(mz, zero); \
2191 d2x = gmx_simd4_mul_f(m0x, m0x); \
2192 d2y = gmx_simd4_mul_f(m0y, m0y); \
2193 d2z = gmx_simd4_mul_f(m0z, m0z); \
2195 d2s = gmx_simd4_add_f(d2x, d2y); \
2196 d2t = gmx_simd4_add_f(d2s, d2z); \
2198 gmx_simd4_store_f(d2+si, d2t); \
2201 /* 4-wide SIMD code for nsi bb distances for bb format xxxxyyyyzzzz */
2202 static void subc_bb_dist2_simd4_xxxx(const float *bb_j,
2203 int nsi, const float *bb_i,
2206 gmx_simd4_float_t xj_l, yj_l, zj_l;
2207 gmx_simd4_float_t xj_h, yj_h, zj_h;
2208 gmx_simd4_float_t xi_l, yi_l, zi_l;
2209 gmx_simd4_float_t xi_h, yi_h, zi_h;
2211 gmx_simd4_float_t zero;
2213 zero = gmx_simd4_setzero_f();
2215 xj_l = gmx_simd4_set1_f(bb_j[0*STRIDE_PBB]);
2216 yj_l = gmx_simd4_set1_f(bb_j[1*STRIDE_PBB]);
2217 zj_l = gmx_simd4_set1_f(bb_j[2*STRIDE_PBB]);
2218 xj_h = gmx_simd4_set1_f(bb_j[3*STRIDE_PBB]);
2219 yj_h = gmx_simd4_set1_f(bb_j[4*STRIDE_PBB]);
2220 zj_h = gmx_simd4_set1_f(bb_j[5*STRIDE_PBB]);
2222 /* Here we "loop" over si (0,STRIDE_PBB) from 0 to nsi with step STRIDE_PBB.
2223 * But as we know the number of iterations is 1 or 2, we unroll manually.
2225 SUBC_BB_DIST2_SIMD4_XXXX_INNER(0, bb_i, d2);
2226 if (STRIDE_PBB < nsi)
2228 SUBC_BB_DIST2_SIMD4_XXXX_INNER(STRIDE_PBB, bb_i, d2);
2232 #endif /* NBNXN_SEARCH_BB_SIMD4 */
2234 /* Plain C function which determines if any atom pair between two cells
2235 * is within distance sqrt(rl2).
2237 static gmx_bool subc_in_range_x(int na_c,
2238 int si, const real *x_i,
2239 int csj, int stride, const real *x_j,
2245 for (i = 0; i < na_c; i++)
2247 i0 = (si*na_c + i)*DIM;
2248 for (j = 0; j < na_c; j++)
2250 j0 = (csj*na_c + j)*stride;
2252 d2 = sqr(x_i[i0 ] - x_j[j0 ]) +
2253 sqr(x_i[i0+1] - x_j[j0+1]) +
2254 sqr(x_i[i0+2] - x_j[j0+2]);
2266 #ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
2268 /* 4-wide SIMD function which determines if any atom pair between two cells,
2269 * both with 8 atoms, is within distance sqrt(rl2).
2270 * Using 8-wide AVX is not faster on Intel Sandy Bridge.
2272 static gmx_bool subc_in_range_simd4(int na_c,
2273 int si, const real *x_i,
2274 int csj, int stride, const real *x_j,
2277 gmx_simd4_real_t ix_S0, iy_S0, iz_S0;
2278 gmx_simd4_real_t ix_S1, iy_S1, iz_S1;
2280 gmx_simd4_real_t rc2_S;
2285 rc2_S = gmx_simd4_set1_r(rl2);
2287 dim_stride = NBNXN_GPU_CLUSTER_SIZE/STRIDE_PBB*DIM;
2288 ix_S0 = gmx_simd4_load_r(x_i+(si*dim_stride+0)*STRIDE_PBB);
2289 iy_S0 = gmx_simd4_load_r(x_i+(si*dim_stride+1)*STRIDE_PBB);
2290 iz_S0 = gmx_simd4_load_r(x_i+(si*dim_stride+2)*STRIDE_PBB);
2291 ix_S1 = gmx_simd4_load_r(x_i+(si*dim_stride+3)*STRIDE_PBB);
2292 iy_S1 = gmx_simd4_load_r(x_i+(si*dim_stride+4)*STRIDE_PBB);
2293 iz_S1 = gmx_simd4_load_r(x_i+(si*dim_stride+5)*STRIDE_PBB);
2295 /* We loop from the outer to the inner particles to maximize
2296 * the chance that we find a pair in range quickly and return.
2302 gmx_simd4_real_t jx0_S, jy0_S, jz0_S;
2303 gmx_simd4_real_t jx1_S, jy1_S, jz1_S;
2305 gmx_simd4_real_t dx_S0, dy_S0, dz_S0;
2306 gmx_simd4_real_t dx_S1, dy_S1, dz_S1;
2307 gmx_simd4_real_t dx_S2, dy_S2, dz_S2;
2308 gmx_simd4_real_t dx_S3, dy_S3, dz_S3;
2310 gmx_simd4_real_t rsq_S0;
2311 gmx_simd4_real_t rsq_S1;
2312 gmx_simd4_real_t rsq_S2;
2313 gmx_simd4_real_t rsq_S3;
2315 gmx_simd4_bool_t wco_S0;
2316 gmx_simd4_bool_t wco_S1;
2317 gmx_simd4_bool_t wco_S2;
2318 gmx_simd4_bool_t wco_S3;
2319 gmx_simd4_bool_t wco_any_S01, wco_any_S23, wco_any_S;
2321 jx0_S = gmx_simd4_set1_r(x_j[j0*stride+0]);
2322 jy0_S = gmx_simd4_set1_r(x_j[j0*stride+1]);
2323 jz0_S = gmx_simd4_set1_r(x_j[j0*stride+2]);
2325 jx1_S = gmx_simd4_set1_r(x_j[j1*stride+0]);
2326 jy1_S = gmx_simd4_set1_r(x_j[j1*stride+1]);
2327 jz1_S = gmx_simd4_set1_r(x_j[j1*stride+2]);
2329 /* Calculate distance */
2330 dx_S0 = gmx_simd4_sub_r(ix_S0, jx0_S);
2331 dy_S0 = gmx_simd4_sub_r(iy_S0, jy0_S);
2332 dz_S0 = gmx_simd4_sub_r(iz_S0, jz0_S);
2333 dx_S1 = gmx_simd4_sub_r(ix_S1, jx0_S);
2334 dy_S1 = gmx_simd4_sub_r(iy_S1, jy0_S);
2335 dz_S1 = gmx_simd4_sub_r(iz_S1, jz0_S);
2336 dx_S2 = gmx_simd4_sub_r(ix_S0, jx1_S);
2337 dy_S2 = gmx_simd4_sub_r(iy_S0, jy1_S);
2338 dz_S2 = gmx_simd4_sub_r(iz_S0, jz1_S);
2339 dx_S3 = gmx_simd4_sub_r(ix_S1, jx1_S);
2340 dy_S3 = gmx_simd4_sub_r(iy_S1, jy1_S);
2341 dz_S3 = gmx_simd4_sub_r(iz_S1, jz1_S);
2343 /* rsq = dx*dx+dy*dy+dz*dz */
2344 rsq_S0 = gmx_simd4_calc_rsq_r(dx_S0, dy_S0, dz_S0);
2345 rsq_S1 = gmx_simd4_calc_rsq_r(dx_S1, dy_S1, dz_S1);
2346 rsq_S2 = gmx_simd4_calc_rsq_r(dx_S2, dy_S2, dz_S2);
2347 rsq_S3 = gmx_simd4_calc_rsq_r(dx_S3, dy_S3, dz_S3);
2349 wco_S0 = gmx_simd4_cmplt_r(rsq_S0, rc2_S);
2350 wco_S1 = gmx_simd4_cmplt_r(rsq_S1, rc2_S);
2351 wco_S2 = gmx_simd4_cmplt_r(rsq_S2, rc2_S);
2352 wco_S3 = gmx_simd4_cmplt_r(rsq_S3, rc2_S);
2354 wco_any_S01 = gmx_simd4_or_b(wco_S0, wco_S1);
2355 wco_any_S23 = gmx_simd4_or_b(wco_S2, wco_S3);
2356 wco_any_S = gmx_simd4_or_b(wco_any_S01, wco_any_S23);
2358 if (gmx_simd4_anytrue_b(wco_any_S))
2372 /* Returns the j sub-cell for index cj_ind */
2373 static int nbl_cj(const nbnxn_pairlist_t *nbl, int cj_ind)
2375 return nbl->cj4[cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG].cj[cj_ind & (NBNXN_GPU_JGROUP_SIZE - 1)];
2378 /* Returns the i-interaction mask of the j sub-cell for index cj_ind */
2379 static unsigned int nbl_imask0(const nbnxn_pairlist_t *nbl, int cj_ind)
2381 return nbl->cj4[cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG].imei[0].imask;
2384 /* Ensures there is enough space for extra extra exclusion masks */
2385 static void check_excl_space(nbnxn_pairlist_t *nbl, int extra)
2387 if (nbl->nexcl+extra > nbl->excl_nalloc)
2389 nbl->excl_nalloc = over_alloc_small(nbl->nexcl+extra);
2390 nbnxn_realloc_void((void **)&nbl->excl,
2391 nbl->nexcl*sizeof(*nbl->excl),
2392 nbl->excl_nalloc*sizeof(*nbl->excl),
2393 nbl->alloc, nbl->free);
2397 /* Ensures there is enough space for ncell extra j-cells in the list */
2398 static void check_subcell_list_space_simple(nbnxn_pairlist_t *nbl,
2403 cj_max = nbl->ncj + ncell;
2405 if (cj_max > nbl->cj_nalloc)
2407 nbl->cj_nalloc = over_alloc_small(cj_max);
2408 nbnxn_realloc_void((void **)&nbl->cj,
2409 nbl->ncj*sizeof(*nbl->cj),
2410 nbl->cj_nalloc*sizeof(*nbl->cj),
2411 nbl->alloc, nbl->free);
2415 /* Ensures there is enough space for ncell extra j-subcells in the list */
2416 static void check_subcell_list_space_supersub(nbnxn_pairlist_t *nbl,
2419 int ncj4_max, j4, j, w, t;
2422 #define WARP_SIZE 32
2424 /* We can have maximally nsupercell*GPU_NSUBCELL sj lists */
2425 /* We can store 4 j-subcell - i-supercell pairs in one struct.
2426 * since we round down, we need one extra entry.
2428 ncj4_max = ((nbl->work->cj_ind + nsupercell*GPU_NSUBCELL + NBNXN_GPU_JGROUP_SIZE - 1) >> NBNXN_GPU_JGROUP_SIZE_2LOG);
2430 if (ncj4_max > nbl->cj4_nalloc)
2432 nbl->cj4_nalloc = over_alloc_small(ncj4_max);
2433 nbnxn_realloc_void((void **)&nbl->cj4,
2434 nbl->work->cj4_init*sizeof(*nbl->cj4),
2435 nbl->cj4_nalloc*sizeof(*nbl->cj4),
2436 nbl->alloc, nbl->free);
2439 if (ncj4_max > nbl->work->cj4_init)
2441 for (j4 = nbl->work->cj4_init; j4 < ncj4_max; j4++)
2443 /* No i-subcells and no excl's in the list initially */
2444 for (w = 0; w < NWARP; w++)
2446 nbl->cj4[j4].imei[w].imask = 0U;
2447 nbl->cj4[j4].imei[w].excl_ind = 0;
2451 nbl->work->cj4_init = ncj4_max;
2455 /* Set all excl masks for one GPU warp no exclusions */
2456 static void set_no_excls(nbnxn_excl_t *excl)
2460 for (t = 0; t < WARP_SIZE; t++)
2462 /* Turn all interaction bits on */
2463 excl->pair[t] = NBNXN_INTERACTION_MASK_ALL;
2467 /* Initializes a single nbnxn_pairlist_t data structure */
2468 static void nbnxn_init_pairlist(nbnxn_pairlist_t *nbl,
2470 nbnxn_alloc_t *alloc,
2475 nbl->alloc = nbnxn_alloc_aligned;
2483 nbl->free = nbnxn_free_aligned;
2490 nbl->bSimple = bSimple;
2501 /* We need one element extra in sj, so alloc initially with 1 */
2502 nbl->cj4_nalloc = 0;
2509 nbl->excl_nalloc = 0;
2511 check_excl_space(nbl, 1);
2513 set_no_excls(&nbl->excl[0]);
2519 snew_aligned(nbl->work->bb_ci, 1, NBNXN_SEARCH_BB_MEM_ALIGN);
2524 snew_aligned(nbl->work->pbb_ci, GPU_NSUBCELL/STRIDE_PBB*NNBSBB_XXXX, NBNXN_SEARCH_BB_MEM_ALIGN);
2526 snew_aligned(nbl->work->bb_ci, GPU_NSUBCELL, NBNXN_SEARCH_BB_MEM_ALIGN);
2529 snew_aligned(nbl->work->x_ci, NBNXN_NA_SC_MAX*DIM, NBNXN_SEARCH_BB_MEM_ALIGN);
2530 #ifdef GMX_NBNXN_SIMD
2531 snew_aligned(nbl->work->x_ci_simd_4xn, 1, NBNXN_MEM_ALIGN);
2532 snew_aligned(nbl->work->x_ci_simd_2xnn, 1, NBNXN_MEM_ALIGN);
2534 snew_aligned(nbl->work->d2, GPU_NSUBCELL, NBNXN_SEARCH_BB_MEM_ALIGN);
2536 nbl->work->sort = NULL;
2537 nbl->work->sort_nalloc = 0;
2538 nbl->work->sci_sort = NULL;
2539 nbl->work->sci_sort_nalloc = 0;
2542 void nbnxn_init_pairlist_set(nbnxn_pairlist_set_t *nbl_list,
2543 gmx_bool bSimple, gmx_bool bCombined,
2544 nbnxn_alloc_t *alloc,
2549 nbl_list->bSimple = bSimple;
2550 nbl_list->bCombined = bCombined;
2552 nbl_list->nnbl = gmx_omp_nthreads_get(emntNonbonded);
2554 if (!nbl_list->bCombined &&
2555 nbl_list->nnbl > NBNXN_BUFFERFLAG_MAX_THREADS)
2557 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.",
2558 nbl_list->nnbl, NBNXN_BUFFERFLAG_MAX_THREADS, NBNXN_BUFFERFLAG_MAX_THREADS);
2561 snew(nbl_list->nbl, nbl_list->nnbl);
2562 snew(nbl_list->nbl_fep, nbl_list->nnbl);
2563 /* Execute in order to avoid memory interleaving between threads */
2564 #pragma omp parallel for num_threads(nbl_list->nnbl) schedule(static)
2565 for (i = 0; i < nbl_list->nnbl; i++)
2567 /* Allocate the nblist data structure locally on each thread
2568 * to optimize memory access for NUMA architectures.
2570 snew(nbl_list->nbl[i], 1);
2572 /* Only list 0 is used on the GPU, use normal allocation for i>0 */
2575 nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, alloc, free);
2579 nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, NULL, NULL);
2582 snew(nbl_list->nbl_fep[i], 1);
2583 nbnxn_init_pairlist_fep(nbl_list->nbl_fep[i]);
2587 /* Print statistics of a pair list, used for debug output */
2588 static void print_nblist_statistics_simple(FILE *fp, const nbnxn_pairlist_t *nbl,
2589 const nbnxn_search_t nbs, real rl)
2591 const nbnxn_grid_t *grid;
2596 /* This code only produces correct statistics with domain decomposition */
2597 grid = &nbs->grid[0];
2599 fprintf(fp, "nbl nci %d ncj %d\n",
2600 nbl->nci, nbl->ncj);
2601 fprintf(fp, "nbl na_sc %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
2602 nbl->na_sc, rl, nbl->ncj, nbl->ncj/(double)grid->nc,
2603 nbl->ncj/(double)grid->nc*grid->na_sc,
2604 nbl->ncj/(double)grid->nc*grid->na_sc/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid->nc*grid->na_sc/det(nbs->box)));
2606 fprintf(fp, "nbl average j cell list length %.1f\n",
2607 0.25*nbl->ncj/(double)nbl->nci);
2609 for (s = 0; s < SHIFTS; s++)
2614 for (i = 0; i < nbl->nci; i++)
2616 cs[nbl->ci[i].shift & NBNXN_CI_SHIFT] +=
2617 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start;
2619 j = nbl->ci[i].cj_ind_start;
2620 while (j < nbl->ci[i].cj_ind_end &&
2621 nbl->cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
2627 fprintf(fp, "nbl cell pairs, total: %d excl: %d %.1f%%\n",
2628 nbl->ncj, npexcl, 100*npexcl/(double)nbl->ncj);
2629 for (s = 0; s < SHIFTS; s++)
2633 fprintf(fp, "nbl shift %2d ncj %3d\n", s, cs[s]);
2638 /* Print statistics of a pair lists, used for debug output */
2639 static void print_nblist_statistics_supersub(FILE *fp, const nbnxn_pairlist_t *nbl,
2640 const nbnxn_search_t nbs, real rl)
2642 const nbnxn_grid_t *grid;
2643 int i, j4, j, si, b;
2644 int c[GPU_NSUBCELL+1];
2646 /* This code only produces correct statistics with domain decomposition */
2647 grid = &nbs->grid[0];
2649 fprintf(fp, "nbl nsci %d ncj4 %d nsi %d excl4 %d\n",
2650 nbl->nsci, nbl->ncj4, nbl->nci_tot, nbl->nexcl);
2651 fprintf(fp, "nbl na_c %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
2652 nbl->na_ci, rl, nbl->nci_tot, nbl->nci_tot/(double)grid->nsubc_tot,
2653 nbl->nci_tot/(double)grid->nsubc_tot*grid->na_c,
2654 nbl->nci_tot/(double)grid->nsubc_tot*grid->na_c/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid->nsubc_tot*grid->na_c/det(nbs->box)));
2656 fprintf(fp, "nbl average j super cell list length %.1f\n",
2657 0.25*nbl->ncj4/(double)nbl->nsci);
2658 fprintf(fp, "nbl average i sub cell list length %.1f\n",
2659 nbl->nci_tot/((double)nbl->ncj4));
2661 for (si = 0; si <= GPU_NSUBCELL; si++)
2665 for (i = 0; i < nbl->nsci; i++)
2667 for (j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
2669 for (j = 0; j < NBNXN_GPU_JGROUP_SIZE; j++)
2672 for (si = 0; si < GPU_NSUBCELL; si++)
2674 if (nbl->cj4[j4].imei[0].imask & (1U << (j*GPU_NSUBCELL + si)))
2683 for (b = 0; b <= GPU_NSUBCELL; b++)
2685 fprintf(fp, "nbl j-list #i-subcell %d %7d %4.1f\n",
2686 b, c[b], 100.0*c[b]/(double)(nbl->ncj4*NBNXN_GPU_JGROUP_SIZE));
2690 /* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp */
2691 static void low_get_nbl_exclusions(nbnxn_pairlist_t *nbl, int cj4,
2692 int warp, nbnxn_excl_t **excl)
2694 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
2696 /* No exclusions set, make a new list entry */
2697 nbl->cj4[cj4].imei[warp].excl_ind = nbl->nexcl;
2699 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
2700 set_no_excls(*excl);
2704 /* We already have some exclusions, new ones can be added to the list */
2705 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
2709 /* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp,
2710 * generates a new element and allocates extra memory, if necessary.
2712 static void get_nbl_exclusions_1(nbnxn_pairlist_t *nbl, int cj4,
2713 int warp, nbnxn_excl_t **excl)
2715 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
2717 /* We need to make a new list entry, check if we have space */
2718 check_excl_space(nbl, 1);
2720 low_get_nbl_exclusions(nbl, cj4, warp, excl);
2723 /* Returns pointers to the exclusion mask for cj4-unit cj4 for both warps,
2724 * generates a new element and allocates extra memory, if necessary.
2726 static void get_nbl_exclusions_2(nbnxn_pairlist_t *nbl, int cj4,
2727 nbnxn_excl_t **excl_w0,
2728 nbnxn_excl_t **excl_w1)
2730 /* Check for space we might need */
2731 check_excl_space(nbl, 2);
2733 low_get_nbl_exclusions(nbl, cj4, 0, excl_w0);
2734 low_get_nbl_exclusions(nbl, cj4, 1, excl_w1);
2737 /* Sets the self exclusions i=j and pair exclusions i>j */
2738 static void set_self_and_newton_excls_supersub(nbnxn_pairlist_t *nbl,
2739 int cj4_ind, int sj_offset,
2742 nbnxn_excl_t *excl[2];
2745 /* Here we only set the set self and double pair exclusions */
2747 get_nbl_exclusions_2(nbl, cj4_ind, &excl[0], &excl[1]);
2749 /* Only minor < major bits set */
2750 for (ej = 0; ej < nbl->na_ci; ej++)
2753 for (ei = ej; ei < nbl->na_ci; ei++)
2755 excl[w]->pair[(ej & (NBNXN_GPU_JGROUP_SIZE-1))*nbl->na_ci + ei] &=
2756 ~(1U << (sj_offset*GPU_NSUBCELL + si));
2761 /* Returns a diagonal or off-diagonal interaction mask for plain C lists */
2762 static unsigned int get_imask(gmx_bool rdiag, int ci, int cj)
2764 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
2767 /* Returns a diagonal or off-diagonal interaction mask for cj-size=2 */
2768 static unsigned int get_imask_simd_j2(gmx_bool rdiag, int ci, int cj)
2770 return (rdiag && ci*2 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_0 :
2771 (rdiag && ci*2+1 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_1 :
2772 NBNXN_INTERACTION_MASK_ALL));
2775 /* Returns a diagonal or off-diagonal interaction mask for cj-size=4 */
2776 static unsigned int get_imask_simd_j4(gmx_bool rdiag, int ci, int cj)
2778 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
2781 /* Returns a diagonal or off-diagonal interaction mask for cj-size=8 */
2782 static unsigned int get_imask_simd_j8(gmx_bool rdiag, int ci, int cj)
2784 return (rdiag && ci == cj*2 ? NBNXN_INTERACTION_MASK_DIAG_J8_0 :
2785 (rdiag && ci == cj*2+1 ? NBNXN_INTERACTION_MASK_DIAG_J8_1 :
2786 NBNXN_INTERACTION_MASK_ALL));
2789 #ifdef GMX_NBNXN_SIMD
2790 #if GMX_SIMD_REAL_WIDTH == 2
2791 #define get_imask_simd_4xn get_imask_simd_j2
2793 #if GMX_SIMD_REAL_WIDTH == 4
2794 #define get_imask_simd_4xn get_imask_simd_j4
2796 #if GMX_SIMD_REAL_WIDTH == 8
2797 #define get_imask_simd_4xn get_imask_simd_j8
2798 #define get_imask_simd_2xnn get_imask_simd_j4
2800 #if GMX_SIMD_REAL_WIDTH == 16
2801 #define get_imask_simd_2xnn get_imask_simd_j8
2805 /* Plain C code for making a pair list of cell ci vs cell cjf-cjl.
2806 * Checks bounding box distances and possibly atom pair distances.
2808 static void make_cluster_list_simple(const nbnxn_grid_t *gridj,
2809 nbnxn_pairlist_t *nbl,
2810 int ci, int cjf, int cjl,
2811 gmx_bool remove_sub_diag,
2813 real rl2, float rbb2,
2816 const nbnxn_list_work_t *work;
2818 const nbnxn_bb_t *bb_ci;
2823 int cjf_gl, cjl_gl, cj;
2827 bb_ci = nbl->work->bb_ci;
2828 x_ci = nbl->work->x_ci;
2831 while (!InRange && cjf <= cjl)
2833 d2 = subc_bb_dist2(0, bb_ci, cjf, gridj->bb);
2836 /* Check if the distance is within the distance where
2837 * we use only the bounding box distance rbb,
2838 * or within the cut-off and there is at least one atom pair
2839 * within the cut-off.
2849 cjf_gl = gridj->cell0 + cjf;
2850 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
2852 for (j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE; j++)
2854 InRange = InRange ||
2855 (sqr(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
2856 sqr(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
2857 sqr(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rl2);
2860 *ndistc += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
2873 while (!InRange && cjl > cjf)
2875 d2 = subc_bb_dist2(0, bb_ci, cjl, gridj->bb);
2878 /* Check if the distance is within the distance where
2879 * we use only the bounding box distance rbb,
2880 * or within the cut-off and there is at least one atom pair
2881 * within the cut-off.
2891 cjl_gl = gridj->cell0 + cjl;
2892 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
2894 for (j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE; j++)
2896 InRange = InRange ||
2897 (sqr(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
2898 sqr(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
2899 sqr(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rl2);
2902 *ndistc += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
2912 for (cj = cjf; cj <= cjl; cj++)
2914 /* Store cj and the interaction mask */
2915 nbl->cj[nbl->ncj].cj = gridj->cell0 + cj;
2916 nbl->cj[nbl->ncj].excl = get_imask(remove_sub_diag, ci, cj);
2919 /* Increase the closing index in i super-cell list */
2920 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
2924 #ifdef GMX_NBNXN_SIMD_4XN
2925 #include "nbnxn_search_simd_4xn.h"
2927 #ifdef GMX_NBNXN_SIMD_2XNN
2928 #include "nbnxn_search_simd_2xnn.h"
2931 /* Plain C or SIMD4 code for making a pair list of super-cell sci vs scj.
2932 * Checks bounding box distances and possibly atom pair distances.
2934 static void make_cluster_list_supersub(const nbnxn_grid_t *gridi,
2935 const nbnxn_grid_t *gridj,
2936 nbnxn_pairlist_t *nbl,
2938 gmx_bool sci_equals_scj,
2939 int stride, const real *x,
2940 real rl2, float rbb2,
2945 int cjo, ci1, ci, cj, cj_gl;
2946 int cj4_ind, cj_offset;
2950 const float *pbb_ci;
2952 const nbnxn_bb_t *bb_ci;
2957 #define PRUNE_LIST_CPU_ONE
2958 #ifdef PRUNE_LIST_CPU_ONE
2962 d2l = nbl->work->d2;
2965 pbb_ci = nbl->work->pbb_ci;
2967 bb_ci = nbl->work->bb_ci;
2969 x_ci = nbl->work->x_ci;
2973 for (cjo = 0; cjo < gridj->nsubc[scj]; cjo++)
2975 cj4_ind = (nbl->work->cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG);
2976 cj_offset = nbl->work->cj_ind - cj4_ind*NBNXN_GPU_JGROUP_SIZE;
2977 cj4 = &nbl->cj4[cj4_ind];
2979 cj = scj*GPU_NSUBCELL + cjo;
2981 cj_gl = gridj->cell0*GPU_NSUBCELL + cj;
2983 /* Initialize this j-subcell i-subcell list */
2984 cj4->cj[cj_offset] = cj_gl;
2993 ci1 = gridi->nsubc[sci];
2997 /* Determine all ci1 bb distances in one call with SIMD4 */
2998 subc_bb_dist2_simd4_xxxx(gridj->pbb+(cj>>STRIDE_PBB_2LOG)*NNBSBB_XXXX+(cj & (STRIDE_PBB-1)),
3004 /* We use a fixed upper-bound instead of ci1 to help optimization */
3005 for (ci = 0; ci < GPU_NSUBCELL; ci++)
3012 #ifndef NBNXN_BBXXXX
3013 /* Determine the bb distance between ci and cj */
3014 d2l[ci] = subc_bb_dist2(ci, bb_ci, cj, gridj->bb);
3019 #ifdef PRUNE_LIST_CPU_ALL
3020 /* Check if the distance is within the distance where
3021 * we use only the bounding box distance rbb,
3022 * or within the cut-off and there is at least one atom pair
3023 * within the cut-off. This check is very costly.
3025 *ndistc += na_c*na_c;
3028 #ifdef NBNXN_PBB_SIMD4
3033 (na_c, ci, x_ci, cj_gl, stride, x, rl2)))
3035 /* Check if the distance between the two bounding boxes
3036 * in within the pair-list cut-off.
3041 /* Flag this i-subcell to be taken into account */
3042 imask |= (1U << (cj_offset*GPU_NSUBCELL+ci));
3044 #ifdef PRUNE_LIST_CPU_ONE
3052 #ifdef PRUNE_LIST_CPU_ONE
3053 /* If we only found 1 pair, check if any atoms are actually
3054 * within the cut-off, so we could get rid of it.
3056 if (npair == 1 && d2l[ci_last] >= rbb2)
3058 /* Avoid using function pointers here, as it's slower */
3060 #ifdef NBNXN_PBB_SIMD4
3061 !subc_in_range_simd4
3065 (na_c, ci_last, x_ci, cj_gl, stride, x, rl2))
3067 imask &= ~(1U << (cj_offset*GPU_NSUBCELL+ci_last));
3075 /* We have a useful sj entry, close it now */
3077 /* Set the exclucions for the ci== sj entry.
3078 * Here we don't bother to check if this entry is actually flagged,
3079 * as it will nearly always be in the list.
3083 set_self_and_newton_excls_supersub(nbl, cj4_ind, cj_offset, cjo);
3086 /* Copy the cluster interaction mask to the list */
3087 for (w = 0; w < NWARP; w++)
3089 cj4->imei[w].imask |= imask;
3092 nbl->work->cj_ind++;
3094 /* Keep the count */
3095 nbl->nci_tot += npair;
3097 /* Increase the closing index in i super-cell list */
3098 nbl->sci[nbl->nsci].cj4_ind_end =
3099 ((nbl->work->cj_ind+NBNXN_GPU_JGROUP_SIZE-1) >> NBNXN_GPU_JGROUP_SIZE_2LOG);
3104 /* Set all atom-pair exclusions from the topology stored in excl
3105 * as masks in the pair-list for simple list i-entry nbl_ci
3107 static void set_ci_top_excls(const nbnxn_search_t nbs,
3108 nbnxn_pairlist_t *nbl,
3109 gmx_bool diagRemoved,
3112 const nbnxn_ci_t *nbl_ci,
3113 const t_blocka *excl)
3117 int cj_ind_first, cj_ind_last;
3118 int cj_first, cj_last;
3120 int i, ai, aj, si, eind, ge, se;
3121 int found, cj_ind_0, cj_ind_1, cj_ind_m;
3125 nbnxn_excl_t *nbl_excl;
3126 int inner_i, inner_e;
3130 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
3138 cj_ind_first = nbl_ci->cj_ind_start;
3139 cj_ind_last = nbl->ncj - 1;
3141 cj_first = nbl->cj[cj_ind_first].cj;
3142 cj_last = nbl->cj[cj_ind_last].cj;
3144 /* Determine how many contiguous j-cells we have starting
3145 * from the first i-cell. This number can be used to directly
3146 * calculate j-cell indices for excluded atoms.
3149 if (na_ci_2log == na_cj_2log)
3151 while (cj_ind_first + ndirect <= cj_ind_last &&
3152 nbl->cj[cj_ind_first+ndirect].cj == ci + ndirect)
3157 #ifdef NBNXN_SEARCH_BB_SIMD4
3160 while (cj_ind_first + ndirect <= cj_ind_last &&
3161 nbl->cj[cj_ind_first+ndirect].cj == ci_to_cj(na_cj_2log, ci) + ndirect)
3168 /* Loop over the atoms in the i super-cell */
3169 for (i = 0; i < nbl->na_sc; i++)
3171 ai = nbs->a[ci*nbl->na_sc+i];
3174 si = (i>>na_ci_2log);
3176 /* Loop over the topology-based exclusions for this i-atom */
3177 for (eind = excl->index[ai]; eind < excl->index[ai+1]; eind++)
3183 /* The self exclusion are already set, save some time */
3189 /* Without shifts we only calculate interactions j>i
3190 * for one-way pair-lists.
3192 if (diagRemoved && ge <= ci*nbl->na_sc + i)
3197 se = (ge >> na_cj_2log);
3199 /* Could the cluster se be in our list? */
3200 if (se >= cj_first && se <= cj_last)
3202 if (se < cj_first + ndirect)
3204 /* We can calculate cj_ind directly from se */
3205 found = cj_ind_first + se - cj_first;
3209 /* Search for se using bisection */
3211 cj_ind_0 = cj_ind_first + ndirect;
3212 cj_ind_1 = cj_ind_last + 1;
3213 while (found == -1 && cj_ind_0 < cj_ind_1)
3215 cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
3217 cj_m = nbl->cj[cj_ind_m].cj;
3225 cj_ind_1 = cj_ind_m;
3229 cj_ind_0 = cj_ind_m + 1;
3236 inner_i = i - (si << na_ci_2log);
3237 inner_e = ge - (se << na_cj_2log);
3239 nbl->cj[found].excl &= ~(1U<<((inner_i<<na_cj_2log) + inner_e));
3240 /* The next code line is usually not needed. We do not want to version
3241 * away the above line, because there is logic that relies on being
3242 * able to detect easily whether any exclusions exist. */
3243 #if (defined GMX_SIMD_IBM_QPX)
3244 nbl->cj[found].interaction_mask_indices[inner_i] &= ~(1U << inner_e);
3253 /* Add a new i-entry to the FEP list and copy the i-properties */
3254 static gmx_inline void fep_list_new_nri_copy(t_nblist *nlist)
3256 /* Add a new i-entry */
3259 assert(nlist->nri < nlist->maxnri);
3261 /* Duplicate the last i-entry, except for jindex, which continues */
3262 nlist->iinr[nlist->nri] = nlist->iinr[nlist->nri-1];
3263 nlist->shift[nlist->nri] = nlist->shift[nlist->nri-1];
3264 nlist->gid[nlist->nri] = nlist->gid[nlist->nri-1];
3265 nlist->jindex[nlist->nri] = nlist->nrj;
3268 /* For load balancing of the free-energy lists over threads, we set
3269 * the maximum nrj size of an i-entry to 40. This leads to good
3270 * load balancing in the worst case scenario of a single perturbed
3271 * particle on 16 threads, while not introducing significant overhead.
3272 * Note that half of the perturbed pairs will anyhow end up in very small lists,
3273 * since non perturbed i-particles will see few perturbed j-particles).
3275 const int max_nrj_fep = 40;
3277 /* Exclude the perturbed pairs from the Verlet list. This is only done to avoid
3278 * singularities for overlapping particles (0/0), since the charges and
3279 * LJ parameters have been zeroed in the nbnxn data structure.
3280 * Simultaneously make a group pair list for the perturbed pairs.
3282 static void make_fep_list(const nbnxn_search_t nbs,
3283 const nbnxn_atomdata_t *nbat,
3284 nbnxn_pairlist_t *nbl,
3285 gmx_bool bDiagRemoved,
3287 const nbnxn_grid_t *gridi,
3288 const nbnxn_grid_t *gridj,
3291 int ci, cj_ind_start, cj_ind_end, cj_ind, cja, cjr;
3293 int ngid, gid_i = 0, gid_j, gid;
3294 int egp_shift, egp_mask;
3296 int i, j, ind_i, ind_j, ai, aj;
3298 gmx_bool bFEP_i, bFEP_i_all;
3300 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
3308 cj_ind_start = nbl_ci->cj_ind_start;
3309 cj_ind_end = nbl_ci->cj_ind_end;
3311 /* In worst case we have alternating energy groups
3312 * and create #atom-pair lists, which means we need the size
3313 * of a cluster pair (na_ci*na_cj) times the number of cj's.
3315 nri_max = nbl->na_ci*nbl->na_cj*(cj_ind_end - cj_ind_start);
3316 if (nlist->nri + nri_max > nlist->maxnri)
3318 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
3319 reallocate_nblist(nlist);
3322 ngid = nbat->nenergrp;
3324 if (ngid*gridj->na_cj > sizeof(gid_cj)*8)
3326 gmx_fatal(FARGS, "The Verlet scheme with %dx%d kernels and free-energy only supports up to %d energy groups",
3327 gridi->na_c, gridj->na_cj, (sizeof(gid_cj)*8)/gridj->na_cj);
3330 egp_shift = nbat->neg_2log;
3331 egp_mask = (1<<nbat->neg_2log) - 1;
3333 /* Loop over the atoms in the i sub-cell */
3335 for (i = 0; i < nbl->na_ci; i++)
3337 ind_i = ci*nbl->na_ci + i;
3342 nlist->jindex[nri+1] = nlist->jindex[nri];
3343 nlist->iinr[nri] = ai;
3344 /* The actual energy group pair index is set later */
3345 nlist->gid[nri] = 0;
3346 nlist->shift[nri] = nbl_ci->shift & NBNXN_CI_SHIFT;
3348 bFEP_i = gridi->fep[ci - gridi->cell0] & (1 << i);
3350 bFEP_i_all = bFEP_i_all && bFEP_i;
3352 if ((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj > nlist->maxnrj)
3354 nlist->maxnrj = over_alloc_small((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj);
3355 srenew(nlist->jjnr, nlist->maxnrj);
3356 srenew(nlist->excl_fep, nlist->maxnrj);
3361 gid_i = (nbat->energrp[ci] >> (egp_shift*i)) & egp_mask;
3364 for (cj_ind = cj_ind_start; cj_ind < cj_ind_end; cj_ind++)
3366 unsigned int fep_cj;
3368 cja = nbl->cj[cj_ind].cj;
3370 if (gridj->na_cj == gridj->na_c)
3372 cjr = cja - gridj->cell0;
3373 fep_cj = gridj->fep[cjr];
3376 gid_cj = nbat->energrp[cja];
3379 else if (2*gridj->na_cj == gridj->na_c)
3381 cjr = cja - gridj->cell0*2;
3382 /* Extract half of the ci fep/energrp mask */
3383 fep_cj = (gridj->fep[cjr>>1] >> ((cjr&1)*gridj->na_cj)) & ((1<<gridj->na_cj) - 1);
3386 gid_cj = nbat->energrp[cja>>1] >> ((cja&1)*gridj->na_cj*egp_shift) & ((1<<(gridj->na_cj*egp_shift)) - 1);
3391 cjr = cja - (gridj->cell0>>1);
3392 /* Combine two ci fep masks/energrp */
3393 fep_cj = gridj->fep[cjr*2] + (gridj->fep[cjr*2+1] << gridj->na_c);
3396 gid_cj = nbat->energrp[cja*2] + (nbat->energrp[cja*2+1] << (gridj->na_c*egp_shift));
3400 if (bFEP_i || fep_cj != 0)
3402 for (j = 0; j < nbl->na_cj; j++)
3404 /* Is this interaction perturbed and not excluded? */
3405 ind_j = cja*nbl->na_cj + j;
3408 (bFEP_i || (fep_cj & (1 << j))) &&
3409 (!bDiagRemoved || ind_j >= ind_i))
3413 gid_j = (gid_cj >> (j*egp_shift)) & egp_mask;
3414 gid = GID(gid_i, gid_j, ngid);
3416 if (nlist->nrj > nlist->jindex[nri] &&
3417 nlist->gid[nri] != gid)
3419 /* Energy group pair changed: new list */
3420 fep_list_new_nri_copy(nlist);
3423 nlist->gid[nri] = gid;
3426 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
3428 fep_list_new_nri_copy(nlist);
3432 /* Add it to the FEP list */
3433 nlist->jjnr[nlist->nrj] = aj;
3434 nlist->excl_fep[nlist->nrj] = (nbl->cj[cj_ind].excl >> (i*nbl->na_cj + j)) & 1;
3437 /* Exclude it from the normal list.
3438 * Note that the charge has been set to zero,
3439 * but we need to avoid 0/0, as perturbed atoms
3440 * can be on top of each other.
3441 * (and the LJ parameters have not been zeroed)
3443 nbl->cj[cj_ind].excl &= ~(1U << (i*nbl->na_cj + j));
3449 if (nlist->nrj > nlist->jindex[nri])
3451 /* Actually add this new, non-empty, list */
3453 nlist->jindex[nlist->nri] = nlist->nrj;
3460 /* All interactions are perturbed, we can skip this entry */
3461 nbl_ci->cj_ind_end = cj_ind_start;
3465 /* Return the index of atom a within a cluster */
3466 static gmx_inline int cj_mod_cj4(int cj)
3468 return cj & (NBNXN_GPU_JGROUP_SIZE - 1);
3471 /* Convert a j-cluster to a cj4 group */
3472 static gmx_inline int cj_to_cj4(int cj)
3474 return cj >> NBNXN_GPU_JGROUP_SIZE_2LOG;
3477 /* Return the index of an j-atom within a warp */
3478 static gmx_inline int a_mod_wj(int a)
3480 return a & (NBNXN_GPU_CLUSTER_SIZE/2 - 1);
3483 /* As make_fep_list above, but for super/sub lists. */
3484 static void make_fep_list_supersub(const nbnxn_search_t nbs,
3485 const nbnxn_atomdata_t *nbat,
3486 nbnxn_pairlist_t *nbl,
3487 gmx_bool bDiagRemoved,
3488 const nbnxn_sci_t *nbl_sci,
3493 const nbnxn_grid_t *gridi,
3494 const nbnxn_grid_t *gridj,
3497 int sci, cj4_ind_start, cj4_ind_end, cj4_ind, gcj, cjr;
3500 int i, j, ind_i, ind_j, ai, aj;
3504 const nbnxn_cj4_t *cj4;
3506 if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
3514 cj4_ind_start = nbl_sci->cj4_ind_start;
3515 cj4_ind_end = nbl_sci->cj4_ind_end;
3517 /* Here we process one super-cell, max #atoms na_sc, versus a list
3518 * cj4 entries, each with max NBNXN_GPU_JGROUP_SIZE cj's, each
3519 * of size na_cj atoms.
3520 * On the GPU we don't support energy groups (yet).
3521 * So for each of the na_sc i-atoms, we need max one FEP list
3522 * for each max_nrj_fep j-atoms.
3524 nri_max = nbl->na_sc*nbl->na_cj*(1 + ((cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE)/max_nrj_fep);
3525 if (nlist->nri + nri_max > nlist->maxnri)
3527 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
3528 reallocate_nblist(nlist);
3531 /* Loop over the atoms in the i super-cluster */
3532 for (c = 0; c < GPU_NSUBCELL; c++)
3534 c_abs = sci*GPU_NSUBCELL + c;
3536 for (i = 0; i < nbl->na_ci; i++)
3538 ind_i = c_abs*nbl->na_ci + i;
3543 nlist->jindex[nri+1] = nlist->jindex[nri];
3544 nlist->iinr[nri] = ai;
3545 /* With GPUs, energy groups are not supported */
3546 nlist->gid[nri] = 0;
3547 nlist->shift[nri] = nbl_sci->shift & NBNXN_CI_SHIFT;
3549 bFEP_i = (gridi->fep[c_abs - gridi->cell0] & (1 << i));
3551 xi = nbat->x[ind_i*nbat->xstride+XX] + shx;
3552 yi = nbat->x[ind_i*nbat->xstride+YY] + shy;
3553 zi = nbat->x[ind_i*nbat->xstride+ZZ] + shz;
3555 if ((nlist->nrj + cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE*nbl->na_cj > nlist->maxnrj)
3557 nlist->maxnrj = over_alloc_small((nlist->nrj + cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE*nbl->na_cj);
3558 srenew(nlist->jjnr, nlist->maxnrj);
3559 srenew(nlist->excl_fep, nlist->maxnrj);
3562 for (cj4_ind = cj4_ind_start; cj4_ind < cj4_ind_end; cj4_ind++)
3564 cj4 = &nbl->cj4[cj4_ind];
3566 for (gcj = 0; gcj < NBNXN_GPU_JGROUP_SIZE; gcj++)
3568 unsigned int fep_cj;
3570 if ((cj4->imei[0].imask & (1U << (gcj*GPU_NSUBCELL + c))) == 0)
3572 /* Skip this ci for this cj */
3576 cjr = cj4->cj[gcj] - gridj->cell0*GPU_NSUBCELL;
3578 fep_cj = gridj->fep[cjr];
3580 if (bFEP_i || fep_cj != 0)
3582 for (j = 0; j < nbl->na_cj; j++)
3584 /* Is this interaction perturbed and not excluded? */
3585 ind_j = (gridj->cell0*GPU_NSUBCELL + cjr)*nbl->na_cj + j;
3588 (bFEP_i || (fep_cj & (1 << j))) &&
3589 (!bDiagRemoved || ind_j >= ind_i))
3593 unsigned int excl_bit;
3596 get_nbl_exclusions_1(nbl, cj4_ind, j>>2, &excl);
3598 excl_pair = a_mod_wj(j)*nbl->na_ci + i;
3599 excl_bit = (1U << (gcj*GPU_NSUBCELL + c));
3601 dx = nbat->x[ind_j*nbat->xstride+XX] - xi;
3602 dy = nbat->x[ind_j*nbat->xstride+YY] - yi;
3603 dz = nbat->x[ind_j*nbat->xstride+ZZ] - zi;
3605 /* The unpruned GPU list has more than 2/3
3606 * of the atom pairs beyond rlist. Using
3607 * this list will cause a lot of overhead
3608 * in the CPU FEP kernels, especially
3609 * relative to the fast GPU kernels.
3610 * So we prune the FEP list here.
3612 if (dx*dx + dy*dy + dz*dz < rlist_fep2)
3614 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
3616 fep_list_new_nri_copy(nlist);
3620 /* Add it to the FEP list */
3621 nlist->jjnr[nlist->nrj] = aj;
3622 nlist->excl_fep[nlist->nrj] = (excl->pair[excl_pair] & excl_bit) ? 1 : 0;
3626 /* Exclude it from the normal list.
3627 * Note that the charge and LJ parameters have
3628 * been set to zero, but we need to avoid 0/0,
3629 * as perturbed atoms can be on top of each other.
3631 excl->pair[excl_pair] &= ~excl_bit;
3635 /* Note that we could mask out this pair in imask
3636 * if all i- and/or all j-particles are perturbed.
3637 * But since the perturbed pairs on the CPU will
3638 * take an order of magnitude more time, the GPU
3639 * will finish before the CPU and there is no gain.
3645 if (nlist->nrj > nlist->jindex[nri])
3647 /* Actually add this new, non-empty, list */
3649 nlist->jindex[nlist->nri] = nlist->nrj;
3656 /* Set all atom-pair exclusions from the topology stored in excl
3657 * as masks in the pair-list for i-super-cell entry nbl_sci
3659 static void set_sci_top_excls(const nbnxn_search_t nbs,
3660 nbnxn_pairlist_t *nbl,
3661 gmx_bool diagRemoved,
3663 const nbnxn_sci_t *nbl_sci,
3664 const t_blocka *excl)
3669 int cj_ind_first, cj_ind_last;
3670 int cj_first, cj_last;
3672 int i, ai, aj, si, eind, ge, se;
3673 int found, cj_ind_0, cj_ind_1, cj_ind_m;
3677 nbnxn_excl_t *nbl_excl;
3678 int inner_i, inner_e, w;
3684 if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
3692 cj_ind_first = nbl_sci->cj4_ind_start*NBNXN_GPU_JGROUP_SIZE;
3693 cj_ind_last = nbl->work->cj_ind - 1;
3695 cj_first = nbl->cj4[nbl_sci->cj4_ind_start].cj[0];
3696 cj_last = nbl_cj(nbl, cj_ind_last);
3698 /* Determine how many contiguous j-clusters we have starting
3699 * from the first i-cluster. This number can be used to directly
3700 * calculate j-cluster indices for excluded atoms.
3703 while (cj_ind_first + ndirect <= cj_ind_last &&
3704 nbl_cj(nbl, cj_ind_first+ndirect) == sci*GPU_NSUBCELL + ndirect)
3709 /* Loop over the atoms in the i super-cell */
3710 for (i = 0; i < nbl->na_sc; i++)
3712 ai = nbs->a[sci*nbl->na_sc+i];
3715 si = (i>>na_c_2log);
3717 /* Loop over the topology-based exclusions for this i-atom */
3718 for (eind = excl->index[ai]; eind < excl->index[ai+1]; eind++)
3724 /* The self exclusion are already set, save some time */
3730 /* Without shifts we only calculate interactions j>i
3731 * for one-way pair-lists.
3733 if (diagRemoved && ge <= sci*nbl->na_sc + i)
3739 /* Could the cluster se be in our list? */
3740 if (se >= cj_first && se <= cj_last)
3742 if (se < cj_first + ndirect)
3744 /* We can calculate cj_ind directly from se */
3745 found = cj_ind_first + se - cj_first;
3749 /* Search for se using bisection */
3751 cj_ind_0 = cj_ind_first + ndirect;
3752 cj_ind_1 = cj_ind_last + 1;
3753 while (found == -1 && cj_ind_0 < cj_ind_1)
3755 cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
3757 cj_m = nbl_cj(nbl, cj_ind_m);
3765 cj_ind_1 = cj_ind_m;
3769 cj_ind_0 = cj_ind_m + 1;
3776 inner_i = i - si*na_c;
3777 inner_e = ge - se*na_c;
3779 if (nbl_imask0(nbl, found) & (1U << (cj_mod_cj4(found)*GPU_NSUBCELL + si)))
3783 get_nbl_exclusions_1(nbl, cj_to_cj4(found), w, &nbl_excl);
3785 nbl_excl->pair[a_mod_wj(inner_e)*nbl->na_ci+inner_i] &=
3786 ~(1U << (cj_mod_cj4(found)*GPU_NSUBCELL + si));
3795 /* Reallocate the simple ci list for at least n entries */
3796 static void nb_realloc_ci(nbnxn_pairlist_t *nbl, int n)
3798 nbl->ci_nalloc = over_alloc_small(n);
3799 nbnxn_realloc_void((void **)&nbl->ci,
3800 nbl->nci*sizeof(*nbl->ci),
3801 nbl->ci_nalloc*sizeof(*nbl->ci),
3802 nbl->alloc, nbl->free);
3805 /* Reallocate the super-cell sci list for at least n entries */
3806 static void nb_realloc_sci(nbnxn_pairlist_t *nbl, int n)
3808 nbl->sci_nalloc = over_alloc_small(n);
3809 nbnxn_realloc_void((void **)&nbl->sci,
3810 nbl->nsci*sizeof(*nbl->sci),
3811 nbl->sci_nalloc*sizeof(*nbl->sci),
3812 nbl->alloc, nbl->free);
3815 /* Make a new ci entry at index nbl->nci */
3816 static void new_ci_entry(nbnxn_pairlist_t *nbl, int ci, int shift, int flags)
3818 if (nbl->nci + 1 > nbl->ci_nalloc)
3820 nb_realloc_ci(nbl, nbl->nci+1);
3822 nbl->ci[nbl->nci].ci = ci;
3823 nbl->ci[nbl->nci].shift = shift;
3824 /* Store the interaction flags along with the shift */
3825 nbl->ci[nbl->nci].shift |= flags;
3826 nbl->ci[nbl->nci].cj_ind_start = nbl->ncj;
3827 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
3830 /* Make a new sci entry at index nbl->nsci */
3831 static void new_sci_entry(nbnxn_pairlist_t *nbl, int sci, int shift)
3833 if (nbl->nsci + 1 > nbl->sci_nalloc)
3835 nb_realloc_sci(nbl, nbl->nsci+1);
3837 nbl->sci[nbl->nsci].sci = sci;
3838 nbl->sci[nbl->nsci].shift = shift;
3839 nbl->sci[nbl->nsci].cj4_ind_start = nbl->ncj4;
3840 nbl->sci[nbl->nsci].cj4_ind_end = nbl->ncj4;
3843 /* Sort the simple j-list cj on exclusions.
3844 * Entries with exclusions will all be sorted to the beginning of the list.
3846 static void sort_cj_excl(nbnxn_cj_t *cj, int ncj,
3847 nbnxn_list_work_t *work)
3851 if (ncj > work->cj_nalloc)
3853 work->cj_nalloc = over_alloc_large(ncj);
3854 srenew(work->cj, work->cj_nalloc);
3857 /* Make a list of the j-cells involving exclusions */
3859 for (j = 0; j < ncj; j++)
3861 if (cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
3863 work->cj[jnew++] = cj[j];
3866 /* Check if there are exclusions at all or not just the first entry */
3867 if (!((jnew == 0) ||
3868 (jnew == 1 && cj[0].excl != NBNXN_INTERACTION_MASK_ALL)))
3870 for (j = 0; j < ncj; j++)
3872 if (cj[j].excl == NBNXN_INTERACTION_MASK_ALL)
3874 work->cj[jnew++] = cj[j];
3877 for (j = 0; j < ncj; j++)
3879 cj[j] = work->cj[j];
3884 /* Close this simple list i entry */
3885 static void close_ci_entry_simple(nbnxn_pairlist_t *nbl)
3889 /* All content of the new ci entry have already been filled correctly,
3890 * we only need to increase the count here (for non empty lists).
3892 jlen = nbl->ci[nbl->nci].cj_ind_end - nbl->ci[nbl->nci].cj_ind_start;
3895 sort_cj_excl(nbl->cj+nbl->ci[nbl->nci].cj_ind_start, jlen, nbl->work);
3897 /* The counts below are used for non-bonded pair/flop counts
3898 * and should therefore match the available kernel setups.
3900 if (!(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_COUL(0)))
3902 nbl->work->ncj_noq += jlen;
3904 else if ((nbl->ci[nbl->nci].shift & NBNXN_CI_HALF_LJ(0)) ||
3905 !(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_LJ(0)))
3907 nbl->work->ncj_hlj += jlen;
3914 /* Split sci entry for load balancing on the GPU.
3915 * Splitting ensures we have enough lists to fully utilize the whole GPU.
3916 * With progBal we generate progressively smaller lists, which improves
3917 * load balancing. As we only know the current count on our own thread,
3918 * we will need to estimate the current total amount of i-entries.
3919 * As the lists get concatenated later, this estimate depends
3920 * both on nthread and our own thread index.
3922 static void split_sci_entry(nbnxn_pairlist_t *nbl,
3923 int nsp_max_av, gmx_bool progBal, int nc_bal,
3924 int thread, int nthread)
3928 int cj4_start, cj4_end, j4len, cj4;
3930 int nsp, nsp_sci, nsp_cj4, nsp_cj4_e, nsp_cj4_p;
3935 /* Estimate the total numbers of ci's of the nblist combined
3936 * over all threads using the target number of ci's.
3938 nsci_est = nc_bal*thread/nthread + nbl->nsci;
3940 /* The first ci blocks should be larger, to avoid overhead.
3941 * The last ci blocks should be smaller, to improve load balancing.
3944 nsp_max_av*nc_bal*3/(2*(nsci_est - 1 + nc_bal)));
3948 nsp_max = nsp_max_av;
3951 cj4_start = nbl->sci[nbl->nsci-1].cj4_ind_start;
3952 cj4_end = nbl->sci[nbl->nsci-1].cj4_ind_end;
3953 j4len = cj4_end - cj4_start;
3955 if (j4len > 1 && j4len*GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE > nsp_max)
3957 /* Remove the last ci entry and process the cj4's again */
3965 for (cj4 = cj4_start; cj4 < cj4_end; cj4++)
3967 nsp_cj4_p = nsp_cj4;
3968 /* Count the number of cluster pairs in this cj4 group */
3970 for (p = 0; p < GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE; p++)
3972 nsp_cj4 += (nbl->cj4[cj4].imei[0].imask >> p) & 1;
3975 if (nsp_cj4 > 0 && nsp + nsp_cj4 > nsp_max)
3977 /* Split the list at cj4 */
3978 nbl->sci[sci].cj4_ind_end = cj4;
3979 /* Create a new sci entry */
3982 if (nbl->nsci+1 > nbl->sci_nalloc)
3984 nb_realloc_sci(nbl, nbl->nsci+1);
3986 nbl->sci[sci].sci = nbl->sci[nbl->nsci-1].sci;
3987 nbl->sci[sci].shift = nbl->sci[nbl->nsci-1].shift;
3988 nbl->sci[sci].cj4_ind_start = cj4;
3990 nsp_cj4_e = nsp_cj4_p;
3996 /* Put the remaining cj4's in the last sci entry */
3997 nbl->sci[sci].cj4_ind_end = cj4_end;
3999 /* Possibly balance out the last two sci's
4000 * by moving the last cj4 of the second last sci.
4002 if (nsp_sci - nsp_cj4_e >= nsp + nsp_cj4_e)
4004 nbl->sci[sci-1].cj4_ind_end--;
4005 nbl->sci[sci].cj4_ind_start--;
4012 /* Clost this super/sub list i entry */
4013 static void close_ci_entry_supersub(nbnxn_pairlist_t *nbl,
4015 gmx_bool progBal, int nc_bal,
4016 int thread, int nthread)
4021 /* All content of the new ci entry have already been filled correctly,
4022 * we only need to increase the count here (for non empty lists).
4024 j4len = nbl->sci[nbl->nsci].cj4_ind_end - nbl->sci[nbl->nsci].cj4_ind_start;
4027 /* We can only have complete blocks of 4 j-entries in a list,
4028 * so round the count up before closing.
4030 nbl->ncj4 = ((nbl->work->cj_ind + NBNXN_GPU_JGROUP_SIZE - 1) >> NBNXN_GPU_JGROUP_SIZE_2LOG);
4031 nbl->work->cj_ind = nbl->ncj4*NBNXN_GPU_JGROUP_SIZE;
4037 /* Measure the size of the new entry and potentially split it */
4038 split_sci_entry(nbl, nsp_max_av, progBal, nc_bal, thread, nthread);
4043 /* Syncs the working array before adding another grid pair to the list */
4044 static void sync_work(nbnxn_pairlist_t *nbl)
4048 nbl->work->cj_ind = nbl->ncj4*NBNXN_GPU_JGROUP_SIZE;
4049 nbl->work->cj4_init = nbl->ncj4;
4053 /* Clears an nbnxn_pairlist_t data structure */
4054 static void clear_pairlist(nbnxn_pairlist_t *nbl)
4063 nbl->work->ncj_noq = 0;
4064 nbl->work->ncj_hlj = 0;
4067 /* Clears a group scheme pair list */
4068 static void clear_pairlist_fep(t_nblist *nl)
4072 if (nl->jindex == NULL)
4074 snew(nl->jindex, 1);
4079 /* Sets a simple list i-cell bounding box, including PBC shift */
4080 static gmx_inline void set_icell_bb_simple(const nbnxn_bb_t *bb, int ci,
4081 real shx, real shy, real shz,
4084 bb_ci->lower[BB_X] = bb[ci].lower[BB_X] + shx;
4085 bb_ci->lower[BB_Y] = bb[ci].lower[BB_Y] + shy;
4086 bb_ci->lower[BB_Z] = bb[ci].lower[BB_Z] + shz;
4087 bb_ci->upper[BB_X] = bb[ci].upper[BB_X] + shx;
4088 bb_ci->upper[BB_Y] = bb[ci].upper[BB_Y] + shy;
4089 bb_ci->upper[BB_Z] = bb[ci].upper[BB_Z] + shz;
4093 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
4094 static void set_icell_bbxxxx_supersub(const float *bb, int ci,
4095 real shx, real shy, real shz,
4100 ia = ci*(GPU_NSUBCELL>>STRIDE_PBB_2LOG)*NNBSBB_XXXX;
4101 for (m = 0; m < (GPU_NSUBCELL>>STRIDE_PBB_2LOG)*NNBSBB_XXXX; m += NNBSBB_XXXX)
4103 for (i = 0; i < STRIDE_PBB; i++)
4105 bb_ci[m+0*STRIDE_PBB+i] = bb[ia+m+0*STRIDE_PBB+i] + shx;
4106 bb_ci[m+1*STRIDE_PBB+i] = bb[ia+m+1*STRIDE_PBB+i] + shy;
4107 bb_ci[m+2*STRIDE_PBB+i] = bb[ia+m+2*STRIDE_PBB+i] + shz;
4108 bb_ci[m+3*STRIDE_PBB+i] = bb[ia+m+3*STRIDE_PBB+i] + shx;
4109 bb_ci[m+4*STRIDE_PBB+i] = bb[ia+m+4*STRIDE_PBB+i] + shy;
4110 bb_ci[m+5*STRIDE_PBB+i] = bb[ia+m+5*STRIDE_PBB+i] + shz;
4116 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
4117 static void set_icell_bb_supersub(const nbnxn_bb_t *bb, int ci,
4118 real shx, real shy, real shz,
4123 for (i = 0; i < GPU_NSUBCELL; i++)
4125 set_icell_bb_simple(bb, ci*GPU_NSUBCELL+i,
4131 /* Copies PBC shifted i-cell atom coordinates x,y,z to working array */
4132 static void icell_set_x_simple(int ci,
4133 real shx, real shy, real shz,
4134 int gmx_unused na_c,
4135 int stride, const real *x,
4136 nbnxn_list_work_t *work)
4140 ia = ci*NBNXN_CPU_CLUSTER_I_SIZE;
4142 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE; i++)
4144 work->x_ci[i*STRIDE_XYZ+XX] = x[(ia+i)*stride+XX] + shx;
4145 work->x_ci[i*STRIDE_XYZ+YY] = x[(ia+i)*stride+YY] + shy;
4146 work->x_ci[i*STRIDE_XYZ+ZZ] = x[(ia+i)*stride+ZZ] + shz;
4150 /* Copies PBC shifted super-cell atom coordinates x,y,z to working array */
4151 static void icell_set_x_supersub(int ci,
4152 real shx, real shy, real shz,
4154 int stride, const real *x,
4155 nbnxn_list_work_t *work)
4162 ia = ci*GPU_NSUBCELL*na_c;
4163 for (i = 0; i < GPU_NSUBCELL*na_c; i++)
4165 x_ci[i*DIM + XX] = x[(ia+i)*stride + XX] + shx;
4166 x_ci[i*DIM + YY] = x[(ia+i)*stride + YY] + shy;
4167 x_ci[i*DIM + ZZ] = x[(ia+i)*stride + ZZ] + shz;
4171 #ifdef NBNXN_SEARCH_BB_SIMD4
4172 /* Copies PBC shifted super-cell packed atom coordinates to working array */
4173 static void icell_set_x_supersub_simd4(int ci,
4174 real shx, real shy, real shz,
4176 int stride, const real *x,
4177 nbnxn_list_work_t *work)
4179 int si, io, ia, i, j;
4184 for (si = 0; si < GPU_NSUBCELL; si++)
4186 for (i = 0; i < na_c; i += STRIDE_PBB)
4189 ia = ci*GPU_NSUBCELL*na_c + io;
4190 for (j = 0; j < STRIDE_PBB; j++)
4192 x_ci[io*DIM + j + XX*STRIDE_PBB] = x[(ia+j)*stride+XX] + shx;
4193 x_ci[io*DIM + j + YY*STRIDE_PBB] = x[(ia+j)*stride+YY] + shy;
4194 x_ci[io*DIM + j + ZZ*STRIDE_PBB] = x[(ia+j)*stride+ZZ] + shz;
4201 static real minimum_subgrid_size_xy(const nbnxn_grid_t *grid)
4205 return min(grid->sx, grid->sy);
4209 return min(grid->sx/GPU_NSUBCELL_X, grid->sy/GPU_NSUBCELL_Y);
4213 static real effective_buffer_1x1_vs_MxN(const nbnxn_grid_t *gridi,
4214 const nbnxn_grid_t *gridj)
4216 const real eff_1x1_buffer_fac_overest = 0.1;
4218 /* Determine an atom-pair list cut-off buffer size for atom pairs,
4219 * to be added to rlist (including buffer) used for MxN.
4220 * This is for converting an MxN list to a 1x1 list. This means we can't
4221 * use the normal buffer estimate, as we have an MxN list in which
4222 * some atom pairs beyond rlist are missing. We want to capture
4223 * the beneficial effect of buffering by extra pairs just outside rlist,
4224 * while removing the useless pairs that are further away from rlist.
4225 * (Also the buffer could have been set manually not using the estimate.)
4226 * This buffer size is an overestimate.
4227 * We add 10% of the smallest grid sub-cell dimensions.
4228 * Note that the z-size differs per cell and we don't use this,
4229 * so we overestimate.
4230 * With PME, the 10% value gives a buffer that is somewhat larger
4231 * than the effective buffer with a tolerance of 0.005 kJ/mol/ps.
4232 * Smaller tolerances or using RF lead to a smaller effective buffer,
4233 * so 10% gives a safe overestimate.
4235 return eff_1x1_buffer_fac_overest*(minimum_subgrid_size_xy(gridi) +
4236 minimum_subgrid_size_xy(gridj));
4239 /* Clusters at the cut-off only increase rlist by 60% of their size */
4240 static real nbnxn_rlist_inc_outside_fac = 0.6;
4242 /* Due to the cluster size the effective pair-list is longer than
4243 * that of a simple atom pair-list. This function gives the extra distance.
4245 real nbnxn_get_rlist_effective_inc(int cluster_size_j, real atom_density)
4248 real vol_inc_i, vol_inc_j;
4250 /* We should get this from the setup, but currently it's the same for
4251 * all setups, including GPUs.
4253 cluster_size_i = NBNXN_CPU_CLUSTER_I_SIZE;
4255 vol_inc_i = (cluster_size_i - 1)/atom_density;
4256 vol_inc_j = (cluster_size_j - 1)/atom_density;
4258 return nbnxn_rlist_inc_outside_fac*pow(vol_inc_i + vol_inc_j, 1.0/3.0);
4261 /* Estimates the interaction volume^2 for non-local interactions */
4262 static real nonlocal_vol2(const gmx_domdec_zones_t *zones, rvec ls, real r)
4271 /* Here we simply add up the volumes of 1, 2 or 3 1D decomposition
4272 * not home interaction volume^2. As these volumes are not additive,
4273 * this is an overestimate, but it would only be significant in the limit
4274 * of small cells, where we anyhow need to split the lists into
4275 * as small parts as possible.
4278 for (z = 0; z < zones->n; z++)
4280 if (zones->shift[z][XX] + zones->shift[z][YY] + zones->shift[z][ZZ] == 1)
4285 for (d = 0; d < DIM; d++)
4287 if (zones->shift[z][d] == 0)
4291 za *= zones->size[z].x1[d] - zones->size[z].x0[d];
4295 /* 4 octants of a sphere */
4296 vold_est = 0.25*M_PI*r*r*r*r;
4297 /* 4 quarter pie slices on the edges */
4298 vold_est += 4*cl*M_PI/6.0*r*r*r;
4299 /* One rectangular volume on a face */
4300 vold_est += ca*0.5*r*r;
4302 vol2_est_tot += vold_est*za;
4306 return vol2_est_tot;
4309 /* Estimates the average size of a full j-list for super/sub setup */
4310 static int get_nsubpair_max(const nbnxn_search_t nbs,
4313 int min_ci_balanced)
4315 const nbnxn_grid_t *grid;
4317 real xy_diag2, r_eff_sup, vol_est, nsp_est, nsp_est_nl;
4320 grid = &nbs->grid[0];
4322 ls[XX] = (grid->c1[XX] - grid->c0[XX])/(grid->ncx*GPU_NSUBCELL_X);
4323 ls[YY] = (grid->c1[YY] - grid->c0[YY])/(grid->ncy*GPU_NSUBCELL_Y);
4324 ls[ZZ] = (grid->c1[ZZ] - grid->c0[ZZ])*grid->ncx*grid->ncy/(grid->nc*GPU_NSUBCELL_Z);
4326 /* The average squared length of the diagonal of a sub cell */
4327 xy_diag2 = ls[XX]*ls[XX] + ls[YY]*ls[YY] + ls[ZZ]*ls[ZZ];
4329 /* The formulas below are a heuristic estimate of the average nsj per si*/
4330 r_eff_sup = rlist + nbnxn_rlist_inc_outside_fac*sqr((grid->na_c - 1.0)/grid->na_c)*sqrt(xy_diag2/3);
4332 if (!nbs->DomDec || nbs->zones->n == 1)
4339 sqr(grid->atom_density/grid->na_c)*
4340 nonlocal_vol2(nbs->zones, ls, r_eff_sup);
4345 /* Sub-cell interacts with itself */
4346 vol_est = ls[XX]*ls[YY]*ls[ZZ];
4347 /* 6/2 rectangular volume on the faces */
4348 vol_est += (ls[XX]*ls[YY] + ls[XX]*ls[ZZ] + ls[YY]*ls[ZZ])*r_eff_sup;
4349 /* 12/2 quarter pie slices on the edges */
4350 vol_est += 2*(ls[XX] + ls[YY] + ls[ZZ])*0.25*M_PI*sqr(r_eff_sup);
4351 /* 4 octants of a sphere */
4352 vol_est += 0.5*4.0/3.0*M_PI*pow(r_eff_sup, 3);
4354 nsp_est = grid->nsubc_tot*vol_est*grid->atom_density/grid->na_c;
4356 /* Subtract the non-local pair count */
4357 nsp_est -= nsp_est_nl;
4361 fprintf(debug, "nsp_est local %5.1f non-local %5.1f\n",
4362 nsp_est, nsp_est_nl);
4367 nsp_est = nsp_est_nl;
4370 if (min_ci_balanced <= 0 || grid->nc >= min_ci_balanced || grid->nc == 0)
4372 /* We don't need to worry */
4377 /* Thus the (average) maximum j-list size should be as follows */
4378 nsubpair_max = max(1, (int)(nsp_est/min_ci_balanced+0.5));
4380 /* Since the target value is a maximum (this avoids high outliers,
4381 * which lead to load imbalance), not average, we add half the
4382 * number of pairs in a cj4 block to get the average about right.
4384 nsubpair_max += GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE/2;
4389 fprintf(debug, "nbl nsp estimate %.1f, nsubpair_max %d\n",
4390 nsp_est, nsubpair_max);
4393 return nsubpair_max;
4396 /* Debug list print function */
4397 static void print_nblist_ci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
4401 for (i = 0; i < nbl->nci; i++)
4403 fprintf(fp, "ci %4d shift %2d ncj %3d\n",
4404 nbl->ci[i].ci, nbl->ci[i].shift,
4405 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start);
4407 for (j = nbl->ci[i].cj_ind_start; j < nbl->ci[i].cj_ind_end; j++)
4409 fprintf(fp, " cj %5d imask %x\n",
4416 /* Debug list print function */
4417 static void print_nblist_sci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
4419 int i, j4, j, ncp, si;
4421 for (i = 0; i < nbl->nsci; i++)
4423 fprintf(fp, "ci %4d shift %2d ncj4 %2d\n",
4424 nbl->sci[i].sci, nbl->sci[i].shift,
4425 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start);
4428 for (j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
4430 for (j = 0; j < NBNXN_GPU_JGROUP_SIZE; j++)
4432 fprintf(fp, " sj %5d imask %x\n",
4434 nbl->cj4[j4].imei[0].imask);
4435 for (si = 0; si < GPU_NSUBCELL; si++)
4437 if (nbl->cj4[j4].imei[0].imask & (1U << (j*GPU_NSUBCELL + si)))
4444 fprintf(fp, "ci %4d shift %2d ncj4 %2d ncp %3d\n",
4445 nbl->sci[i].sci, nbl->sci[i].shift,
4446 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start,
4451 /* Combine pair lists *nbl generated on multiple threads nblc */
4452 static void combine_nblists(int nnbl, nbnxn_pairlist_t **nbl,
4453 nbnxn_pairlist_t *nblc)
4455 int nsci, ncj4, nexcl;
4460 gmx_incons("combine_nblists does not support simple lists");
4465 nexcl = nblc->nexcl;
4466 for (i = 0; i < nnbl; i++)
4468 nsci += nbl[i]->nsci;
4469 ncj4 += nbl[i]->ncj4;
4470 nexcl += nbl[i]->nexcl;
4473 if (nsci > nblc->sci_nalloc)
4475 nb_realloc_sci(nblc, nsci);
4477 if (ncj4 > nblc->cj4_nalloc)
4479 nblc->cj4_nalloc = over_alloc_small(ncj4);
4480 nbnxn_realloc_void((void **)&nblc->cj4,
4481 nblc->ncj4*sizeof(*nblc->cj4),
4482 nblc->cj4_nalloc*sizeof(*nblc->cj4),
4483 nblc->alloc, nblc->free);
4485 if (nexcl > nblc->excl_nalloc)
4487 nblc->excl_nalloc = over_alloc_small(nexcl);
4488 nbnxn_realloc_void((void **)&nblc->excl,
4489 nblc->nexcl*sizeof(*nblc->excl),
4490 nblc->excl_nalloc*sizeof(*nblc->excl),
4491 nblc->alloc, nblc->free);
4494 /* Each thread should copy its own data to the combined arrays,
4495 * as otherwise data will go back and forth between different caches.
4497 #pragma omp parallel for num_threads(gmx_omp_nthreads_get(emntPairsearch)) schedule(static)
4498 for (n = 0; n < nnbl; n++)
4505 const nbnxn_pairlist_t *nbli;
4507 /* Determine the offset in the combined data for our thread */
4508 sci_offset = nblc->nsci;
4509 cj4_offset = nblc->ncj4;
4510 ci_offset = nblc->nci_tot;
4511 excl_offset = nblc->nexcl;
4513 for (i = 0; i < n; i++)
4515 sci_offset += nbl[i]->nsci;
4516 cj4_offset += nbl[i]->ncj4;
4517 ci_offset += nbl[i]->nci_tot;
4518 excl_offset += nbl[i]->nexcl;
4523 for (i = 0; i < nbli->nsci; i++)
4525 nblc->sci[sci_offset+i] = nbli->sci[i];
4526 nblc->sci[sci_offset+i].cj4_ind_start += cj4_offset;
4527 nblc->sci[sci_offset+i].cj4_ind_end += cj4_offset;
4530 for (j4 = 0; j4 < nbli->ncj4; j4++)
4532 nblc->cj4[cj4_offset+j4] = nbli->cj4[j4];
4533 nblc->cj4[cj4_offset+j4].imei[0].excl_ind += excl_offset;
4534 nblc->cj4[cj4_offset+j4].imei[1].excl_ind += excl_offset;
4537 for (j4 = 0; j4 < nbli->nexcl; j4++)
4539 nblc->excl[excl_offset+j4] = nbli->excl[j4];
4543 for (n = 0; n < nnbl; n++)
4545 nblc->nsci += nbl[n]->nsci;
4546 nblc->ncj4 += nbl[n]->ncj4;
4547 nblc->nci_tot += nbl[n]->nci_tot;
4548 nblc->nexcl += nbl[n]->nexcl;
4552 static void balance_fep_lists(const nbnxn_search_t nbs,
4553 nbnxn_pairlist_set_t *nbl_lists)
4556 int nri_tot, nrj_tot, nrj_target;
4560 nnbl = nbl_lists->nnbl;
4564 /* Nothing to balance */
4568 /* Count the total i-lists and pairs */
4571 for (th = 0; th < nnbl; th++)
4573 nri_tot += nbl_lists->nbl_fep[th]->nri;
4574 nrj_tot += nbl_lists->nbl_fep[th]->nrj;
4577 nrj_target = (nrj_tot + nnbl - 1)/nnbl;
4579 assert(gmx_omp_nthreads_get(emntNonbonded) == nnbl);
4581 #pragma omp parallel for schedule(static) num_threads(nnbl)
4582 for (th = 0; th < nnbl; th++)
4586 nbl = nbs->work[th].nbl_fep;
4588 /* Note that here we allocate for the total size, instead of
4589 * a per-thread esimate (which is hard to obtain).
4591 if (nri_tot > nbl->maxnri)
4593 nbl->maxnri = over_alloc_large(nri_tot);
4594 reallocate_nblist(nbl);
4596 if (nri_tot > nbl->maxnri || nrj_tot > nbl->maxnrj)
4598 nbl->maxnrj = over_alloc_small(nrj_tot);
4599 srenew(nbl->jjnr, nbl->maxnrj);
4600 srenew(nbl->excl_fep, nbl->maxnrj);
4603 clear_pairlist_fep(nbl);
4606 /* Loop over the source lists and assign and copy i-entries */
4608 nbld = nbs->work[th_dest].nbl_fep;
4609 for (th = 0; th < nnbl; th++)
4614 nbls = nbl_lists->nbl_fep[th];
4616 for (i = 0; i < nbls->nri; i++)
4620 /* The number of pairs in this i-entry */
4621 nrj = nbls->jindex[i+1] - nbls->jindex[i];
4623 /* Decide if list th_dest is too large and we should procede
4624 * to the next destination list.
4626 if (th_dest+1 < nnbl && nbld->nrj > 0 &&
4627 nbld->nrj + nrj - nrj_target > nrj_target - nbld->nrj)
4630 nbld = nbs->work[th_dest].nbl_fep;
4633 nbld->iinr[nbld->nri] = nbls->iinr[i];
4634 nbld->gid[nbld->nri] = nbls->gid[i];
4635 nbld->shift[nbld->nri] = nbls->shift[i];
4637 for (j = nbls->jindex[i]; j < nbls->jindex[i+1]; j++)
4639 nbld->jjnr[nbld->nrj] = nbls->jjnr[j];
4640 nbld->excl_fep[nbld->nrj] = nbls->excl_fep[j];
4644 nbld->jindex[nbld->nri] = nbld->nrj;
4648 /* Swap the list pointers */
4649 for (th = 0; th < nnbl; th++)
4653 nbl_tmp = nbl_lists->nbl_fep[th];
4654 nbl_lists->nbl_fep[th] = nbs->work[th].nbl_fep;
4655 nbs->work[th].nbl_fep = nbl_tmp;
4659 fprintf(debug, "nbl_fep[%d] nri %4d nrj %4d\n",
4661 nbl_lists->nbl_fep[th]->nri,
4662 nbl_lists->nbl_fep[th]->nrj);
4667 /* Returns the next ci to be processes by our thread */
4668 static gmx_bool next_ci(const nbnxn_grid_t *grid,
4670 int nth, int ci_block,
4671 int *ci_x, int *ci_y,
4677 if (*ci_b == ci_block)
4679 /* Jump to the next block assigned to this task */
4680 *ci += (nth - 1)*ci_block;
4684 if (*ci >= grid->nc*conv)
4689 while (*ci >= grid->cxy_ind[*ci_x*grid->ncy + *ci_y + 1]*conv)
4692 if (*ci_y == grid->ncy)
4702 /* Returns the distance^2 for which we put cell pairs in the list
4703 * without checking atom pair distances. This is usually < rlist^2.
4705 static float boundingbox_only_distance2(const nbnxn_grid_t *gridi,
4706 const nbnxn_grid_t *gridj,
4710 /* If the distance between two sub-cell bounding boxes is less
4711 * than this distance, do not check the distance between
4712 * all particle pairs in the sub-cell, since then it is likely
4713 * that the box pair has atom pairs within the cut-off.
4714 * We use the nblist cut-off minus 0.5 times the average x/y diagonal
4715 * spacing of the sub-cells. Around 40% of the checked pairs are pruned.
4716 * Using more than 0.5 gains at most 0.5%.
4717 * If forces are calculated more than twice, the performance gain
4718 * in the force calculation outweighs the cost of checking.
4719 * Note that with subcell lists, the atom-pair distance check
4720 * is only performed when only 1 out of 8 sub-cells in within range,
4721 * this is because the GPU is much faster than the cpu.
4726 bbx = 0.5*(gridi->sx + gridj->sx);
4727 bby = 0.5*(gridi->sy + gridj->sy);
4730 bbx /= GPU_NSUBCELL_X;
4731 bby /= GPU_NSUBCELL_Y;
4734 rbb2 = sqr(max(0, rlist - 0.5*sqrt(bbx*bbx + bby*bby)));
4739 return (float)((1+GMX_FLOAT_EPS)*rbb2);
4743 static int get_ci_block_size(const nbnxn_grid_t *gridi,
4744 gmx_bool bDomDec, int nth)
4746 const int ci_block_enum = 5;
4747 const int ci_block_denom = 11;
4748 const int ci_block_min_atoms = 16;
4751 /* Here we decide how to distribute the blocks over the threads.
4752 * We use prime numbers to try to avoid that the grid size becomes
4753 * a multiple of the number of threads, which would lead to some
4754 * threads getting "inner" pairs and others getting boundary pairs,
4755 * which in turns will lead to load imbalance between threads.
4756 * Set the block size as 5/11/ntask times the average number of cells
4757 * in a y,z slab. This should ensure a quite uniform distribution
4758 * of the grid parts of the different thread along all three grid
4759 * zone boundaries with 3D domain decomposition. At the same time
4760 * the blocks will not become too small.
4762 ci_block = (gridi->nc*ci_block_enum)/(ci_block_denom*gridi->ncx*nth);
4764 /* Ensure the blocks are not too small: avoids cache invalidation */
4765 if (ci_block*gridi->na_sc < ci_block_min_atoms)
4767 ci_block = (ci_block_min_atoms + gridi->na_sc - 1)/gridi->na_sc;
4770 /* Without domain decomposition
4771 * or with less than 3 blocks per task, divide in nth blocks.
4773 if (!bDomDec || ci_block*3*nth > gridi->nc)
4775 ci_block = (gridi->nc + nth - 1)/nth;
4781 /* Generates the part of pair-list nbl assigned to our thread */
4782 static void nbnxn_make_pairlist_part(const nbnxn_search_t nbs,
4783 const nbnxn_grid_t *gridi,
4784 const nbnxn_grid_t *gridj,
4785 nbnxn_search_work_t *work,
4786 const nbnxn_atomdata_t *nbat,
4787 const t_blocka *excl,
4791 gmx_bool bFBufferFlag,
4794 int min_ci_balanced,
4796 nbnxn_pairlist_t *nbl,
4801 real rl2, rl_fep2 = 0;
4804 int ci_b, ci, ci_x, ci_y, ci_xy, cj;
4810 int conv_i, cell0_i;
4811 const nbnxn_bb_t *bb_i = NULL;
4813 const float *pbb_i = NULL;
4815 const float *bbcz_i, *bbcz_j;
4817 real bx0, bx1, by0, by1, bz0, bz1;
4819 real d2cx, d2z, d2z_cx, d2z_cy, d2zx, d2zxy, d2xy;
4820 int cxf, cxl, cyf, cyf_x, cyl;
4822 int c0, c1, cs, cf, cl;
4825 int gridi_flag_shift = 0, gridj_flag_shift = 0;
4826 unsigned int *gridj_flag = NULL;
4827 int ncj_old_i, ncj_old_j;
4829 nbs_cycle_start(&work->cc[enbsCCsearch]);
4831 if (gridj->bSimple != nbl->bSimple)
4833 gmx_incons("Grid incompatible with pair-list");
4837 nbl->na_sc = gridj->na_sc;
4838 nbl->na_ci = gridj->na_c;
4839 nbl->na_cj = nbnxn_kernel_to_cj_size(nb_kernel_type);
4840 na_cj_2log = get_2log(nbl->na_cj);
4846 /* Determine conversion of clusters to flag blocks */
4847 gridi_flag_shift = 0;
4848 while ((nbl->na_ci<<gridi_flag_shift) < NBNXN_BUFFERFLAG_SIZE)
4852 gridj_flag_shift = 0;
4853 while ((nbl->na_cj<<gridj_flag_shift) < NBNXN_BUFFERFLAG_SIZE)
4858 gridj_flag = work->buffer_flags.flag;
4861 copy_mat(nbs->box, box);
4863 rl2 = nbl->rlist*nbl->rlist;
4865 if (nbs->bFEP && !nbl->bSimple)
4867 /* Determine an atom-pair list cut-off distance for FEP atom pairs.
4868 * We should not simply use rlist, since then we would not have
4869 * the small, effective buffering of the NxN lists.
4870 * The buffer is on overestimate, but the resulting cost for pairs
4871 * beyond rlist is neglible compared to the FEP pairs within rlist.
4873 rl_fep2 = nbl->rlist + effective_buffer_1x1_vs_MxN(gridi, gridj);
4877 fprintf(debug, "nbl_fep atom-pair rlist %f\n", rl_fep2);
4879 rl_fep2 = rl_fep2*rl_fep2;
4882 rbb2 = boundingbox_only_distance2(gridi, gridj, nbl->rlist, nbl->bSimple);
4886 fprintf(debug, "nbl bounding box only distance %f\n", sqrt(rbb2));
4889 /* Set the shift range */
4890 for (d = 0; d < DIM; d++)
4892 /* Check if we need periodicity shifts.
4893 * Without PBC or with domain decomposition we don't need them.
4895 if (d >= ePBC2npbcdim(nbs->ePBC) || nbs->dd_dim[d])
4902 box[XX][XX] - fabs(box[YY][XX]) - fabs(box[ZZ][XX]) < sqrt(rl2))
4913 if (nbl->bSimple && !gridi->bSimple)
4915 conv_i = gridi->na_sc/gridj->na_sc;
4916 bb_i = gridi->bb_simple;
4917 bbcz_i = gridi->bbcz_simple;
4918 flags_i = gridi->flags_simple;
4933 /* We use the normal bounding box format for both grid types */
4936 bbcz_i = gridi->bbcz;
4937 flags_i = gridi->flags;
4939 cell0_i = gridi->cell0*conv_i;
4941 bbcz_j = gridj->bbcz;
4945 /* Blocks of the conversion factor - 1 give a large repeat count
4946 * combined with a small block size. This should result in good
4947 * load balancing for both small and large domains.
4949 ci_block = conv_i - 1;
4953 fprintf(debug, "nbl nc_i %d col.av. %.1f ci_block %d\n",
4954 gridi->nc, gridi->nc/(double)(gridi->ncx*gridi->ncy), ci_block);
4960 /* Initially ci_b and ci to 1 before where we want them to start,
4961 * as they will both be incremented in next_ci.
4964 ci = th*ci_block - 1;
4967 while (next_ci(gridi, conv_i, nth, ci_block, &ci_x, &ci_y, &ci_b, &ci))
4969 if (nbl->bSimple && flags_i[ci] == 0)
4974 ncj_old_i = nbl->ncj;
4977 if (gridj != gridi && shp[XX] == 0)
4981 bx1 = bb_i[ci].upper[BB_X];
4985 bx1 = gridi->c0[XX] + (ci_x+1)*gridi->sx;
4987 if (bx1 < gridj->c0[XX])
4989 d2cx = sqr(gridj->c0[XX] - bx1);
4998 ci_xy = ci_x*gridi->ncy + ci_y;
5000 /* Loop over shift vectors in three dimensions */
5001 for (tz = -shp[ZZ]; tz <= shp[ZZ]; tz++)
5003 shz = tz*box[ZZ][ZZ];
5005 bz0 = bbcz_i[ci*NNBSBB_D ] + shz;
5006 bz1 = bbcz_i[ci*NNBSBB_D+1] + shz;
5018 d2z = sqr(bz0 - box[ZZ][ZZ]);
5021 d2z_cx = d2z + d2cx;
5029 bz1/((real)(gridi->cxy_ind[ci_xy+1] - gridi->cxy_ind[ci_xy]));
5034 /* The check with bz1_frac close to or larger than 1 comes later */
5036 for (ty = -shp[YY]; ty <= shp[YY]; ty++)
5038 shy = ty*box[YY][YY] + tz*box[ZZ][YY];
5042 by0 = bb_i[ci].lower[BB_Y] + shy;
5043 by1 = bb_i[ci].upper[BB_Y] + shy;
5047 by0 = gridi->c0[YY] + (ci_y )*gridi->sy + shy;
5048 by1 = gridi->c0[YY] + (ci_y+1)*gridi->sy + shy;
5051 get_cell_range(by0, by1,
5052 gridj->ncy, gridj->c0[YY], gridj->sy, gridj->inv_sy,
5062 if (by1 < gridj->c0[YY])
5064 d2z_cy += sqr(gridj->c0[YY] - by1);
5066 else if (by0 > gridj->c1[YY])
5068 d2z_cy += sqr(by0 - gridj->c1[YY]);
5071 for (tx = -shp[XX]; tx <= shp[XX]; tx++)
5073 shift = XYZ2IS(tx, ty, tz);
5075 #ifdef NBNXN_SHIFT_BACKWARD
5076 if (gridi == gridj && shift > CENTRAL)
5082 shx = tx*box[XX][XX] + ty*box[YY][XX] + tz*box[ZZ][XX];
5086 bx0 = bb_i[ci].lower[BB_X] + shx;
5087 bx1 = bb_i[ci].upper[BB_X] + shx;
5091 bx0 = gridi->c0[XX] + (ci_x )*gridi->sx + shx;
5092 bx1 = gridi->c0[XX] + (ci_x+1)*gridi->sx + shx;
5095 get_cell_range(bx0, bx1,
5096 gridj->ncx, gridj->c0[XX], gridj->sx, gridj->inv_sx,
5107 new_ci_entry(nbl, cell0_i+ci, shift, flags_i[ci]);
5111 new_sci_entry(nbl, cell0_i+ci, shift);
5114 #ifndef NBNXN_SHIFT_BACKWARD
5117 if (shift == CENTRAL && gridi == gridj &&
5121 /* Leave the pairs with i > j.
5122 * x is the major index, so skip half of it.
5129 set_icell_bb_simple(bb_i, ci, shx, shy, shz,
5135 set_icell_bbxxxx_supersub(pbb_i, ci, shx, shy, shz,
5138 set_icell_bb_supersub(bb_i, ci, shx, shy, shz,
5143 nbs->icell_set_x(cell0_i+ci, shx, shy, shz,
5144 gridi->na_c, nbat->xstride, nbat->x,
5147 for (cx = cxf; cx <= cxl; cx++)
5150 if (gridj->c0[XX] + cx*gridj->sx > bx1)
5152 d2zx += sqr(gridj->c0[XX] + cx*gridj->sx - bx1);
5154 else if (gridj->c0[XX] + (cx+1)*gridj->sx < bx0)
5156 d2zx += sqr(gridj->c0[XX] + (cx+1)*gridj->sx - bx0);
5159 #ifndef NBNXN_SHIFT_BACKWARD
5160 if (gridi == gridj &&
5161 cx == 0 && cyf < ci_y)
5163 if (gridi == gridj &&
5164 cx == 0 && shift == CENTRAL && cyf < ci_y)
5167 /* Leave the pairs with i > j.
5168 * Skip half of y when i and j have the same x.
5177 for (cy = cyf_x; cy <= cyl; cy++)
5179 c0 = gridj->cxy_ind[cx*gridj->ncy+cy];
5180 c1 = gridj->cxy_ind[cx*gridj->ncy+cy+1];
5181 #ifdef NBNXN_SHIFT_BACKWARD
5182 if (gridi == gridj &&
5183 shift == CENTRAL && c0 < ci)
5190 if (gridj->c0[YY] + cy*gridj->sy > by1)
5192 d2zxy += sqr(gridj->c0[YY] + cy*gridj->sy - by1);
5194 else if (gridj->c0[YY] + (cy+1)*gridj->sy < by0)
5196 d2zxy += sqr(gridj->c0[YY] + (cy+1)*gridj->sy - by0);
5198 if (c1 > c0 && d2zxy < rl2)
5200 cs = c0 + (int)(bz1_frac*(c1 - c0));
5208 /* Find the lowest cell that can possibly
5213 (bbcz_j[cf*NNBSBB_D+1] >= bz0 ||
5214 d2xy + sqr(bbcz_j[cf*NNBSBB_D+1] - bz0) < rl2))
5219 /* Find the highest cell that can possibly
5224 (bbcz_j[cl*NNBSBB_D] <= bz1 ||
5225 d2xy + sqr(bbcz_j[cl*NNBSBB_D] - bz1) < rl2))
5230 #ifdef NBNXN_REFCODE
5232 /* Simple reference code, for debugging,
5233 * overrides the more complex code above.
5238 for (k = c0; k < c1; k++)
5240 if (box_dist2(bx0, bx1, by0, by1, bz0, bz1, bb+k) < rl2 &&
5245 if (box_dist2(bx0, bx1, by0, by1, bz0, bz1, bb+k) < rl2 &&
5256 /* We want each atom/cell pair only once,
5257 * only use cj >= ci.
5259 #ifndef NBNXN_SHIFT_BACKWARD
5262 if (shift == CENTRAL)
5271 /* For f buffer flags with simple lists */
5272 ncj_old_j = nbl->ncj;
5274 switch (nb_kernel_type)
5276 case nbnxnk4x4_PlainC:
5277 check_subcell_list_space_simple(nbl, cl-cf+1);
5279 make_cluster_list_simple(gridj,
5281 (gridi == gridj && shift == CENTRAL),
5286 #ifdef GMX_NBNXN_SIMD_4XN
5287 case nbnxnk4xN_SIMD_4xN:
5288 check_subcell_list_space_simple(nbl, ci_to_cj(na_cj_2log, cl-cf)+2);
5289 make_cluster_list_simd_4xn(gridj,
5291 (gridi == gridj && shift == CENTRAL),
5297 #ifdef GMX_NBNXN_SIMD_2XNN
5298 case nbnxnk4xN_SIMD_2xNN:
5299 check_subcell_list_space_simple(nbl, ci_to_cj(na_cj_2log, cl-cf)+2);
5300 make_cluster_list_simd_2xnn(gridj,
5302 (gridi == gridj && shift == CENTRAL),
5308 case nbnxnk8x8x8_PlainC:
5309 case nbnxnk8x8x8_CUDA:
5310 check_subcell_list_space_supersub(nbl, cl-cf+1);
5311 for (cj = cf; cj <= cl; cj++)
5313 make_cluster_list_supersub(gridi, gridj,
5315 (gridi == gridj && shift == CENTRAL && ci == cj),
5316 nbat->xstride, nbat->x,
5322 ncpcheck += cl - cf + 1;
5324 if (bFBufferFlag && nbl->ncj > ncj_old_j)
5328 cbf = nbl->cj[ncj_old_j].cj >> gridj_flag_shift;
5329 cbl = nbl->cj[nbl->ncj-1].cj >> gridj_flag_shift;
5330 for (cb = cbf; cb <= cbl; cb++)
5332 gridj_flag[cb] = 1U<<th;
5340 /* Set the exclusions for this ci list */
5343 set_ci_top_excls(nbs,
5345 shift == CENTRAL && gridi == gridj,
5348 &(nbl->ci[nbl->nci]),
5353 make_fep_list(nbs, nbat, nbl,
5354 shift == CENTRAL && gridi == gridj,
5355 &(nbl->ci[nbl->nci]),
5356 gridi, gridj, nbl_fep);
5361 set_sci_top_excls(nbs,
5363 shift == CENTRAL && gridi == gridj,
5365 &(nbl->sci[nbl->nsci]),
5370 make_fep_list_supersub(nbs, nbat, nbl,
5371 shift == CENTRAL && gridi == gridj,
5372 &(nbl->sci[nbl->nsci]),
5375 gridi, gridj, nbl_fep);
5379 /* Close this ci list */
5382 close_ci_entry_simple(nbl);
5386 close_ci_entry_supersub(nbl,
5388 progBal, min_ci_balanced,
5395 if (bFBufferFlag && nbl->ncj > ncj_old_i)
5397 work->buffer_flags.flag[(gridi->cell0+ci)>>gridi_flag_shift] = 1U<<th;
5401 work->ndistc = ndistc;
5403 nbs_cycle_stop(&work->cc[enbsCCsearch]);
5407 fprintf(debug, "number of distance checks %d\n", ndistc);
5408 fprintf(debug, "ncpcheck %s %d\n", gridi == gridj ? "local" : "non-local",
5413 print_nblist_statistics_simple(debug, nbl, nbs, rlist);
5417 print_nblist_statistics_supersub(debug, nbl, nbs, rlist);
5422 fprintf(debug, "nbl FEP list pairs: %d\n", nbl_fep->nrj);
5427 static void reduce_buffer_flags(const nbnxn_search_t nbs,
5429 const nbnxn_buffer_flags_t *dest)
5432 const unsigned int *flag;
5434 for (s = 0; s < nsrc; s++)
5436 flag = nbs->work[s].buffer_flags.flag;
5438 for (b = 0; b < dest->nflag; b++)
5440 dest->flag[b] |= flag[b];
5445 static void print_reduction_cost(const nbnxn_buffer_flags_t *flags, int nout)
5447 int nelem, nkeep, ncopy, nred, b, c, out;
5453 for (b = 0; b < flags->nflag; b++)
5455 if (flags->flag[b] == 1)
5457 /* Only flag 0 is set, no copy of reduction required */
5461 else if (flags->flag[b] > 0)
5464 for (out = 0; out < nout; out++)
5466 if (flags->flag[b] & (1U<<out))
5483 fprintf(debug, "nbnxn reduction: #flag %d #list %d elem %4.2f, keep %4.2f copy %4.2f red %4.2f\n",
5485 nelem/(double)(flags->nflag),
5486 nkeep/(double)(flags->nflag),
5487 ncopy/(double)(flags->nflag),
5488 nred/(double)(flags->nflag));
5491 /* Perform a count (linear) sort to sort the smaller lists to the end.
5492 * This avoids load imbalance on the GPU, as large lists will be
5493 * scheduled and executed first and the smaller lists later.
5494 * Load balancing between multi-processors only happens at the end
5495 * and there smaller lists lead to more effective load balancing.
5496 * The sorting is done on the cj4 count, not on the actual pair counts.
5497 * Not only does this make the sort faster, but it also results in
5498 * better load balancing than using a list sorted on exact load.
5499 * This function swaps the pointer in the pair list to avoid a copy operation.
5501 static void sort_sci(nbnxn_pairlist_t *nbl)
5503 nbnxn_list_work_t *work;
5504 int m, i, s, s0, s1;
5505 nbnxn_sci_t *sci_sort;
5507 if (nbl->ncj4 <= nbl->nsci)
5509 /* nsci = 0 or all sci have size 1, sorting won't change the order */
5515 /* We will distinguish differences up to double the average */
5516 m = (2*nbl->ncj4)/nbl->nsci;
5518 if (m + 1 > work->sort_nalloc)
5520 work->sort_nalloc = over_alloc_large(m + 1);
5521 srenew(work->sort, work->sort_nalloc);
5524 if (work->sci_sort_nalloc != nbl->sci_nalloc)
5526 work->sci_sort_nalloc = nbl->sci_nalloc;
5527 nbnxn_realloc_void((void **)&work->sci_sort,
5529 work->sci_sort_nalloc*sizeof(*work->sci_sort),
5530 nbl->alloc, nbl->free);
5533 /* Count the entries of each size */
5534 for (i = 0; i <= m; i++)
5538 for (s = 0; s < nbl->nsci; s++)
5540 i = min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start);
5543 /* Calculate the offset for each count */
5546 for (i = m - 1; i >= 0; i--)
5549 work->sort[i] = work->sort[i + 1] + s0;
5553 /* Sort entries directly into place */
5554 sci_sort = work->sci_sort;
5555 for (s = 0; s < nbl->nsci; s++)
5557 i = min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start);
5558 sci_sort[work->sort[i]++] = nbl->sci[s];
5561 /* Swap the sci pointers so we use the new, sorted list */
5562 work->sci_sort = nbl->sci;
5563 nbl->sci = sci_sort;
5566 /* Make a local or non-local pair-list, depending on iloc */
5567 void nbnxn_make_pairlist(const nbnxn_search_t nbs,
5568 nbnxn_atomdata_t *nbat,
5569 const t_blocka *excl,
5571 int min_ci_balanced,
5572 nbnxn_pairlist_set_t *nbl_list,
5577 nbnxn_grid_t *gridi, *gridj;
5579 int nzi, zi, zj0, zj1, zj;
5583 nbnxn_pairlist_t **nbl;
5585 gmx_bool CombineNBLists;
5587 int np_tot, np_noq, np_hlj, nap;
5589 /* Check if we are running hybrid GPU + CPU nbnxn mode */
5590 bGPUCPU = (!nbs->grid[0].bSimple && nbl_list->bSimple);
5592 nnbl = nbl_list->nnbl;
5593 nbl = nbl_list->nbl;
5594 CombineNBLists = nbl_list->bCombined;
5598 fprintf(debug, "ns making %d nblists\n", nnbl);
5601 nbat->bUseBufferFlags = (nbat->nout > 1);
5602 /* We should re-init the flags before making the first list */
5603 if (nbat->bUseBufferFlags && (LOCAL_I(iloc) || bGPUCPU))
5605 init_buffer_flags(&nbat->buffer_flags, nbat->natoms);
5608 if (nbl_list->bSimple)
5610 switch (nb_kernel_type)
5612 #ifdef GMX_NBNXN_SIMD_4XN
5613 case nbnxnk4xN_SIMD_4xN:
5614 nbs->icell_set_x = icell_set_x_simd_4xn;
5617 #ifdef GMX_NBNXN_SIMD_2XNN
5618 case nbnxnk4xN_SIMD_2xNN:
5619 nbs->icell_set_x = icell_set_x_simd_2xnn;
5623 nbs->icell_set_x = icell_set_x_simple;
5629 #ifdef NBNXN_SEARCH_BB_SIMD4
5630 nbs->icell_set_x = icell_set_x_supersub_simd4;
5632 nbs->icell_set_x = icell_set_x_supersub;
5638 /* Only zone (grid) 0 vs 0 */
5645 nzi = nbs->zones->nizone;
5648 if (!nbl_list->bSimple && min_ci_balanced > 0)
5650 nsubpair_max = get_nsubpair_max(nbs, iloc, rlist, min_ci_balanced);
5657 /* Clear all pair-lists */
5658 for (th = 0; th < nnbl; th++)
5660 clear_pairlist(nbl[th]);
5664 clear_pairlist_fep(nbl_list->nbl_fep[th]);
5668 for (zi = 0; zi < nzi; zi++)
5670 gridi = &nbs->grid[zi];
5672 if (NONLOCAL_I(iloc))
5674 zj0 = nbs->zones->izone[zi].j0;
5675 zj1 = nbs->zones->izone[zi].j1;
5681 for (zj = zj0; zj < zj1; zj++)
5683 gridj = &nbs->grid[zj];
5687 fprintf(debug, "ns search grid %d vs %d\n", zi, zj);
5690 nbs_cycle_start(&nbs->cc[enbsCCsearch]);
5692 if (nbl[0]->bSimple && !gridi->bSimple)
5694 /* Hybrid list, determine blocking later */
5699 ci_block = get_ci_block_size(gridi, nbs->DomDec, nnbl);
5702 /* With GPU: generate progressively smaller lists for
5703 * load balancing for local only or non-local with 2 zones.
5705 progBal = (LOCAL_I(iloc) || nbs->zones->n <= 2);
5707 #pragma omp parallel for num_threads(nnbl) schedule(static)
5708 for (th = 0; th < nnbl; th++)
5710 /* Re-init the thread-local work flag data before making
5711 * the first list (not an elegant conditional).
5713 if (nbat->bUseBufferFlags && ((zi == 0 && zj == 0) ||
5714 (bGPUCPU && zi == 0 && zj == 1)))
5716 init_buffer_flags(&nbs->work[th].buffer_flags, nbat->natoms);
5719 if (CombineNBLists && th > 0)
5721 clear_pairlist(nbl[th]);
5724 /* Divide the i super cell equally over the nblists */
5725 nbnxn_make_pairlist_part(nbs, gridi, gridj,
5726 &nbs->work[th], nbat, excl,
5730 nbat->bUseBufferFlags,
5732 progBal, min_ci_balanced,
5735 nbl_list->nbl_fep[th]);
5737 nbs_cycle_stop(&nbs->cc[enbsCCsearch]);
5742 for (th = 0; th < nnbl; th++)
5744 inc_nrnb(nrnb, eNR_NBNXN_DIST2, nbs->work[th].ndistc);
5746 if (nbl_list->bSimple)
5748 np_tot += nbl[th]->ncj;
5749 np_noq += nbl[th]->work->ncj_noq;
5750 np_hlj += nbl[th]->work->ncj_hlj;
5754 /* This count ignores potential subsequent pair pruning */
5755 np_tot += nbl[th]->nci_tot;
5758 nap = nbl[0]->na_ci*nbl[0]->na_cj;
5759 nbl_list->natpair_ljq = (np_tot - np_noq)*nap - np_hlj*nap/2;
5760 nbl_list->natpair_lj = np_noq*nap;
5761 nbl_list->natpair_q = np_hlj*nap/2;
5763 if (CombineNBLists && nnbl > 1)
5765 nbs_cycle_start(&nbs->cc[enbsCCcombine]);
5767 combine_nblists(nnbl-1, nbl+1, nbl[0]);
5769 nbs_cycle_stop(&nbs->cc[enbsCCcombine]);
5774 if (!nbl_list->bSimple)
5776 /* Sort the entries on size, large ones first */
5777 if (CombineNBLists || nnbl == 1)
5783 #pragma omp parallel for num_threads(nnbl) schedule(static)
5784 for (th = 0; th < nnbl; th++)
5791 if (nbat->bUseBufferFlags)
5793 reduce_buffer_flags(nbs, nnbl, &nbat->buffer_flags);
5798 /* Balance the free-energy lists over all the threads */
5799 balance_fep_lists(nbs, nbl_list);
5802 /* Special performance logging stuff (env.var. GMX_NBNXN_CYCLE) */
5805 nbs->search_count++;
5807 if (nbs->print_cycles &&
5808 (!nbs->DomDec || (nbs->DomDec && !LOCAL_I(iloc))) &&
5809 nbs->search_count % 100 == 0)
5811 nbs_cycle_print(stderr, nbs);
5814 if (debug && (CombineNBLists && nnbl > 1))
5816 if (nbl[0]->bSimple)
5818 print_nblist_statistics_simple(debug, nbl[0], nbs, rlist);
5822 print_nblist_statistics_supersub(debug, nbl[0], nbs, rlist);
5830 if (nbl[0]->bSimple)
5832 print_nblist_ci_cj(debug, nbl[0]);
5836 print_nblist_sci_cj(debug, nbl[0]);
5840 if (nbat->bUseBufferFlags)
5842 print_reduction_cost(&nbat->buffer_flags, nnbl);