2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
38 #include "nbnxn_search.h"
48 #include "gromacs/domdec/domdec_struct.h"
49 #include "gromacs/gmxlib/nrnb.h"
50 #include "gromacs/math/functions.h"
51 #include "gromacs/math/utilities.h"
52 #include "gromacs/math/vec.h"
53 #include "gromacs/mdlib/gmx_omp_nthreads.h"
54 #include "gromacs/mdlib/nb_verlet.h"
55 #include "gromacs/mdlib/nbnxn_atomdata.h"
56 #include "gromacs/mdlib/nbnxn_consts.h"
57 #include "gromacs/mdlib/nbnxn_grid.h"
58 #include "gromacs/mdlib/nbnxn_internal.h"
59 #include "gromacs/mdlib/nbnxn_simd.h"
60 #include "gromacs/mdlib/nbnxn_util.h"
61 #include "gromacs/mdlib/ns.h"
62 #include "gromacs/mdtypes/group.h"
63 #include "gromacs/mdtypes/md_enums.h"
64 #include "gromacs/pbcutil/ishift.h"
65 #include "gromacs/pbcutil/pbc.h"
66 #include "gromacs/simd/simd.h"
67 #include "gromacs/simd/vector_operations.h"
68 #include "gromacs/topology/block.h"
69 #include "gromacs/utility/exceptions.h"
70 #include "gromacs/utility/fatalerror.h"
71 #include "gromacs/utility/gmxomp.h"
72 #include "gromacs/utility/smalloc.h"
74 using namespace gmx; // TODO: Remove when this file is moved into gmx namespace
77 /* We shift the i-particles backward for PBC.
78 * This leads to more conditionals than shifting forward.
79 * We do this to get more balanced pair lists.
81 constexpr bool c_pbcShiftBackward = true;
84 static void nbs_cycle_clear(nbnxn_cycle_t *cc)
86 for (int i = 0; i < enbsCCnr; i++)
93 static double Mcyc_av(const nbnxn_cycle_t *cc)
95 return static_cast<double>(cc->c)*1e-6/cc->count;
98 static void nbs_cycle_print(FILE *fp, const nbnxn_search *nbs)
101 fprintf(fp, "ns %4d grid %4.1f search %4.1f red.f %5.3f",
102 nbs->cc[enbsCCgrid].count,
103 Mcyc_av(&nbs->cc[enbsCCgrid]),
104 Mcyc_av(&nbs->cc[enbsCCsearch]),
105 Mcyc_av(&nbs->cc[enbsCCreducef]));
107 if (nbs->work.size() > 1)
109 if (nbs->cc[enbsCCcombine].count > 0)
111 fprintf(fp, " comb %5.2f",
112 Mcyc_av(&nbs->cc[enbsCCcombine]));
114 fprintf(fp, " s. th");
115 for (const nbnxn_search_work_t &work : nbs->work)
117 fprintf(fp, " %4.1f",
118 Mcyc_av(&work.cc[enbsCCsearch]));
124 /* Layout for the nonbonded NxN pair lists */
125 enum class NbnxnLayout
127 NoSimd4x4, // i-cluster size 4, j-cluster size 4
128 Simd4xN, // i-cluster size 4, j-cluster size SIMD width
129 Simd2xNN, // i-cluster size 4, j-cluster size half SIMD width
130 Gpu8x8x8 // i-cluster size 8, j-cluster size 8 + super-clustering
134 /* Returns the j-cluster size */
135 template <NbnxnLayout layout>
136 static constexpr int jClusterSize()
138 static_assert(layout == NbnxnLayout::NoSimd4x4 || layout == NbnxnLayout::Simd4xN || layout == NbnxnLayout::Simd2xNN, "Currently jClusterSize only supports CPU layouts");
140 return layout == NbnxnLayout::Simd4xN ? GMX_SIMD_REAL_WIDTH : (layout == NbnxnLayout::Simd2xNN ? GMX_SIMD_REAL_WIDTH/2 : NBNXN_CPU_CLUSTER_I_SIZE);
143 /* Returns the j-cluster index given the i-cluster index */
144 template <int jClusterSize>
145 static inline int cjFromCi(int ci)
147 static_assert(jClusterSize == NBNXN_CPU_CLUSTER_I_SIZE/2 || jClusterSize == NBNXN_CPU_CLUSTER_I_SIZE || jClusterSize == NBNXN_CPU_CLUSTER_I_SIZE*2, "Only j-cluster sizes 2, 4 and 8 are currently implemented");
149 if (jClusterSize == NBNXN_CPU_CLUSTER_I_SIZE/2)
153 else if (jClusterSize == NBNXN_CPU_CLUSTER_I_SIZE)
163 /* Returns the j-cluster index given the i-cluster index */
164 template <NbnxnLayout layout>
165 static inline int cjFromCi(int ci)
167 constexpr int clusterSize = jClusterSize<layout>();
169 return cjFromCi<clusterSize>(ci);
172 /* Returns the nbnxn coordinate data index given the i-cluster index */
173 template <NbnxnLayout layout>
174 static inline int xIndexFromCi(int ci)
176 constexpr int clusterSize = jClusterSize<layout>();
178 static_assert(clusterSize == NBNXN_CPU_CLUSTER_I_SIZE/2 || clusterSize == NBNXN_CPU_CLUSTER_I_SIZE || clusterSize == NBNXN_CPU_CLUSTER_I_SIZE*2, "Only j-cluster sizes 2, 4 and 8 are currently implemented");
180 if (clusterSize <= NBNXN_CPU_CLUSTER_I_SIZE)
182 /* Coordinates are stored packed in groups of 4 */
187 /* Coordinates packed in 8, i-cluster size is half the packing width */
188 return (ci >> 1)*STRIDE_P8 + (ci & 1)*(c_packX8 >> 1);
192 /* Returns the nbnxn coordinate data index given the j-cluster index */
193 template <NbnxnLayout layout>
194 static inline int xIndexFromCj(int cj)
196 constexpr int clusterSize = jClusterSize<layout>();
198 static_assert(clusterSize == NBNXN_CPU_CLUSTER_I_SIZE/2 || clusterSize == NBNXN_CPU_CLUSTER_I_SIZE || clusterSize == NBNXN_CPU_CLUSTER_I_SIZE*2, "Only j-cluster sizes 2, 4 and 8 are currently implemented");
200 if (clusterSize == NBNXN_CPU_CLUSTER_I_SIZE/2)
202 /* Coordinates are stored packed in groups of 4 */
203 return (cj >> 1)*STRIDE_P4 + (cj & 1)*(c_packX4 >> 1);
205 else if (clusterSize == NBNXN_CPU_CLUSTER_I_SIZE)
207 /* Coordinates are stored packed in groups of 4 */
212 /* Coordinates are stored packed in groups of 8 */
218 gmx_bool nbnxn_kernel_pairlist_simple(int nb_kernel_type)
220 if (nb_kernel_type == nbnxnkNotSet)
222 gmx_fatal(FARGS, "Non-bonded kernel type not set for Verlet-style pair-list.");
225 switch (nb_kernel_type)
227 case nbnxnk8x8x8_GPU:
228 case nbnxnk8x8x8_PlainC:
231 case nbnxnk4x4_PlainC:
232 case nbnxnk4xN_SIMD_4xN:
233 case nbnxnk4xN_SIMD_2xNN:
237 gmx_incons("Invalid nonbonded kernel type passed!");
242 /* Initializes a single nbnxn_pairlist_t data structure */
243 static void nbnxn_init_pairlist_fep(t_nblist *nl)
245 nl->type = GMX_NBLIST_INTERACTION_FREE_ENERGY;
246 nl->igeometry = GMX_NBLIST_GEOMETRY_PARTICLE_PARTICLE;
247 /* The interaction functions are set in the free energy kernel fuction */
260 nl->jindex = nullptr;
262 nl->excl_fep = nullptr;
266 static void free_nblist(t_nblist *nl)
276 nbnxn_search_work_t::nbnxn_search_work_t() :
279 buffer_flags({0, nullptr, 0}),
281 nbl_fep(new t_nblist),
284 nbnxn_init_pairlist_fep(nbl_fep.get());
289 nbnxn_search_work_t::~nbnxn_search_work_t()
291 sfree(buffer_flags.flag);
293 free_nblist(nbl_fep.get());
296 nbnxn_search::nbnxn_search(const ivec *n_dd_cells,
297 const gmx_domdec_zones_t *zones,
301 ePBC(epbcNONE), // The correct value will be set during the gridding
308 // The correct value will be set during the gridding
312 DomDec = n_dd_cells != nullptr;
315 for (int d = 0; d < DIM; d++)
317 if ((*n_dd_cells)[d] > 1)
320 /* Each grid matches a DD zone */
326 grid.resize(numGrids);
328 /* Initialize detailed nbsearch cycle counting */
329 print_cycles = (getenv("GMX_NBNXN_CYCLE") != nullptr);
333 nbnxn_search *nbnxn_init_search(const ivec *n_dd_cells,
334 const gmx_domdec_zones_t *zones,
338 return new nbnxn_search(n_dd_cells, zones, bFEP, nthread_max);
341 static void init_buffer_flags(nbnxn_buffer_flags_t *flags,
344 flags->nflag = (natoms + NBNXN_BUFFERFLAG_SIZE - 1)/NBNXN_BUFFERFLAG_SIZE;
345 if (flags->nflag > flags->flag_nalloc)
347 flags->flag_nalloc = over_alloc_large(flags->nflag);
348 srenew(flags->flag, flags->flag_nalloc);
350 for (int b = 0; b < flags->nflag; b++)
352 bitmask_clear(&(flags->flag[b]));
356 /* Returns the pair-list cutoff between a bounding box and a grid cell given an atom-to-atom pair-list cutoff
358 * Given a cutoff distance between atoms, this functions returns the cutoff
359 * distance2 between a bounding box of a group of atoms and a grid cell.
360 * Since atoms can be geometrically outside of the cell they have been
361 * assigned to (when atom groups instead of individual atoms are assigned
362 * to cells), this distance returned can be larger than the input.
364 static real listRangeForBoundingBoxToGridCell(real rlist,
365 const nbnxn_grid_t &grid)
367 return rlist + grid.maxAtomGroupRadius;
370 /* Returns the pair-list cutoff between a grid cells given an atom-to-atom pair-list cutoff
372 * Given a cutoff distance between atoms, this functions returns the cutoff
373 * distance2 between two grid cells.
374 * Since atoms can be geometrically outside of the cell they have been
375 * assigned to (when atom groups instead of individual atoms are assigned
376 * to cells), this distance returned can be larger than the input.
378 static real listRangeForGridCellToGridCell(real rlist,
379 const nbnxn_grid_t &gridi,
380 const nbnxn_grid_t &gridj)
382 return rlist + gridi.maxAtomGroupRadius + gridj.maxAtomGroupRadius;
385 /* Determines the cell range along one dimension that
386 * the bounding box b0 - b1 sees.
389 static void get_cell_range(real b0, real b1,
390 const nbnxn_grid_t &gridj,
391 real d2, real rlist, int *cf, int *cl)
393 real listRangeBBToCell2 = gmx::square(listRangeForBoundingBoxToGridCell(rlist, gridj));
394 real distanceInCells = (b0 - gridj.c0[dim])*gridj.invCellSize[dim];
395 *cf = std::max(static_cast<int>(distanceInCells), 0);
398 d2 + gmx::square((b0 - gridj.c0[dim]) - (*cf - 1 + 1)*gridj.cellSize[dim]) < listRangeBBToCell2)
403 *cl = std::min(static_cast<int>((b1 - gridj.c0[dim])*gridj.invCellSize[dim]), gridj.numCells[dim] - 1);
404 while (*cl < gridj.numCells[dim] - 1 &&
405 d2 + gmx::square((*cl + 1)*gridj.cellSize[dim] - (b1 - gridj.c0[dim])) < listRangeBBToCell2)
411 /* Reference code calculating the distance^2 between two bounding boxes */
413 static float box_dist2(float bx0, float bx1, float by0,
414 float by1, float bz0, float bz1,
415 const nbnxn_bb_t *bb)
418 float dl, dh, dm, dm0;
422 dl = bx0 - bb->upper[BB_X];
423 dh = bb->lower[BB_X] - bx1;
424 dm = std::max(dl, dh);
425 dm0 = std::max(dm, 0.0f);
428 dl = by0 - bb->upper[BB_Y];
429 dh = bb->lower[BB_Y] - by1;
430 dm = std::max(dl, dh);
431 dm0 = std::max(dm, 0.0f);
434 dl = bz0 - bb->upper[BB_Z];
435 dh = bb->lower[BB_Z] - bz1;
436 dm = std::max(dl, dh);
437 dm0 = std::max(dm, 0.0f);
444 /* Plain C code calculating the distance^2 between two bounding boxes */
445 static float subc_bb_dist2(int si,
446 const nbnxn_bb_t *bb_i_ci,
448 gmx::ArrayRef<const nbnxn_bb_t> bb_j_all)
450 const nbnxn_bb_t *bb_i = bb_i_ci + si;
451 const nbnxn_bb_t *bb_j = bb_j_all.data() + csj;
454 float dl, dh, dm, dm0;
456 dl = bb_i->lower[BB_X] - bb_j->upper[BB_X];
457 dh = bb_j->lower[BB_X] - bb_i->upper[BB_X];
458 dm = std::max(dl, dh);
459 dm0 = std::max(dm, 0.0f);
462 dl = bb_i->lower[BB_Y] - bb_j->upper[BB_Y];
463 dh = bb_j->lower[BB_Y] - bb_i->upper[BB_Y];
464 dm = std::max(dl, dh);
465 dm0 = std::max(dm, 0.0f);
468 dl = bb_i->lower[BB_Z] - bb_j->upper[BB_Z];
469 dh = bb_j->lower[BB_Z] - bb_i->upper[BB_Z];
470 dm = std::max(dl, dh);
471 dm0 = std::max(dm, 0.0f);
477 #if NBNXN_SEARCH_BB_SIMD4
479 /* 4-wide SIMD code for bb distance for bb format xyz0 */
480 static float subc_bb_dist2_simd4(int si,
481 const nbnxn_bb_t *bb_i_ci,
483 gmx::ArrayRef<const nbnxn_bb_t> bb_j_all)
485 // TODO: During SIMDv2 transition only some archs use namespace (remove when done)
488 Simd4Float bb_i_S0, bb_i_S1;
489 Simd4Float bb_j_S0, bb_j_S1;
495 bb_i_S0 = load4(&bb_i_ci[si].lower[0]);
496 bb_i_S1 = load4(&bb_i_ci[si].upper[0]);
497 bb_j_S0 = load4(&bb_j_all[csj].lower[0]);
498 bb_j_S1 = load4(&bb_j_all[csj].upper[0]);
500 dl_S = bb_i_S0 - bb_j_S1;
501 dh_S = bb_j_S0 - bb_i_S1;
503 dm_S = max(dl_S, dh_S);
504 dm0_S = max(dm_S, simd4SetZeroF());
506 return dotProduct(dm0_S, dm0_S);
509 /* Calculate bb bounding distances of bb_i[si,...,si+3] and store them in d2 */
510 #define SUBC_BB_DIST2_SIMD4_XXXX_INNER(si, bb_i, d2) \
514 Simd4Float dx_0, dy_0, dz_0; \
515 Simd4Float dx_1, dy_1, dz_1; \
517 Simd4Float mx, my, mz; \
518 Simd4Float m0x, m0y, m0z; \
520 Simd4Float d2x, d2y, d2z; \
521 Simd4Float d2s, d2t; \
523 shi = (si)*NNBSBB_D*DIM; \
525 xi_l = load4((bb_i)+shi+0*STRIDE_PBB); \
526 yi_l = load4((bb_i)+shi+1*STRIDE_PBB); \
527 zi_l = load4((bb_i)+shi+2*STRIDE_PBB); \
528 xi_h = load4((bb_i)+shi+3*STRIDE_PBB); \
529 yi_h = load4((bb_i)+shi+4*STRIDE_PBB); \
530 zi_h = load4((bb_i)+shi+5*STRIDE_PBB); \
532 dx_0 = xi_l - xj_h; \
533 dy_0 = yi_l - yj_h; \
534 dz_0 = zi_l - zj_h; \
536 dx_1 = xj_l - xi_h; \
537 dy_1 = yj_l - yi_h; \
538 dz_1 = zj_l - zi_h; \
540 mx = max(dx_0, dx_1); \
541 my = max(dy_0, dy_1); \
542 mz = max(dz_0, dz_1); \
544 m0x = max(mx, zero); \
545 m0y = max(my, zero); \
546 m0z = max(mz, zero); \
555 store4((d2)+(si), d2t); \
558 /* 4-wide SIMD code for nsi bb distances for bb format xxxxyyyyzzzz */
559 static void subc_bb_dist2_simd4_xxxx(const float *bb_j,
560 int nsi, const float *bb_i,
563 // TODO: During SIMDv2 transition only some archs use namespace (remove when done)
566 Simd4Float xj_l, yj_l, zj_l;
567 Simd4Float xj_h, yj_h, zj_h;
568 Simd4Float xi_l, yi_l, zi_l;
569 Simd4Float xi_h, yi_h, zi_h;
575 xj_l = Simd4Float(bb_j[0*STRIDE_PBB]);
576 yj_l = Simd4Float(bb_j[1*STRIDE_PBB]);
577 zj_l = Simd4Float(bb_j[2*STRIDE_PBB]);
578 xj_h = Simd4Float(bb_j[3*STRIDE_PBB]);
579 yj_h = Simd4Float(bb_j[4*STRIDE_PBB]);
580 zj_h = Simd4Float(bb_j[5*STRIDE_PBB]);
582 /* Here we "loop" over si (0,STRIDE_PBB) from 0 to nsi with step STRIDE_PBB.
583 * But as we know the number of iterations is 1 or 2, we unroll manually.
585 SUBC_BB_DIST2_SIMD4_XXXX_INNER(0, bb_i, d2);
586 if (STRIDE_PBB < nsi)
588 SUBC_BB_DIST2_SIMD4_XXXX_INNER(STRIDE_PBB, bb_i, d2);
592 #endif /* NBNXN_SEARCH_BB_SIMD4 */
595 /* Returns if any atom pair from two clusters is within distance sqrt(rlist2) */
596 static inline gmx_bool
597 clusterpair_in_range(const nbnxn_list_work_t *work,
599 int csj, int stride, const real *x_j,
602 #if !GMX_SIMD4_HAVE_REAL
605 * All coordinates are stored as xyzxyz...
608 const real *x_i = work->x_ci;
610 for (int i = 0; i < c_nbnxnGpuClusterSize; i++)
612 int i0 = (si*c_nbnxnGpuClusterSize + i)*DIM;
613 for (int j = 0; j < c_nbnxnGpuClusterSize; j++)
615 int j0 = (csj*c_nbnxnGpuClusterSize + j)*stride;
617 real d2 = gmx::square(x_i[i0 ] - x_j[j0 ]) + gmx::square(x_i[i0+1] - x_j[j0+1]) + gmx::square(x_i[i0+2] - x_j[j0+2]);
628 #else /* !GMX_SIMD4_HAVE_REAL */
630 /* 4-wide SIMD version.
631 * The coordinates x_i are stored as xxxxyyyy..., x_j is stored xyzxyz...
632 * Using 8-wide AVX(2) is not faster on Intel Sandy Bridge and Haswell.
634 static_assert(c_nbnxnGpuClusterSize == 8 || c_nbnxnGpuClusterSize == 4,
635 "A cluster is hard-coded to 4/8 atoms.");
637 Simd4Real rc2_S = Simd4Real(rlist2);
639 const real *x_i = work->x_ci_simd;
641 int dim_stride = c_nbnxnGpuClusterSize*DIM;
642 Simd4Real ix_S0 = load4(x_i + si*dim_stride + 0*GMX_SIMD4_WIDTH);
643 Simd4Real iy_S0 = load4(x_i + si*dim_stride + 1*GMX_SIMD4_WIDTH);
644 Simd4Real iz_S0 = load4(x_i + si*dim_stride + 2*GMX_SIMD4_WIDTH);
646 Simd4Real ix_S1, iy_S1, iz_S1;
647 if (c_nbnxnGpuClusterSize == 8)
649 ix_S1 = load4(x_i + si*dim_stride + 3*GMX_SIMD4_WIDTH);
650 iy_S1 = load4(x_i + si*dim_stride + 4*GMX_SIMD4_WIDTH);
651 iz_S1 = load4(x_i + si*dim_stride + 5*GMX_SIMD4_WIDTH);
653 /* We loop from the outer to the inner particles to maximize
654 * the chance that we find a pair in range quickly and return.
656 int j0 = csj*c_nbnxnGpuClusterSize;
657 int j1 = j0 + c_nbnxnGpuClusterSize - 1;
660 Simd4Real jx0_S, jy0_S, jz0_S;
661 Simd4Real jx1_S, jy1_S, jz1_S;
663 Simd4Real dx_S0, dy_S0, dz_S0;
664 Simd4Real dx_S1, dy_S1, dz_S1;
665 Simd4Real dx_S2, dy_S2, dz_S2;
666 Simd4Real dx_S3, dy_S3, dz_S3;
677 Simd4Bool wco_any_S01, wco_any_S23, wco_any_S;
679 jx0_S = Simd4Real(x_j[j0*stride+0]);
680 jy0_S = Simd4Real(x_j[j0*stride+1]);
681 jz0_S = Simd4Real(x_j[j0*stride+2]);
683 jx1_S = Simd4Real(x_j[j1*stride+0]);
684 jy1_S = Simd4Real(x_j[j1*stride+1]);
685 jz1_S = Simd4Real(x_j[j1*stride+2]);
687 /* Calculate distance */
688 dx_S0 = ix_S0 - jx0_S;
689 dy_S0 = iy_S0 - jy0_S;
690 dz_S0 = iz_S0 - jz0_S;
691 dx_S2 = ix_S0 - jx1_S;
692 dy_S2 = iy_S0 - jy1_S;
693 dz_S2 = iz_S0 - jz1_S;
694 if (c_nbnxnGpuClusterSize == 8)
696 dx_S1 = ix_S1 - jx0_S;
697 dy_S1 = iy_S1 - jy0_S;
698 dz_S1 = iz_S1 - jz0_S;
699 dx_S3 = ix_S1 - jx1_S;
700 dy_S3 = iy_S1 - jy1_S;
701 dz_S3 = iz_S1 - jz1_S;
704 /* rsq = dx*dx+dy*dy+dz*dz */
705 rsq_S0 = norm2(dx_S0, dy_S0, dz_S0);
706 rsq_S2 = norm2(dx_S2, dy_S2, dz_S2);
707 if (c_nbnxnGpuClusterSize == 8)
709 rsq_S1 = norm2(dx_S1, dy_S1, dz_S1);
710 rsq_S3 = norm2(dx_S3, dy_S3, dz_S3);
713 wco_S0 = (rsq_S0 < rc2_S);
714 wco_S2 = (rsq_S2 < rc2_S);
715 if (c_nbnxnGpuClusterSize == 8)
717 wco_S1 = (rsq_S1 < rc2_S);
718 wco_S3 = (rsq_S3 < rc2_S);
720 if (c_nbnxnGpuClusterSize == 8)
722 wco_any_S01 = wco_S0 || wco_S1;
723 wco_any_S23 = wco_S2 || wco_S3;
724 wco_any_S = wco_any_S01 || wco_any_S23;
728 wco_any_S = wco_S0 || wco_S2;
731 if (anyTrue(wco_any_S))
742 #endif /* !GMX_SIMD4_HAVE_REAL */
745 /* Returns the j-cluster index for index cjIndex in a cj list */
746 static inline int nblCj(const nbnxn_cj_t *cjList, int cjIndex)
748 return cjList[cjIndex].cj;
751 /* Returns the j-cluster index for index cjIndex in a cj4 list */
752 static inline int nblCj(const nbnxn_cj4_t *cj4List, int cjIndex)
754 return cj4List[cjIndex/c_nbnxnGpuJgroupSize].cj[cjIndex & (c_nbnxnGpuJgroupSize - 1)];
757 /* Returns the i-interaction mask of the j sub-cell for index cj_ind */
758 static unsigned int nbl_imask0(const nbnxn_pairlist_t *nbl, int cj_ind)
760 return nbl->cj4[cj_ind/c_nbnxnGpuJgroupSize].imei[0].imask;
763 /* Ensures there is enough space for extra extra exclusion masks */
764 static void check_excl_space(nbnxn_pairlist_t *nbl, int extra)
766 if (nbl->nexcl+extra > nbl->excl_nalloc)
768 nbl->excl_nalloc = over_alloc_small(nbl->nexcl+extra);
769 nbnxn_realloc_void(reinterpret_cast<void **>(&nbl->excl),
770 nbl->nexcl*sizeof(*nbl->excl),
771 nbl->excl_nalloc*sizeof(*nbl->excl),
772 nbl->alloc, nbl->free);
776 /* Ensures there is enough space for maxNumExtraClusters extra j-clusters in the list */
777 static void check_cell_list_space_simple(nbnxn_pairlist_t *nbl,
778 int maxNumExtraClusters)
782 cj_max = nbl->ncj + maxNumExtraClusters;
784 if (cj_max > nbl->cj_nalloc)
786 nbl->cj_nalloc = over_alloc_small(cj_max);
787 nbnxn_realloc_void(reinterpret_cast<void **>(&nbl->cj),
788 nbl->ncj*sizeof(*nbl->cj),
789 nbl->cj_nalloc*sizeof(*nbl->cj),
790 nbl->alloc, nbl->free);
792 nbnxn_realloc_void(reinterpret_cast<void **>(&nbl->cjOuter),
793 nbl->ncj*sizeof(*nbl->cjOuter),
794 nbl->cj_nalloc*sizeof(*nbl->cjOuter),
795 nbl->alloc, nbl->free);
799 /* Ensures there is enough space for ncell extra j-clusters in the list */
800 static void check_cell_list_space_supersub(nbnxn_pairlist_t *nbl,
805 /* We can have maximally nsupercell*c_gpuNumClusterPerCell sj lists */
806 /* We can store 4 j-subcell - i-supercell pairs in one struct.
807 * since we round down, we need one extra entry.
809 ncj4_max = ((nbl->work->cj_ind + ncell*c_gpuNumClusterPerCell + c_nbnxnGpuJgroupSize - 1)/c_nbnxnGpuJgroupSize);
811 if (ncj4_max > nbl->cj4_nalloc)
813 nbl->cj4_nalloc = over_alloc_small(ncj4_max);
814 nbnxn_realloc_void(reinterpret_cast<void **>(&nbl->cj4),
815 nbl->work->cj4_init*sizeof(*nbl->cj4),
816 nbl->cj4_nalloc*sizeof(*nbl->cj4),
817 nbl->alloc, nbl->free);
820 if (ncj4_max > nbl->work->cj4_init)
822 for (int j4 = nbl->work->cj4_init; j4 < ncj4_max; j4++)
824 /* No i-subcells and no excl's in the list initially */
825 for (w = 0; w < c_nbnxnGpuClusterpairSplit; w++)
827 nbl->cj4[j4].imei[w].imask = 0U;
828 nbl->cj4[j4].imei[w].excl_ind = 0;
832 nbl->work->cj4_init = ncj4_max;
836 /* Set all excl masks for one GPU warp no exclusions */
837 static void set_no_excls(nbnxn_excl_t *excl)
839 for (int t = 0; t < c_nbnxnGpuExclSize; t++)
841 /* Turn all interaction bits on */
842 excl->pair[t] = NBNXN_INTERACTION_MASK_ALL;
846 /* Initializes a single nbnxn_pairlist_t data structure */
847 static void nbnxn_init_pairlist(nbnxn_pairlist_t *nbl,
849 nbnxn_alloc_t *alloc,
852 if (alloc == nullptr)
854 nbl->alloc = nbnxn_alloc_aligned;
862 nbl->free = nbnxn_free_aligned;
869 nbl->bSimple = bSimple;
884 /* We need one element extra in sj, so alloc initially with 1 */
891 GMX_ASSERT(c_nbnxnGpuNumClusterPerSupercluster == c_gpuNumClusterPerCell, "The search code assumes that the a super-cluster matches a search grid cell");
893 GMX_ASSERT(sizeof(nbl->cj4[0].imei[0].imask)*8 >= c_nbnxnGpuJgroupSize*c_gpuNumClusterPerCell, "The i super-cluster cluster interaction mask does not contain a sufficient number of bits");
894 GMX_ASSERT(sizeof(nbl->excl[0])*8 >= c_nbnxnGpuJgroupSize*c_gpuNumClusterPerCell, "The GPU exclusion mask does not contain a sufficient number of bits");
897 nbl->excl_nalloc = 0;
899 check_excl_space(nbl, 1);
901 set_no_excls(&nbl->excl[0]);
907 snew_aligned(nbl->work->bb_ci, 1, NBNXN_SEARCH_BB_MEM_ALIGN);
912 snew_aligned(nbl->work->pbb_ci, c_gpuNumClusterPerCell/STRIDE_PBB*NNBSBB_XXXX, NBNXN_SEARCH_BB_MEM_ALIGN);
914 snew_aligned(nbl->work->bb_ci, c_gpuNumClusterPerCell, NBNXN_SEARCH_BB_MEM_ALIGN);
917 int gpu_clusterpair_nc = c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize*DIM;
918 snew(nbl->work->x_ci, gpu_clusterpair_nc);
920 snew_aligned(nbl->work->x_ci_simd,
921 std::max(NBNXN_CPU_CLUSTER_I_SIZE*DIM*GMX_SIMD_REAL_WIDTH,
923 GMX_SIMD_REAL_WIDTH);
925 snew_aligned(nbl->work->d2, c_gpuNumClusterPerCell, NBNXN_SEARCH_BB_MEM_ALIGN);
927 nbl->work->sort = nullptr;
928 nbl->work->sort_nalloc = 0;
929 nbl->work->sci_sort = nullptr;
930 nbl->work->sci_sort_nalloc = 0;
933 void nbnxn_init_pairlist_set(nbnxn_pairlist_set_t *nbl_list,
934 gmx_bool bSimple, gmx_bool bCombined,
935 nbnxn_alloc_t *alloc,
938 nbl_list->bSimple = bSimple;
939 nbl_list->bCombined = bCombined;
941 nbl_list->nnbl = gmx_omp_nthreads_get(emntNonbonded);
943 if (!nbl_list->bCombined &&
944 nbl_list->nnbl > NBNXN_BUFFERFLAG_MAX_THREADS)
946 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.",
947 nbl_list->nnbl, NBNXN_BUFFERFLAG_MAX_THREADS, NBNXN_BUFFERFLAG_MAX_THREADS);
950 snew(nbl_list->nbl, nbl_list->nnbl);
951 if (bSimple && nbl_list->nnbl > 1)
953 snew(nbl_list->nbl_work, nbl_list->nnbl);
955 snew(nbl_list->nbl_fep, nbl_list->nnbl);
956 /* Execute in order to avoid memory interleaving between threads */
957 #pragma omp parallel for num_threads(nbl_list->nnbl) schedule(static)
958 for (int i = 0; i < nbl_list->nnbl; i++)
962 /* Allocate the nblist data structure locally on each thread
963 * to optimize memory access for NUMA architectures.
965 snew(nbl_list->nbl[i], 1);
967 /* Only list 0 is used on the GPU, use normal allocation for i>0 */
968 if (!bSimple && i == 0)
970 nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, alloc, free);
974 nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, nullptr, nullptr);
975 if (bSimple && nbl_list->nnbl > 1)
977 snew(nbl_list->nbl_work[i], 1);
978 nbnxn_init_pairlist(nbl_list->nbl_work[i], nbl_list->bSimple, nullptr, nullptr);
982 snew(nbl_list->nbl_fep[i], 1);
983 nbnxn_init_pairlist_fep(nbl_list->nbl_fep[i]);
985 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
989 /* Print statistics of a pair list, used for debug output */
990 static void print_nblist_statistics_simple(FILE *fp, const nbnxn_pairlist_t *nbl,
991 const nbnxn_search *nbs, real rl)
993 const nbnxn_grid_t *grid;
997 grid = &nbs->grid[0];
999 fprintf(fp, "nbl nci %d ncj %d\n",
1000 nbl->nci, nbl->ncjInUse);
1001 fprintf(fp, "nbl na_sc %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
1002 nbl->na_sc, rl, nbl->ncjInUse, nbl->ncjInUse/static_cast<double>(grid->nc),
1003 nbl->ncjInUse/static_cast<double>(grid->nc)*grid->na_sc,
1004 nbl->ncjInUse/static_cast<double>(grid->nc)*grid->na_sc/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid->nc*grid->na_sc/(grid->size[XX]*grid->size[YY]*grid->size[ZZ])));
1006 fprintf(fp, "nbl average j cell list length %.1f\n",
1007 0.25*nbl->ncjInUse/static_cast<double>(std::max(nbl->nci, 1)));
1009 for (int s = 0; s < SHIFTS; s++)
1014 for (int i = 0; i < nbl->nci; i++)
1016 cs[nbl->ci[i].shift & NBNXN_CI_SHIFT] +=
1017 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start;
1019 int j = nbl->ci[i].cj_ind_start;
1020 while (j < nbl->ci[i].cj_ind_end &&
1021 nbl->cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
1027 fprintf(fp, "nbl cell pairs, total: %d excl: %d %.1f%%\n",
1028 nbl->ncj, npexcl, 100*npexcl/static_cast<double>(std::max(nbl->ncj, 1)));
1029 for (int s = 0; s < SHIFTS; s++)
1033 fprintf(fp, "nbl shift %2d ncj %3d\n", s, cs[s]);
1038 /* Print statistics of a pair lists, used for debug output */
1039 static void print_nblist_statistics_supersub(FILE *fp, const nbnxn_pairlist_t *nbl,
1040 const nbnxn_search *nbs, real rl)
1042 const nbnxn_grid_t *grid;
1044 int c[c_gpuNumClusterPerCell + 1];
1045 double sum_nsp, sum_nsp2;
1048 /* This code only produces correct statistics with domain decomposition */
1049 grid = &nbs->grid[0];
1051 fprintf(fp, "nbl nsci %d ncj4 %d nsi %d excl4 %d\n",
1052 nbl->nsci, nbl->ncj4, nbl->nci_tot, nbl->nexcl);
1053 fprintf(fp, "nbl na_c %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
1054 nbl->na_ci, rl, nbl->nci_tot, nbl->nci_tot/static_cast<double>(grid->nsubc_tot),
1055 nbl->nci_tot/static_cast<double>(grid->nsubc_tot)*grid->na_c,
1056 nbl->nci_tot/static_cast<double>(grid->nsubc_tot)*grid->na_c/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid->nsubc_tot*grid->na_c/(grid->size[XX]*grid->size[YY]*grid->size[ZZ])));
1061 for (int si = 0; si <= c_gpuNumClusterPerCell; si++)
1065 for (int i = 0; i < nbl->nsci; i++)
1070 for (int j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
1072 for (int j = 0; j < c_nbnxnGpuJgroupSize; j++)
1075 for (int si = 0; si < c_gpuNumClusterPerCell; si++)
1077 if (nbl->cj4[j4].imei[0].imask & (1U << (j*c_gpuNumClusterPerCell + si)))
1087 sum_nsp2 += nsp*nsp;
1088 nsp_max = std::max(nsp_max, nsp);
1092 sum_nsp /= nbl->nsci;
1093 sum_nsp2 /= nbl->nsci;
1095 fprintf(fp, "nbl #cluster-pairs: av %.1f stddev %.1f max %d\n",
1096 sum_nsp, std::sqrt(sum_nsp2 - sum_nsp*sum_nsp), nsp_max);
1100 for (b = 0; b <= c_gpuNumClusterPerCell; b++)
1102 fprintf(fp, "nbl j-list #i-subcell %d %7d %4.1f\n",
1104 100.0*c[b]/int{nbl->ncj4*c_nbnxnGpuJgroupSize});
1109 /* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp */
1110 static void low_get_nbl_exclusions(nbnxn_pairlist_t *nbl, int cj4,
1111 int warp, nbnxn_excl_t **excl)
1113 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
1115 /* No exclusions set, make a new list entry */
1116 nbl->cj4[cj4].imei[warp].excl_ind = nbl->nexcl;
1118 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
1119 set_no_excls(*excl);
1123 /* We already have some exclusions, new ones can be added to the list */
1124 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
1128 /* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp,
1129 * generates a new element and allocates extra memory, if necessary.
1131 static void get_nbl_exclusions_1(nbnxn_pairlist_t *nbl, int cj4,
1132 int warp, nbnxn_excl_t **excl)
1134 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
1136 /* We need to make a new list entry, check if we have space */
1137 check_excl_space(nbl, 1);
1139 low_get_nbl_exclusions(nbl, cj4, warp, excl);
1142 /* Returns pointers to the exclusion masks for cj4-unit cj4 for both warps,
1143 * generates a new element and allocates extra memory, if necessary.
1145 static void get_nbl_exclusions_2(nbnxn_pairlist_t *nbl, int cj4,
1146 nbnxn_excl_t **excl_w0,
1147 nbnxn_excl_t **excl_w1)
1149 /* Check for space we might need */
1150 check_excl_space(nbl, 2);
1152 low_get_nbl_exclusions(nbl, cj4, 0, excl_w0);
1153 low_get_nbl_exclusions(nbl, cj4, 1, excl_w1);
1156 /* Sets the self exclusions i=j and pair exclusions i>j */
1157 static void set_self_and_newton_excls_supersub(nbnxn_pairlist_t *nbl,
1158 int cj4_ind, int sj_offset,
1159 int i_cluster_in_cell)
1161 nbnxn_excl_t *excl[c_nbnxnGpuClusterpairSplit];
1163 /* Here we only set the set self and double pair exclusions */
1165 static_assert(c_nbnxnGpuClusterpairSplit == 2, "");
1167 get_nbl_exclusions_2(nbl, cj4_ind, &excl[0], &excl[1]);
1169 /* Only minor < major bits set */
1170 for (int ej = 0; ej < nbl->na_ci; ej++)
1173 for (int ei = ej; ei < nbl->na_ci; ei++)
1175 excl[w]->pair[(ej & (c_nbnxnGpuJgroupSize-1))*nbl->na_ci + ei] &=
1176 ~(1U << (sj_offset*c_gpuNumClusterPerCell + i_cluster_in_cell));
1181 /* Returns a diagonal or off-diagonal interaction mask for plain C lists */
1182 static unsigned int get_imask(gmx_bool rdiag, int ci, int cj)
1184 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
1187 /* Returns a diagonal or off-diagonal interaction mask for cj-size=2 */
1188 gmx_unused static unsigned int get_imask_simd_j2(gmx_bool rdiag, int ci, int cj)
1190 return (rdiag && ci*2 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_0 :
1191 (rdiag && ci*2+1 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_1 :
1192 NBNXN_INTERACTION_MASK_ALL));
1195 /* Returns a diagonal or off-diagonal interaction mask for cj-size=4 */
1196 gmx_unused static unsigned int get_imask_simd_j4(gmx_bool rdiag, int ci, int cj)
1198 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
1201 /* Returns a diagonal or off-diagonal interaction mask for cj-size=8 */
1202 gmx_unused static unsigned int get_imask_simd_j8(gmx_bool rdiag, int ci, int cj)
1204 return (rdiag && ci == cj*2 ? NBNXN_INTERACTION_MASK_DIAG_J8_0 :
1205 (rdiag && ci == cj*2+1 ? NBNXN_INTERACTION_MASK_DIAG_J8_1 :
1206 NBNXN_INTERACTION_MASK_ALL));
1210 #if GMX_SIMD_REAL_WIDTH == 2
1211 #define get_imask_simd_4xn get_imask_simd_j2
1213 #if GMX_SIMD_REAL_WIDTH == 4
1214 #define get_imask_simd_4xn get_imask_simd_j4
1216 #if GMX_SIMD_REAL_WIDTH == 8
1217 #define get_imask_simd_4xn get_imask_simd_j8
1218 #define get_imask_simd_2xnn get_imask_simd_j4
1220 #if GMX_SIMD_REAL_WIDTH == 16
1221 #define get_imask_simd_2xnn get_imask_simd_j8
1225 /* Plain C code for checking and adding cluster-pairs to the list.
1227 * \param[in] gridj The j-grid
1228 * \param[in,out] nbl The pair-list to store the cluster pairs in
1229 * \param[in] icluster The index of the i-cluster
1230 * \param[in] jclusterFirst The first cluster in the j-range
1231 * \param[in] jclusterLast The last cluster in the j-range
1232 * \param[in] excludeSubDiagonal Exclude atom pairs with i-index > j-index
1233 * \param[in] x_j Coordinates for the j-atom, in xyz format
1234 * \param[in] rlist2 The squared list cut-off
1235 * \param[in] rbb2 The squared cut-off for putting cluster-pairs in the list based on bounding box distance only
1236 * \param[in,out] numDistanceChecks The number of distance checks performed
1239 makeClusterListSimple(const nbnxn_grid_t * gridj,
1240 nbnxn_pairlist_t * nbl,
1244 bool excludeSubDiagonal,
1245 const real * gmx_restrict x_j,
1248 int * gmx_restrict numDistanceChecks)
1250 const nbnxn_bb_t * gmx_restrict bb_ci = nbl->work->bb_ci;
1251 const real * gmx_restrict x_ci = nbl->work->x_ci;
1256 while (!InRange && jclusterFirst <= jclusterLast)
1258 real d2 = subc_bb_dist2(0, bb_ci, jclusterFirst, gridj->bb);
1259 *numDistanceChecks += 2;
1261 /* Check if the distance is within the distance where
1262 * we use only the bounding box distance rbb,
1263 * or within the cut-off and there is at least one atom pair
1264 * within the cut-off.
1270 else if (d2 < rlist2)
1272 int cjf_gl = gridj->cell0 + jclusterFirst;
1273 for (int i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
1275 for (int j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE; j++)
1277 InRange = InRange ||
1278 (gmx::square(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
1279 gmx::square(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
1280 gmx::square(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rlist2);
1283 *numDistanceChecks += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
1296 while (!InRange && jclusterLast > jclusterFirst)
1298 real d2 = subc_bb_dist2(0, bb_ci, jclusterLast, gridj->bb);
1299 *numDistanceChecks += 2;
1301 /* Check if the distance is within the distance where
1302 * we use only the bounding box distance rbb,
1303 * or within the cut-off and there is at least one atom pair
1304 * within the cut-off.
1310 else if (d2 < rlist2)
1312 int cjl_gl = gridj->cell0 + jclusterLast;
1313 for (int i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
1315 for (int j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE; j++)
1317 InRange = InRange ||
1318 (gmx::square(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
1319 gmx::square(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
1320 gmx::square(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rlist2);
1323 *numDistanceChecks += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
1331 if (jclusterFirst <= jclusterLast)
1333 for (int jcluster = jclusterFirst; jcluster <= jclusterLast; jcluster++)
1335 /* Store cj and the interaction mask */
1336 nbl->cj[nbl->ncj].cj = gridj->cell0 + jcluster;
1337 nbl->cj[nbl->ncj].excl = get_imask(excludeSubDiagonal, icluster, jcluster);
1340 /* Increase the closing index in i super-cell list */
1341 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
1345 #ifdef GMX_NBNXN_SIMD_4XN
1346 #include "gromacs/mdlib/nbnxn_search_simd_4xn.h"
1348 #ifdef GMX_NBNXN_SIMD_2XNN
1349 #include "gromacs/mdlib/nbnxn_search_simd_2xnn.h"
1352 /* Plain C or SIMD4 code for making a pair list of super-cell sci vs scj.
1353 * Checks bounding box distances and possibly atom pair distances.
1355 static void make_cluster_list_supersub(const nbnxn_grid_t *gridi,
1356 const nbnxn_grid_t *gridj,
1357 nbnxn_pairlist_t *nbl,
1359 gmx_bool sci_equals_scj,
1360 int stride, const real *x,
1361 real rlist2, float rbb2,
1362 int *numDistanceChecks)
1364 nbnxn_list_work_t *work = nbl->work;
1367 const float *pbb_ci = work->pbb_ci;
1369 const nbnxn_bb_t *bb_ci = work->bb_ci;
1372 assert(c_nbnxnGpuClusterSize == gridi->na_c);
1373 assert(c_nbnxnGpuClusterSize == gridj->na_c);
1375 /* We generate the pairlist mainly based on bounding-box distances
1376 * and do atom pair distance based pruning on the GPU.
1377 * Only if a j-group contains a single cluster-pair, we try to prune
1378 * that pair based on atom distances on the CPU to avoid empty j-groups.
1380 #define PRUNE_LIST_CPU_ONE 1
1381 #define PRUNE_LIST_CPU_ALL 0
1383 #if PRUNE_LIST_CPU_ONE
1387 float *d2l = work->d2;
1389 for (int subc = 0; subc < gridj->nsubc[scj]; subc++)
1391 int cj4_ind = nbl->work->cj_ind/c_nbnxnGpuJgroupSize;
1392 int cj_offset = nbl->work->cj_ind - cj4_ind*c_nbnxnGpuJgroupSize;
1393 nbnxn_cj4_t *cj4 = &nbl->cj4[cj4_ind];
1395 int cj = scj*c_gpuNumClusterPerCell + subc;
1397 int cj_gl = gridj->cell0*c_gpuNumClusterPerCell + cj;
1399 /* Initialize this j-subcell i-subcell list */
1400 cj4->cj[cj_offset] = cj_gl;
1409 ci1 = gridi->nsubc[sci];
1413 /* Determine all ci1 bb distances in one call with SIMD4 */
1414 subc_bb_dist2_simd4_xxxx(gridj->pbb.data() + (cj >> STRIDE_PBB_2LOG)*NNBSBB_XXXX + (cj & (STRIDE_PBB-1)),
1416 *numDistanceChecks += c_nbnxnGpuClusterSize*2;
1420 unsigned int imask = 0;
1421 /* We use a fixed upper-bound instead of ci1 to help optimization */
1422 for (int ci = 0; ci < c_gpuNumClusterPerCell; ci++)
1430 /* Determine the bb distance between ci and cj */
1431 d2l[ci] = subc_bb_dist2(ci, bb_ci, cj, gridj->bb);
1432 *numDistanceChecks += 2;
1436 #if PRUNE_LIST_CPU_ALL
1437 /* Check if the distance is within the distance where
1438 * we use only the bounding box distance rbb,
1439 * or within the cut-off and there is at least one atom pair
1440 * within the cut-off. This check is very costly.
1442 *numDistanceChecks += c_nbnxnGpuClusterSize*c_nbnxnGpuClusterSize;
1445 clusterpair_in_range(work, ci, cj_gl, stride, x, rlist2)))
1447 /* Check if the distance between the two bounding boxes
1448 * in within the pair-list cut-off.
1453 /* Flag this i-subcell to be taken into account */
1454 imask |= (1U << (cj_offset*c_gpuNumClusterPerCell + ci));
1456 #if PRUNE_LIST_CPU_ONE
1464 #if PRUNE_LIST_CPU_ONE
1465 /* If we only found 1 pair, check if any atoms are actually
1466 * within the cut-off, so we could get rid of it.
1468 if (npair == 1 && d2l[ci_last] >= rbb2 &&
1469 !clusterpair_in_range(work, ci_last, cj_gl, stride, x, rlist2))
1471 imask &= ~(1U << (cj_offset*c_gpuNumClusterPerCell + ci_last));
1478 /* We have a useful sj entry, close it now */
1480 /* Set the exclusions for the ci==sj entry.
1481 * Here we don't bother to check if this entry is actually flagged,
1482 * as it will nearly always be in the list.
1486 set_self_and_newton_excls_supersub(nbl, cj4_ind, cj_offset, subc);
1489 /* Copy the cluster interaction mask to the list */
1490 for (int w = 0; w < c_nbnxnGpuClusterpairSplit; w++)
1492 cj4->imei[w].imask |= imask;
1495 nbl->work->cj_ind++;
1497 /* Keep the count */
1498 nbl->nci_tot += npair;
1500 /* Increase the closing index in i super-cell list */
1501 nbl->sci[nbl->nsci].cj4_ind_end =
1502 (nbl->work->cj_ind + c_nbnxnGpuJgroupSize - 1)/c_nbnxnGpuJgroupSize;
1507 /* Returns how many contiguous j-clusters we have starting in the i-list */
1508 template <typename CjListType>
1509 static int numContiguousJClusters(const int cjIndexStart,
1510 const int cjIndexEnd,
1511 const CjListType &cjList)
1513 const int firstJCluster = nblCj(cjList, cjIndexStart);
1515 int numContiguous = 0;
1517 while (cjIndexStart + numContiguous < cjIndexEnd &&
1518 nblCj(cjList, cjIndexStart + numContiguous) == firstJCluster + numContiguous)
1523 return numContiguous;
1526 /* Helper struct for efficient searching for excluded atoms in a j-list */
1530 template <typename CjListType>
1531 JListRanges(int cjIndexStart,
1533 const CjListType &cjList);
1535 int cjIndexStart; // The start index in the j-list
1536 int cjIndexEnd; // The end index in the j-list
1537 int cjFirst; // The j-cluster with index cjIndexStart
1538 int cjLast; // The j-cluster with index cjIndexEnd-1
1539 int numDirect; // Up to cjIndexStart+numDirect the j-clusters are cjFirst + the index offset
1542 template <typename CjListType>
1543 JListRanges::JListRanges(int cjIndexStart,
1545 const CjListType &cjList) :
1546 cjIndexStart(cjIndexStart),
1547 cjIndexEnd(cjIndexEnd)
1549 GMX_ASSERT(cjIndexEnd > cjIndexStart, "JListRanges should only be called with non-empty lists");
1551 cjFirst = nblCj(cjList, cjIndexStart);
1552 cjLast = nblCj(cjList, cjIndexEnd - 1);
1554 /* Determine how many contiguous j-cells we have starting
1555 * from the first i-cell. This number can be used to directly
1556 * calculate j-cell indices for excluded atoms.
1558 numDirect = numContiguousJClusters(cjIndexStart, cjIndexEnd, cjList);
1561 /* Return the index of \p jCluster in the given range or -1 when not present
1563 * Note: This code is executed very often and therefore performance is
1564 * important. It should be inlined and fully optimized.
1566 template <typename CjListType>
1567 static inline int findJClusterInJList(int jCluster,
1568 const JListRanges &ranges,
1569 const CjListType &cjList)
1573 if (jCluster < ranges.cjFirst + ranges.numDirect)
1575 /* We can calculate the index directly using the offset */
1576 index = ranges.cjIndexStart + jCluster - ranges.cjFirst;
1580 /* Search for jCluster using bisection */
1582 int rangeStart = ranges.cjIndexStart + ranges.numDirect;
1583 int rangeEnd = ranges.cjIndexEnd;
1585 while (index == -1 && rangeStart < rangeEnd)
1587 rangeMiddle = (rangeStart + rangeEnd) >> 1;
1589 const int clusterMiddle = nblCj(cjList, rangeMiddle);
1591 if (jCluster == clusterMiddle)
1593 index = rangeMiddle;
1595 else if (jCluster < clusterMiddle)
1597 rangeEnd = rangeMiddle;
1601 rangeStart = rangeMiddle + 1;
1609 /* Set all atom-pair exclusions for a simple type list i-entry
1611 * Set all atom-pair exclusions from the topology stored in exclusions
1612 * as masks in the pair-list for simple list entry iEntry.
1615 setExclusionsForSimpleIentry(const nbnxn_search *nbs,
1616 nbnxn_pairlist_t *nbl,
1617 gmx_bool diagRemoved,
1619 const nbnxn_ci_t &iEntry,
1620 const t_blocka &exclusions)
1622 if (iEntry.cj_ind_end == iEntry.cj_ind_start)
1624 /* Empty list: no exclusions */
1628 const JListRanges ranges(iEntry.cj_ind_start, iEntry.cj_ind_end, nbl->cj);
1630 const int iCluster = iEntry.ci;
1632 gmx::ArrayRef<const int> cell = nbs->cell;
1634 /* Loop over the atoms in the i-cluster */
1635 for (int i = 0; i < nbl->na_sc; i++)
1637 const int iIndex = iCluster*nbl->na_sc + i;
1638 const int iAtom = nbs->a[iIndex];
1641 /* Loop over the topology-based exclusions for this i-atom */
1642 for (int exclIndex = exclusions.index[iAtom]; exclIndex < exclusions.index[iAtom + 1]; exclIndex++)
1644 const int jAtom = exclusions.a[exclIndex];
1648 /* The self exclusion are already set, save some time */
1652 /* Get the index of the j-atom in the nbnxn atom data */
1653 const int jIndex = cell[jAtom];
1655 /* Without shifts we only calculate interactions j>i
1656 * for one-way pair-lists.
1658 if (diagRemoved && jIndex <= iIndex)
1663 const int jCluster = (jIndex >> na_cj_2log);
1665 /* Could the cluster se be in our list? */
1666 if (jCluster >= ranges.cjFirst && jCluster <= ranges.cjLast)
1669 findJClusterInJList(jCluster, ranges, nbl->cj);
1673 /* We found an exclusion, clear the corresponding
1676 const int innerJ = jIndex - (jCluster << na_cj_2log);
1678 nbl->cj[index].excl &= ~(1U << ((i << na_cj_2log) + innerJ));
1686 /* Add a new i-entry to the FEP list and copy the i-properties */
1687 static inline void fep_list_new_nri_copy(t_nblist *nlist)
1689 /* Add a new i-entry */
1692 assert(nlist->nri < nlist->maxnri);
1694 /* Duplicate the last i-entry, except for jindex, which continues */
1695 nlist->iinr[nlist->nri] = nlist->iinr[nlist->nri-1];
1696 nlist->shift[nlist->nri] = nlist->shift[nlist->nri-1];
1697 nlist->gid[nlist->nri] = nlist->gid[nlist->nri-1];
1698 nlist->jindex[nlist->nri] = nlist->nrj;
1701 /* For load balancing of the free-energy lists over threads, we set
1702 * the maximum nrj size of an i-entry to 40. This leads to good
1703 * load balancing in the worst case scenario of a single perturbed
1704 * particle on 16 threads, while not introducing significant overhead.
1705 * Note that half of the perturbed pairs will anyhow end up in very small lists,
1706 * since non perturbed i-particles will see few perturbed j-particles).
1708 const int max_nrj_fep = 40;
1710 /* Exclude the perturbed pairs from the Verlet list. This is only done to avoid
1711 * singularities for overlapping particles (0/0), since the charges and
1712 * LJ parameters have been zeroed in the nbnxn data structure.
1713 * Simultaneously make a group pair list for the perturbed pairs.
1715 static void make_fep_list(const nbnxn_search *nbs,
1716 const nbnxn_atomdata_t *nbat,
1717 nbnxn_pairlist_t *nbl,
1718 gmx_bool bDiagRemoved,
1720 const nbnxn_grid_t *gridi,
1721 const nbnxn_grid_t *gridj,
1724 int ci, cj_ind_start, cj_ind_end, cja, cjr;
1726 int ngid, gid_i = 0, gid_j, gid;
1727 int egp_shift, egp_mask;
1729 int ind_i, ind_j, ai, aj;
1731 gmx_bool bFEP_i, bFEP_i_all;
1733 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
1741 cj_ind_start = nbl_ci->cj_ind_start;
1742 cj_ind_end = nbl_ci->cj_ind_end;
1744 /* In worst case we have alternating energy groups
1745 * and create #atom-pair lists, which means we need the size
1746 * of a cluster pair (na_ci*na_cj) times the number of cj's.
1748 nri_max = nbl->na_ci*nbl->na_cj*(cj_ind_end - cj_ind_start);
1749 if (nlist->nri + nri_max > nlist->maxnri)
1751 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
1752 reallocate_nblist(nlist);
1755 ngid = nbat->nenergrp;
1757 if (ngid*gridj->na_cj > gmx::index(sizeof(gid_cj)*8))
1759 gmx_fatal(FARGS, "The Verlet scheme with %dx%d kernels and free-energy only supports up to %zu energy groups",
1760 gridi->na_c, gridj->na_cj, (sizeof(gid_cj)*8)/gridj->na_cj);
1763 egp_shift = nbat->neg_2log;
1764 egp_mask = (1<<nbat->neg_2log) - 1;
1766 /* Loop over the atoms in the i sub-cell */
1768 for (int i = 0; i < nbl->na_ci; i++)
1770 ind_i = ci*nbl->na_ci + i;
1775 nlist->jindex[nri+1] = nlist->jindex[nri];
1776 nlist->iinr[nri] = ai;
1777 /* The actual energy group pair index is set later */
1778 nlist->gid[nri] = 0;
1779 nlist->shift[nri] = nbl_ci->shift & NBNXN_CI_SHIFT;
1781 bFEP_i = ((gridi->fep[ci - gridi->cell0] & (1 << i)) != 0u);
1783 bFEP_i_all = bFEP_i_all && bFEP_i;
1785 if (nlist->nrj + (cj_ind_end - cj_ind_start)*nbl->na_cj > nlist->maxnrj)
1787 nlist->maxnrj = over_alloc_small(nlist->nrj + (cj_ind_end - cj_ind_start)*nbl->na_cj);
1788 srenew(nlist->jjnr, nlist->maxnrj);
1789 srenew(nlist->excl_fep, nlist->maxnrj);
1794 gid_i = (nbat->energrp[ci] >> (egp_shift*i)) & egp_mask;
1797 for (int cj_ind = cj_ind_start; cj_ind < cj_ind_end; cj_ind++)
1799 unsigned int fep_cj;
1801 cja = nbl->cj[cj_ind].cj;
1803 if (gridj->na_cj == gridj->na_c)
1805 cjr = cja - gridj->cell0;
1806 fep_cj = gridj->fep[cjr];
1809 gid_cj = nbat->energrp[cja];
1812 else if (2*gridj->na_cj == gridj->na_c)
1814 cjr = cja - gridj->cell0*2;
1815 /* Extract half of the ci fep/energrp mask */
1816 fep_cj = (gridj->fep[cjr>>1] >> ((cjr&1)*gridj->na_cj)) & ((1<<gridj->na_cj) - 1);
1819 gid_cj = nbat->energrp[cja>>1] >> ((cja&1)*gridj->na_cj*egp_shift) & ((1<<(gridj->na_cj*egp_shift)) - 1);
1824 cjr = cja - (gridj->cell0>>1);
1825 /* Combine two ci fep masks/energrp */
1826 fep_cj = gridj->fep[cjr*2] + (gridj->fep[cjr*2+1] << gridj->na_c);
1829 gid_cj = nbat->energrp[cja*2] + (nbat->energrp[cja*2+1] << (gridj->na_c*egp_shift));
1833 if (bFEP_i || fep_cj != 0)
1835 for (int j = 0; j < nbl->na_cj; j++)
1837 /* Is this interaction perturbed and not excluded? */
1838 ind_j = cja*nbl->na_cj + j;
1841 (bFEP_i || (fep_cj & (1 << j))) &&
1842 (!bDiagRemoved || ind_j >= ind_i))
1846 gid_j = (gid_cj >> (j*egp_shift)) & egp_mask;
1847 gid = GID(gid_i, gid_j, ngid);
1849 if (nlist->nrj > nlist->jindex[nri] &&
1850 nlist->gid[nri] != gid)
1852 /* Energy group pair changed: new list */
1853 fep_list_new_nri_copy(nlist);
1856 nlist->gid[nri] = gid;
1859 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
1861 fep_list_new_nri_copy(nlist);
1865 /* Add it to the FEP list */
1866 nlist->jjnr[nlist->nrj] = aj;
1867 nlist->excl_fep[nlist->nrj] = (nbl->cj[cj_ind].excl >> (i*nbl->na_cj + j)) & 1;
1870 /* Exclude it from the normal list.
1871 * Note that the charge has been set to zero,
1872 * but we need to avoid 0/0, as perturbed atoms
1873 * can be on top of each other.
1875 nbl->cj[cj_ind].excl &= ~(1U << (i*nbl->na_cj + j));
1881 if (nlist->nrj > nlist->jindex[nri])
1883 /* Actually add this new, non-empty, list */
1885 nlist->jindex[nlist->nri] = nlist->nrj;
1892 /* All interactions are perturbed, we can skip this entry */
1893 nbl_ci->cj_ind_end = cj_ind_start;
1894 nbl->ncjInUse -= cj_ind_end - cj_ind_start;
1898 /* Return the index of atom a within a cluster */
1899 static inline int cj_mod_cj4(int cj)
1901 return cj & (c_nbnxnGpuJgroupSize - 1);
1904 /* Convert a j-cluster to a cj4 group */
1905 static inline int cj_to_cj4(int cj)
1907 return cj/c_nbnxnGpuJgroupSize;
1910 /* Return the index of an j-atom within a warp */
1911 static inline int a_mod_wj(int a)
1913 return a & (c_nbnxnGpuClusterSize/c_nbnxnGpuClusterpairSplit - 1);
1916 /* As make_fep_list above, but for super/sub lists. */
1917 static void make_fep_list_supersub(const nbnxn_search *nbs,
1918 const nbnxn_atomdata_t *nbat,
1919 nbnxn_pairlist_t *nbl,
1920 gmx_bool bDiagRemoved,
1921 const nbnxn_sci_t *nbl_sci,
1926 const nbnxn_grid_t *gridi,
1927 const nbnxn_grid_t *gridj,
1930 int sci, cj4_ind_start, cj4_ind_end, cjr;
1933 int ind_i, ind_j, ai, aj;
1937 const nbnxn_cj4_t *cj4;
1939 if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
1947 cj4_ind_start = nbl_sci->cj4_ind_start;
1948 cj4_ind_end = nbl_sci->cj4_ind_end;
1950 /* Here we process one super-cell, max #atoms na_sc, versus a list
1951 * cj4 entries, each with max c_nbnxnGpuJgroupSize cj's, each
1952 * of size na_cj atoms.
1953 * On the GPU we don't support energy groups (yet).
1954 * So for each of the na_sc i-atoms, we need max one FEP list
1955 * for each max_nrj_fep j-atoms.
1957 nri_max = nbl->na_sc*nbl->na_cj*(1 + ((cj4_ind_end - cj4_ind_start)*c_nbnxnGpuJgroupSize)/max_nrj_fep);
1958 if (nlist->nri + nri_max > nlist->maxnri)
1960 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
1961 reallocate_nblist(nlist);
1964 /* Loop over the atoms in the i super-cluster */
1965 for (int c = 0; c < c_gpuNumClusterPerCell; c++)
1967 c_abs = sci*c_gpuNumClusterPerCell + c;
1969 for (int i = 0; i < nbl->na_ci; i++)
1971 ind_i = c_abs*nbl->na_ci + i;
1976 nlist->jindex[nri+1] = nlist->jindex[nri];
1977 nlist->iinr[nri] = ai;
1978 /* With GPUs, energy groups are not supported */
1979 nlist->gid[nri] = 0;
1980 nlist->shift[nri] = nbl_sci->shift & NBNXN_CI_SHIFT;
1982 bFEP_i = ((gridi->fep[c_abs - gridi->cell0*c_gpuNumClusterPerCell] & (1 << i)) != 0u);
1984 xi = nbat->x[ind_i*nbat->xstride+XX] + shx;
1985 yi = nbat->x[ind_i*nbat->xstride+YY] + shy;
1986 zi = nbat->x[ind_i*nbat->xstride+ZZ] + shz;
1988 if ((nlist->nrj + cj4_ind_end - cj4_ind_start)*c_nbnxnGpuJgroupSize*nbl->na_cj > nlist->maxnrj)
1990 nlist->maxnrj = over_alloc_small((nlist->nrj + cj4_ind_end - cj4_ind_start)*c_nbnxnGpuJgroupSize*nbl->na_cj);
1991 srenew(nlist->jjnr, nlist->maxnrj);
1992 srenew(nlist->excl_fep, nlist->maxnrj);
1995 for (int cj4_ind = cj4_ind_start; cj4_ind < cj4_ind_end; cj4_ind++)
1997 cj4 = &nbl->cj4[cj4_ind];
1999 for (int gcj = 0; gcj < c_nbnxnGpuJgroupSize; gcj++)
2001 unsigned int fep_cj;
2003 if ((cj4->imei[0].imask & (1U << (gcj*c_gpuNumClusterPerCell + c))) == 0)
2005 /* Skip this ci for this cj */
2009 cjr = cj4->cj[gcj] - gridj->cell0*c_gpuNumClusterPerCell;
2011 fep_cj = gridj->fep[cjr];
2013 if (bFEP_i || fep_cj != 0)
2015 for (int j = 0; j < nbl->na_cj; j++)
2017 /* Is this interaction perturbed and not excluded? */
2018 ind_j = (gridj->cell0*c_gpuNumClusterPerCell + cjr)*nbl->na_cj + j;
2021 (bFEP_i || (fep_cj & (1 << j))) &&
2022 (!bDiagRemoved || ind_j >= ind_i))
2026 unsigned int excl_bit;
2029 const int jHalf = j/(c_nbnxnGpuClusterSize/c_nbnxnGpuClusterpairSplit);
2030 get_nbl_exclusions_1(nbl, cj4_ind, jHalf, &excl);
2032 excl_pair = a_mod_wj(j)*nbl->na_ci + i;
2033 excl_bit = (1U << (gcj*c_gpuNumClusterPerCell + c));
2035 dx = nbat->x[ind_j*nbat->xstride+XX] - xi;
2036 dy = nbat->x[ind_j*nbat->xstride+YY] - yi;
2037 dz = nbat->x[ind_j*nbat->xstride+ZZ] - zi;
2039 /* The unpruned GPU list has more than 2/3
2040 * of the atom pairs beyond rlist. Using
2041 * this list will cause a lot of overhead
2042 * in the CPU FEP kernels, especially
2043 * relative to the fast GPU kernels.
2044 * So we prune the FEP list here.
2046 if (dx*dx + dy*dy + dz*dz < rlist_fep2)
2048 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
2050 fep_list_new_nri_copy(nlist);
2054 /* Add it to the FEP list */
2055 nlist->jjnr[nlist->nrj] = aj;
2056 nlist->excl_fep[nlist->nrj] = (excl->pair[excl_pair] & excl_bit) ? 1 : 0;
2060 /* Exclude it from the normal list.
2061 * Note that the charge and LJ parameters have
2062 * been set to zero, but we need to avoid 0/0,
2063 * as perturbed atoms can be on top of each other.
2065 excl->pair[excl_pair] &= ~excl_bit;
2069 /* Note that we could mask out this pair in imask
2070 * if all i- and/or all j-particles are perturbed.
2071 * But since the perturbed pairs on the CPU will
2072 * take an order of magnitude more time, the GPU
2073 * will finish before the CPU and there is no gain.
2079 if (nlist->nrj > nlist->jindex[nri])
2081 /* Actually add this new, non-empty, list */
2083 nlist->jindex[nlist->nri] = nlist->nrj;
2090 /* Set all atom-pair exclusions for a GPU type list i-entry
2092 * Sets all atom-pair exclusions from the topology stored in exclusions
2093 * as masks in the pair-list for i-super-cluster list entry iEntry.
2096 setExclusionsForGpuIentry(const nbnxn_search *nbs,
2097 nbnxn_pairlist_t *nbl,
2098 gmx_bool diagRemoved,
2099 const nbnxn_sci_t &iEntry,
2100 const t_blocka &exclusions)
2102 if (iEntry.cj4_ind_end == iEntry.cj4_ind_start)
2108 /* Set the search ranges using start and end j-cluster indices.
2109 * Note that here we can not use cj4_ind_end, since the last cj4
2110 * can be only partially filled, so we use cj_ind.
2112 const JListRanges ranges(iEntry.cj4_ind_start*c_nbnxnGpuJgroupSize,
2116 GMX_ASSERT(nbl->na_ci == c_nbnxnGpuClusterSize, "na_ci should match the GPU cluster size");
2117 constexpr int c_clusterSize = c_nbnxnGpuClusterSize;
2118 constexpr int c_superClusterSize = c_nbnxnGpuNumClusterPerSupercluster*c_nbnxnGpuClusterSize;
2120 const int iSuperCluster = iEntry.sci;
2122 gmx::ArrayRef<const int> cell = nbs->cell;
2124 /* Loop over the atoms in the i super-cluster */
2125 for (int i = 0; i < c_superClusterSize; i++)
2127 const int iIndex = iSuperCluster*c_superClusterSize + i;
2128 const int iAtom = nbs->a[iIndex];
2131 const int iCluster = i/c_clusterSize;
2133 /* Loop over the topology-based exclusions for this i-atom */
2134 for (int exclIndex = exclusions.index[iAtom]; exclIndex < exclusions.index[iAtom + 1]; exclIndex++)
2136 const int jAtom = exclusions.a[exclIndex];
2140 /* The self exclusions are already set, save some time */
2144 /* Get the index of the j-atom in the nbnxn atom data */
2145 const int jIndex = cell[jAtom];
2147 /* Without shifts we only calculate interactions j>i
2148 * for one-way pair-lists.
2150 /* NOTE: We would like to use iIndex on the right hand side,
2151 * but that makes this routine 25% slower with gcc6/7.
2152 * Even using c_superClusterSize makes it slower.
2153 * Either of these changes triggers peeling of the exclIndex
2154 * loop, which apparently leads to far less efficient code.
2156 if (diagRemoved && jIndex <= iSuperCluster*nbl->na_sc + i)
2161 const int jCluster = jIndex/c_clusterSize;
2163 /* Check whether the cluster is in our list? */
2164 if (jCluster >= ranges.cjFirst && jCluster <= ranges.cjLast)
2167 findJClusterInJList(jCluster, ranges, nbl->cj4);
2171 /* We found an exclusion, clear the corresponding
2174 const unsigned int pairMask = (1U << (cj_mod_cj4(index)*c_gpuNumClusterPerCell + iCluster));
2175 /* Check if the i-cluster interacts with the j-cluster */
2176 if (nbl_imask0(nbl, index) & pairMask)
2178 const int innerI = (i & (c_clusterSize - 1));
2179 const int innerJ = (jIndex & (c_clusterSize - 1));
2181 /* Determine which j-half (CUDA warp) we are in */
2182 const int jHalf = innerJ/(c_clusterSize/c_nbnxnGpuClusterpairSplit);
2184 nbnxn_excl_t *interactionMask;
2185 get_nbl_exclusions_1(nbl, cj_to_cj4(index), jHalf, &interactionMask);
2187 interactionMask->pair[a_mod_wj(innerJ)*c_clusterSize + innerI] &= ~pairMask;
2196 /* Reallocate the simple ci list for at least n entries */
2197 static void nb_realloc_ci(nbnxn_pairlist_t *nbl, int n)
2199 nbl->ci_nalloc = over_alloc_small(n);
2200 nbnxn_realloc_void(reinterpret_cast<void **>(&nbl->ci),
2201 nbl->nci*sizeof(*nbl->ci),
2202 nbl->ci_nalloc*sizeof(*nbl->ci),
2203 nbl->alloc, nbl->free);
2205 nbnxn_realloc_void(reinterpret_cast<void **>(&nbl->ciOuter),
2206 nbl->nci*sizeof(*nbl->ciOuter),
2207 nbl->ci_nalloc*sizeof(*nbl->ciOuter),
2208 nbl->alloc, nbl->free);
2211 /* Reallocate the super-cell sci list for at least n entries */
2212 static void nb_realloc_sci(nbnxn_pairlist_t *nbl, int n)
2214 nbl->sci_nalloc = over_alloc_small(n);
2215 nbnxn_realloc_void(reinterpret_cast<void **>(&nbl->sci),
2216 nbl->nsci*sizeof(*nbl->sci),
2217 nbl->sci_nalloc*sizeof(*nbl->sci),
2218 nbl->alloc, nbl->free);
2221 /* Make a new ci entry at index nbl->nci */
2222 static void new_ci_entry(nbnxn_pairlist_t *nbl, int ci, int shift, int flags)
2224 if (nbl->nci + 1 > nbl->ci_nalloc)
2226 nb_realloc_ci(nbl, nbl->nci+1);
2228 nbl->ci[nbl->nci].ci = ci;
2229 nbl->ci[nbl->nci].shift = shift;
2230 /* Store the interaction flags along with the shift */
2231 nbl->ci[nbl->nci].shift |= flags;
2232 nbl->ci[nbl->nci].cj_ind_start = nbl->ncj;
2233 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
2236 /* Make a new sci entry at index nbl->nsci */
2237 static void new_sci_entry(nbnxn_pairlist_t *nbl, int sci, int shift)
2239 if (nbl->nsci + 1 > nbl->sci_nalloc)
2241 nb_realloc_sci(nbl, nbl->nsci+1);
2243 nbl->sci[nbl->nsci].sci = sci;
2244 nbl->sci[nbl->nsci].shift = shift;
2245 nbl->sci[nbl->nsci].cj4_ind_start = nbl->ncj4;
2246 nbl->sci[nbl->nsci].cj4_ind_end = nbl->ncj4;
2249 /* Sort the simple j-list cj on exclusions.
2250 * Entries with exclusions will all be sorted to the beginning of the list.
2252 static void sort_cj_excl(nbnxn_cj_t *cj, int ncj,
2253 nbnxn_list_work_t *work)
2257 if (ncj > work->cj_nalloc)
2259 work->cj_nalloc = over_alloc_large(ncj);
2260 srenew(work->cj, work->cj_nalloc);
2263 /* Make a list of the j-cells involving exclusions */
2265 for (int j = 0; j < ncj; j++)
2267 if (cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
2269 work->cj[jnew++] = cj[j];
2272 /* Check if there are exclusions at all or not just the first entry */
2273 if (!((jnew == 0) ||
2274 (jnew == 1 && cj[0].excl != NBNXN_INTERACTION_MASK_ALL)))
2276 for (int j = 0; j < ncj; j++)
2278 if (cj[j].excl == NBNXN_INTERACTION_MASK_ALL)
2280 work->cj[jnew++] = cj[j];
2283 for (int j = 0; j < ncj; j++)
2285 cj[j] = work->cj[j];
2290 /* Close this simple list i entry */
2291 static void close_ci_entry_simple(nbnxn_pairlist_t *nbl)
2295 /* All content of the new ci entry have already been filled correctly,
2296 * we only need to increase the count here (for non empty lists).
2298 jlen = nbl->ci[nbl->nci].cj_ind_end - nbl->ci[nbl->nci].cj_ind_start;
2301 sort_cj_excl(nbl->cj+nbl->ci[nbl->nci].cj_ind_start, jlen, nbl->work);
2303 /* The counts below are used for non-bonded pair/flop counts
2304 * and should therefore match the available kernel setups.
2306 if (!(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_COUL(0)))
2308 nbl->work->ncj_noq += jlen;
2310 else if ((nbl->ci[nbl->nci].shift & NBNXN_CI_HALF_LJ(0)) ||
2311 !(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_LJ(0)))
2313 nbl->work->ncj_hlj += jlen;
2320 /* Split sci entry for load balancing on the GPU.
2321 * Splitting ensures we have enough lists to fully utilize the whole GPU.
2322 * With progBal we generate progressively smaller lists, which improves
2323 * load balancing. As we only know the current count on our own thread,
2324 * we will need to estimate the current total amount of i-entries.
2325 * As the lists get concatenated later, this estimate depends
2326 * both on nthread and our own thread index.
2328 static void split_sci_entry(nbnxn_pairlist_t *nbl,
2330 gmx_bool progBal, float nsp_tot_est,
2331 int thread, int nthread)
2334 int cj4_start, cj4_end, j4len;
2336 int nsp, nsp_sci, nsp_cj4, nsp_cj4_e, nsp_cj4_p;
2342 /* Estimate the total numbers of ci's of the nblist combined
2343 * over all threads using the target number of ci's.
2345 nsp_est = (nsp_tot_est*thread)/nthread + nbl->nci_tot;
2347 /* The first ci blocks should be larger, to avoid overhead.
2348 * The last ci blocks should be smaller, to improve load balancing.
2349 * The factor 3/2 makes the first block 3/2 times the target average
2350 * and ensures that the total number of blocks end up equal to
2351 * that of equally sized blocks of size nsp_target_av.
2353 nsp_max = static_cast<int>(nsp_target_av*(nsp_tot_est*1.5/(nsp_est + nsp_tot_est)));
2357 nsp_max = nsp_target_av;
2360 cj4_start = nbl->sci[nbl->nsci-1].cj4_ind_start;
2361 cj4_end = nbl->sci[nbl->nsci-1].cj4_ind_end;
2362 j4len = cj4_end - cj4_start;
2364 if (j4len > 1 && j4len*c_gpuNumClusterPerCell*c_nbnxnGpuJgroupSize > nsp_max)
2366 /* Remove the last ci entry and process the cj4's again */
2374 for (int cj4 = cj4_start; cj4 < cj4_end; cj4++)
2376 nsp_cj4_p = nsp_cj4;
2377 /* Count the number of cluster pairs in this cj4 group */
2379 for (int p = 0; p < c_gpuNumClusterPerCell*c_nbnxnGpuJgroupSize; p++)
2381 nsp_cj4 += (nbl->cj4[cj4].imei[0].imask >> p) & 1;
2384 /* If adding the current cj4 with nsp_cj4 pairs get us further
2385 * away from our target nsp_max, split the list before this cj4.
2387 if (nsp > 0 && nsp_max - nsp < nsp + nsp_cj4 - nsp_max)
2389 /* Split the list at cj4 */
2390 nbl->sci[sci].cj4_ind_end = cj4;
2391 /* Create a new sci entry */
2394 if (nbl->nsci+1 > nbl->sci_nalloc)
2396 nb_realloc_sci(nbl, nbl->nsci+1);
2398 nbl->sci[sci].sci = nbl->sci[nbl->nsci-1].sci;
2399 nbl->sci[sci].shift = nbl->sci[nbl->nsci-1].shift;
2400 nbl->sci[sci].cj4_ind_start = cj4;
2402 nsp_cj4_e = nsp_cj4_p;
2408 /* Put the remaining cj4's in the last sci entry */
2409 nbl->sci[sci].cj4_ind_end = cj4_end;
2411 /* Possibly balance out the last two sci's
2412 * by moving the last cj4 of the second last sci.
2414 if (nsp_sci - nsp_cj4_e >= nsp + nsp_cj4_e)
2416 nbl->sci[sci-1].cj4_ind_end--;
2417 nbl->sci[sci].cj4_ind_start--;
2424 /* Clost this super/sub list i entry */
2425 static void close_ci_entry_supersub(nbnxn_pairlist_t *nbl,
2427 gmx_bool progBal, float nsp_tot_est,
2428 int thread, int nthread)
2430 /* All content of the new ci entry have already been filled correctly,
2431 * we only need to increase the count here (for non empty lists).
2433 int j4len = nbl->sci[nbl->nsci].cj4_ind_end - nbl->sci[nbl->nsci].cj4_ind_start;
2436 /* We can only have complete blocks of 4 j-entries in a list,
2437 * so round the count up before closing.
2439 nbl->ncj4 = (nbl->work->cj_ind + c_nbnxnGpuJgroupSize - 1)/c_nbnxnGpuJgroupSize;
2440 nbl->work->cj_ind = nbl->ncj4*c_nbnxnGpuJgroupSize;
2446 /* Measure the size of the new entry and potentially split it */
2447 split_sci_entry(nbl, nsp_max_av, progBal, nsp_tot_est,
2453 /* Syncs the working array before adding another grid pair to the list */
2454 static void sync_work(nbnxn_pairlist_t *nbl)
2458 nbl->work->cj_ind = nbl->ncj4*c_nbnxnGpuJgroupSize;
2459 nbl->work->cj4_init = nbl->ncj4;
2463 /* Clears an nbnxn_pairlist_t data structure */
2464 static void clear_pairlist(nbnxn_pairlist_t *nbl)
2475 nbl->work->ncj_noq = 0;
2476 nbl->work->ncj_hlj = 0;
2479 /* Clears a group scheme pair list */
2480 static void clear_pairlist_fep(t_nblist *nl)
2484 if (nl->jindex == nullptr)
2486 snew(nl->jindex, 1);
2491 /* Sets a simple list i-cell bounding box, including PBC shift */
2492 static inline void set_icell_bb_simple(gmx::ArrayRef<const nbnxn_bb_t> bb,
2494 real shx, real shy, real shz,
2497 bb_ci->lower[BB_X] = bb[ci].lower[BB_X] + shx;
2498 bb_ci->lower[BB_Y] = bb[ci].lower[BB_Y] + shy;
2499 bb_ci->lower[BB_Z] = bb[ci].lower[BB_Z] + shz;
2500 bb_ci->upper[BB_X] = bb[ci].upper[BB_X] + shx;
2501 bb_ci->upper[BB_Y] = bb[ci].upper[BB_Y] + shy;
2502 bb_ci->upper[BB_Z] = bb[ci].upper[BB_Z] + shz;
2506 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
2507 static void set_icell_bbxxxx_supersub(gmx::ArrayRef<const float> bb,
2509 real shx, real shy, real shz,
2512 int ia = ci*(c_gpuNumClusterPerCell >> STRIDE_PBB_2LOG)*NNBSBB_XXXX;
2513 for (int m = 0; m < (c_gpuNumClusterPerCell >> STRIDE_PBB_2LOG)*NNBSBB_XXXX; m += NNBSBB_XXXX)
2515 for (int i = 0; i < STRIDE_PBB; i++)
2517 bb_ci[m+0*STRIDE_PBB+i] = bb[ia+m+0*STRIDE_PBB+i] + shx;
2518 bb_ci[m+1*STRIDE_PBB+i] = bb[ia+m+1*STRIDE_PBB+i] + shy;
2519 bb_ci[m+2*STRIDE_PBB+i] = bb[ia+m+2*STRIDE_PBB+i] + shz;
2520 bb_ci[m+3*STRIDE_PBB+i] = bb[ia+m+3*STRIDE_PBB+i] + shx;
2521 bb_ci[m+4*STRIDE_PBB+i] = bb[ia+m+4*STRIDE_PBB+i] + shy;
2522 bb_ci[m+5*STRIDE_PBB+i] = bb[ia+m+5*STRIDE_PBB+i] + shz;
2528 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
2529 gmx_unused static void set_icell_bb_supersub(gmx::ArrayRef<const nbnxn_bb_t> bb,
2531 real shx, real shy, real shz,
2534 for (int i = 0; i < c_gpuNumClusterPerCell; i++)
2536 set_icell_bb_simple(bb, ci*c_gpuNumClusterPerCell+i,
2542 /* Copies PBC shifted i-cell atom coordinates x,y,z to working array */
2543 static void icell_set_x_simple(int ci,
2544 real shx, real shy, real shz,
2545 int stride, const real *x,
2546 nbnxn_list_work_t *work)
2548 int ia = ci*NBNXN_CPU_CLUSTER_I_SIZE;
2550 for (int i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE; i++)
2552 work->x_ci[i*STRIDE_XYZ+XX] = x[(ia+i)*stride+XX] + shx;
2553 work->x_ci[i*STRIDE_XYZ+YY] = x[(ia+i)*stride+YY] + shy;
2554 work->x_ci[i*STRIDE_XYZ+ZZ] = x[(ia+i)*stride+ZZ] + shz;
2558 /* Copies PBC shifted super-cell atom coordinates x,y,z to working array */
2559 static void icell_set_x_supersub(int ci,
2560 real shx, real shy, real shz,
2561 int stride, const real *x,
2562 nbnxn_list_work_t *work)
2564 #if !GMX_SIMD4_HAVE_REAL
2566 real * x_ci = work->x_ci;
2568 int ia = ci*c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize;
2569 for (int i = 0; i < c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize; i++)
2571 x_ci[i*DIM + XX] = x[(ia+i)*stride + XX] + shx;
2572 x_ci[i*DIM + YY] = x[(ia+i)*stride + YY] + shy;
2573 x_ci[i*DIM + ZZ] = x[(ia+i)*stride + ZZ] + shz;
2576 #else /* !GMX_SIMD4_HAVE_REAL */
2578 real * x_ci = work->x_ci_simd;
2580 for (int si = 0; si < c_gpuNumClusterPerCell; si++)
2582 for (int i = 0; i < c_nbnxnGpuClusterSize; i += GMX_SIMD4_WIDTH)
2584 int io = si*c_nbnxnGpuClusterSize + i;
2585 int ia = ci*c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize + io;
2586 for (int j = 0; j < GMX_SIMD4_WIDTH; j++)
2588 x_ci[io*DIM + j + XX*GMX_SIMD4_WIDTH] = x[(ia + j)*stride + XX] + shx;
2589 x_ci[io*DIM + j + YY*GMX_SIMD4_WIDTH] = x[(ia + j)*stride + YY] + shy;
2590 x_ci[io*DIM + j + ZZ*GMX_SIMD4_WIDTH] = x[(ia + j)*stride + ZZ] + shz;
2595 #endif /* !GMX_SIMD4_HAVE_REAL */
2598 static real minimum_subgrid_size_xy(const nbnxn_grid_t *grid)
2602 return std::min(grid->cellSize[XX], grid->cellSize[YY]);
2606 return std::min(grid->cellSize[XX]/c_gpuNumClusterPerCellX,
2607 grid->cellSize[YY]/c_gpuNumClusterPerCellY);
2611 static real effective_buffer_1x1_vs_MxN(const nbnxn_grid_t *gridi,
2612 const nbnxn_grid_t *gridj)
2614 const real eff_1x1_buffer_fac_overest = 0.1;
2616 /* Determine an atom-pair list cut-off buffer size for atom pairs,
2617 * to be added to rlist (including buffer) used for MxN.
2618 * This is for converting an MxN list to a 1x1 list. This means we can't
2619 * use the normal buffer estimate, as we have an MxN list in which
2620 * some atom pairs beyond rlist are missing. We want to capture
2621 * the beneficial effect of buffering by extra pairs just outside rlist,
2622 * while removing the useless pairs that are further away from rlist.
2623 * (Also the buffer could have been set manually not using the estimate.)
2624 * This buffer size is an overestimate.
2625 * We add 10% of the smallest grid sub-cell dimensions.
2626 * Note that the z-size differs per cell and we don't use this,
2627 * so we overestimate.
2628 * With PME, the 10% value gives a buffer that is somewhat larger
2629 * than the effective buffer with a tolerance of 0.005 kJ/mol/ps.
2630 * Smaller tolerances or using RF lead to a smaller effective buffer,
2631 * so 10% gives a safe overestimate.
2633 return eff_1x1_buffer_fac_overest*(minimum_subgrid_size_xy(gridi) +
2634 minimum_subgrid_size_xy(gridj));
2637 /* Clusters at the cut-off only increase rlist by 60% of their size */
2638 static real nbnxn_rlist_inc_outside_fac = 0.6;
2640 /* Due to the cluster size the effective pair-list is longer than
2641 * that of a simple atom pair-list. This function gives the extra distance.
2643 real nbnxn_get_rlist_effective_inc(int cluster_size_j, real atom_density)
2646 real vol_inc_i, vol_inc_j;
2648 /* We should get this from the setup, but currently it's the same for
2649 * all setups, including GPUs.
2651 cluster_size_i = NBNXN_CPU_CLUSTER_I_SIZE;
2653 vol_inc_i = (cluster_size_i - 1)/atom_density;
2654 vol_inc_j = (cluster_size_j - 1)/atom_density;
2656 return nbnxn_rlist_inc_outside_fac*std::cbrt(vol_inc_i + vol_inc_j);
2659 /* Estimates the interaction volume^2 for non-local interactions */
2660 static real nonlocal_vol2(const struct gmx_domdec_zones_t *zones, const rvec ls, real r)
2668 /* Here we simply add up the volumes of 1, 2 or 3 1D decomposition
2669 * not home interaction volume^2. As these volumes are not additive,
2670 * this is an overestimate, but it would only be significant in the limit
2671 * of small cells, where we anyhow need to split the lists into
2672 * as small parts as possible.
2675 for (int z = 0; z < zones->n; z++)
2677 if (zones->shift[z][XX] + zones->shift[z][YY] + zones->shift[z][ZZ] == 1)
2682 for (int d = 0; d < DIM; d++)
2684 if (zones->shift[z][d] == 0)
2688 za *= zones->size[z].x1[d] - zones->size[z].x0[d];
2692 /* 4 octants of a sphere */
2693 vold_est = 0.25*M_PI*r*r*r*r;
2694 /* 4 quarter pie slices on the edges */
2695 vold_est += 4*cl*M_PI/6.0*r*r*r;
2696 /* One rectangular volume on a face */
2697 vold_est += ca*0.5*r*r;
2699 vol2_est_tot += vold_est*za;
2703 return vol2_est_tot;
2706 /* Estimates the average size of a full j-list for super/sub setup */
2707 static void get_nsubpair_target(const nbnxn_search *nbs,
2710 int min_ci_balanced,
2711 int *nsubpair_target,
2712 float *nsubpair_tot_est)
2714 /* The target value of 36 seems to be the optimum for Kepler.
2715 * Maxwell is less sensitive to the exact value.
2717 const int nsubpair_target_min = 36;
2718 const nbnxn_grid_t *grid;
2720 real r_eff_sup, vol_est, nsp_est, nsp_est_nl;
2722 grid = &nbs->grid[0];
2724 /* We don't need to balance list sizes if:
2725 * - We didn't request balancing.
2726 * - The number of grid cells >= the number of lists requested,
2727 * since we will always generate at least #cells lists.
2728 * - We don't have any cells, since then there won't be any lists.
2730 if (min_ci_balanced <= 0 || grid->nc >= min_ci_balanced || grid->nc == 0)
2732 /* nsubpair_target==0 signals no balancing */
2733 *nsubpair_target = 0;
2734 *nsubpair_tot_est = 0;
2739 ls[XX] = (grid->c1[XX] - grid->c0[XX])/(grid->numCells[XX]*c_gpuNumClusterPerCellX);
2740 ls[YY] = (grid->c1[YY] - grid->c0[YY])/(grid->numCells[YY]*c_gpuNumClusterPerCellY);
2741 ls[ZZ] = grid->na_c/(grid->atom_density*ls[XX]*ls[YY]);
2743 /* The average length of the diagonal of a sub cell */
2744 real diagonal = std::sqrt(ls[XX]*ls[XX] + ls[YY]*ls[YY] + ls[ZZ]*ls[ZZ]);
2746 /* The formulas below are a heuristic estimate of the average nsj per si*/
2747 r_eff_sup = rlist + nbnxn_rlist_inc_outside_fac*gmx::square((grid->na_c - 1.0)/grid->na_c)*0.5*diagonal;
2749 if (!nbs->DomDec || nbs->zones->n == 1)
2756 gmx::square(grid->atom_density/grid->na_c)*
2757 nonlocal_vol2(nbs->zones, ls, r_eff_sup);
2762 /* Sub-cell interacts with itself */
2763 vol_est = ls[XX]*ls[YY]*ls[ZZ];
2764 /* 6/2 rectangular volume on the faces */
2765 vol_est += (ls[XX]*ls[YY] + ls[XX]*ls[ZZ] + ls[YY]*ls[ZZ])*r_eff_sup;
2766 /* 12/2 quarter pie slices on the edges */
2767 vol_est += 2*(ls[XX] + ls[YY] + ls[ZZ])*0.25*M_PI*gmx::square(r_eff_sup);
2768 /* 4 octants of a sphere */
2769 vol_est += 0.5*4.0/3.0*M_PI*gmx::power3(r_eff_sup);
2771 /* Estimate the number of cluster pairs as the local number of
2772 * clusters times the volume they interact with times the density.
2774 nsp_est = grid->nsubc_tot*vol_est*grid->atom_density/grid->na_c;
2776 /* Subtract the non-local pair count */
2777 nsp_est -= nsp_est_nl;
2779 /* For small cut-offs nsp_est will be an underesimate.
2780 * With DD nsp_est_nl is an overestimate so nsp_est can get negative.
2781 * So to avoid too small or negative nsp_est we set a minimum of
2782 * all cells interacting with all 3^3 direct neighbors (3^3-1)/2+1=14.
2783 * This might be a slight overestimate for small non-periodic groups of
2784 * atoms as will occur for a local domain with DD, but for small
2785 * groups of atoms we'll anyhow be limited by nsubpair_target_min,
2786 * so this overestimation will not matter.
2788 nsp_est = std::max(nsp_est, grid->nsubc_tot*14._real);
2792 fprintf(debug, "nsp_est local %5.1f non-local %5.1f\n",
2793 nsp_est, nsp_est_nl);
2798 nsp_est = nsp_est_nl;
2801 /* Thus the (average) maximum j-list size should be as follows.
2802 * Since there is overhead, we shouldn't make the lists too small
2803 * (and we can't chop up j-groups) so we use a minimum target size of 36.
2805 *nsubpair_target = std::max(nsubpair_target_min,
2806 roundToInt(nsp_est/min_ci_balanced));
2807 *nsubpair_tot_est = static_cast<int>(nsp_est);
2811 fprintf(debug, "nbl nsp estimate %.1f, nsubpair_target %d\n",
2812 nsp_est, *nsubpair_target);
2816 /* Debug list print function */
2817 static void print_nblist_ci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
2819 for (int i = 0; i < nbl->nci; i++)
2821 fprintf(fp, "ci %4d shift %2d ncj %3d\n",
2822 nbl->ci[i].ci, nbl->ci[i].shift,
2823 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start);
2825 for (int j = nbl->ci[i].cj_ind_start; j < nbl->ci[i].cj_ind_end; j++)
2827 fprintf(fp, " cj %5d imask %x\n",
2834 /* Debug list print function */
2835 static void print_nblist_sci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
2837 for (int i = 0; i < nbl->nsci; i++)
2839 fprintf(fp, "ci %4d shift %2d ncj4 %2d\n",
2840 nbl->sci[i].sci, nbl->sci[i].shift,
2841 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start);
2844 for (int j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
2846 for (int j = 0; j < c_nbnxnGpuJgroupSize; j++)
2848 fprintf(fp, " sj %5d imask %x\n",
2850 nbl->cj4[j4].imei[0].imask);
2851 for (int si = 0; si < c_gpuNumClusterPerCell; si++)
2853 if (nbl->cj4[j4].imei[0].imask & (1U << (j*c_gpuNumClusterPerCell + si)))
2860 fprintf(fp, "ci %4d shift %2d ncj4 %2d ncp %3d\n",
2861 nbl->sci[i].sci, nbl->sci[i].shift,
2862 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start,
2867 /* Combine pair lists *nbl generated on multiple threads nblc */
2868 static void combine_nblists(int nnbl, nbnxn_pairlist_t **nbl,
2869 nbnxn_pairlist_t *nblc)
2871 int nsci, ncj4, nexcl;
2875 gmx_incons("combine_nblists does not support simple lists");
2880 nexcl = nblc->nexcl;
2881 for (int i = 0; i < nnbl; i++)
2883 nsci += nbl[i]->nsci;
2884 ncj4 += nbl[i]->ncj4;
2885 nexcl += nbl[i]->nexcl;
2888 if (nsci > nblc->sci_nalloc)
2890 nb_realloc_sci(nblc, nsci);
2892 if (ncj4 > nblc->cj4_nalloc)
2894 nblc->cj4_nalloc = over_alloc_small(ncj4);
2895 nbnxn_realloc_void(reinterpret_cast<void **>(&nblc->cj4),
2896 nblc->ncj4*sizeof(*nblc->cj4),
2897 nblc->cj4_nalloc*sizeof(*nblc->cj4),
2898 nblc->alloc, nblc->free);
2900 if (nexcl > nblc->excl_nalloc)
2902 nblc->excl_nalloc = over_alloc_small(nexcl);
2903 nbnxn_realloc_void(reinterpret_cast<void **>(&nblc->excl),
2904 nblc->nexcl*sizeof(*nblc->excl),
2905 nblc->excl_nalloc*sizeof(*nblc->excl),
2906 nblc->alloc, nblc->free);
2909 /* Each thread should copy its own data to the combined arrays,
2910 * as otherwise data will go back and forth between different caches.
2912 #if GMX_OPENMP && !(defined __clang_analyzer__)
2913 int nthreads = gmx_omp_nthreads_get(emntPairsearch);
2916 #pragma omp parallel for num_threads(nthreads) schedule(static)
2917 for (int n = 0; n < nnbl; n++)
2924 const nbnxn_pairlist_t *nbli;
2926 /* Determine the offset in the combined data for our thread */
2927 sci_offset = nblc->nsci;
2928 cj4_offset = nblc->ncj4;
2929 excl_offset = nblc->nexcl;
2931 for (int i = 0; i < n; i++)
2933 sci_offset += nbl[i]->nsci;
2934 cj4_offset += nbl[i]->ncj4;
2935 excl_offset += nbl[i]->nexcl;
2940 for (int i = 0; i < nbli->nsci; i++)
2942 nblc->sci[sci_offset+i] = nbli->sci[i];
2943 nblc->sci[sci_offset+i].cj4_ind_start += cj4_offset;
2944 nblc->sci[sci_offset+i].cj4_ind_end += cj4_offset;
2947 for (int j4 = 0; j4 < nbli->ncj4; j4++)
2949 nblc->cj4[cj4_offset+j4] = nbli->cj4[j4];
2950 nblc->cj4[cj4_offset+j4].imei[0].excl_ind += excl_offset;
2951 nblc->cj4[cj4_offset+j4].imei[1].excl_ind += excl_offset;
2954 for (int j4 = 0; j4 < nbli->nexcl; j4++)
2956 nblc->excl[excl_offset+j4] = nbli->excl[j4];
2959 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
2962 for (int n = 0; n < nnbl; n++)
2964 nblc->nsci += nbl[n]->nsci;
2965 nblc->ncj4 += nbl[n]->ncj4;
2966 nblc->nci_tot += nbl[n]->nci_tot;
2967 nblc->nexcl += nbl[n]->nexcl;
2971 static void balance_fep_lists(const nbnxn_search *nbs,
2972 nbnxn_pairlist_set_t *nbl_lists)
2975 int nri_tot, nrj_tot, nrj_target;
2979 nnbl = nbl_lists->nnbl;
2983 /* Nothing to balance */
2987 /* Count the total i-lists and pairs */
2990 for (int th = 0; th < nnbl; th++)
2992 nri_tot += nbl_lists->nbl_fep[th]->nri;
2993 nrj_tot += nbl_lists->nbl_fep[th]->nrj;
2996 nrj_target = (nrj_tot + nnbl - 1)/nnbl;
2998 assert(gmx_omp_nthreads_get(emntNonbonded) == nnbl);
3000 #pragma omp parallel for schedule(static) num_threads(nnbl)
3001 for (int th = 0; th < nnbl; th++)
3005 t_nblist *nbl = nbs->work[th].nbl_fep.get();
3007 /* Note that here we allocate for the total size, instead of
3008 * a per-thread esimate (which is hard to obtain).
3010 if (nri_tot > nbl->maxnri)
3012 nbl->maxnri = over_alloc_large(nri_tot);
3013 reallocate_nblist(nbl);
3015 if (nri_tot > nbl->maxnri || nrj_tot > nbl->maxnrj)
3017 nbl->maxnrj = over_alloc_small(nrj_tot);
3018 srenew(nbl->jjnr, nbl->maxnrj);
3019 srenew(nbl->excl_fep, nbl->maxnrj);
3022 clear_pairlist_fep(nbl);
3024 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
3027 /* Loop over the source lists and assign and copy i-entries */
3029 nbld = nbs->work[th_dest].nbl_fep.get();
3030 for (int th = 0; th < nnbl; th++)
3034 nbls = nbl_lists->nbl_fep[th];
3036 for (int i = 0; i < nbls->nri; i++)
3040 /* The number of pairs in this i-entry */
3041 nrj = nbls->jindex[i+1] - nbls->jindex[i];
3043 /* Decide if list th_dest is too large and we should procede
3044 * to the next destination list.
3046 if (th_dest+1 < nnbl && nbld->nrj > 0 &&
3047 nbld->nrj + nrj - nrj_target > nrj_target - nbld->nrj)
3050 nbld = nbs->work[th_dest].nbl_fep.get();
3053 nbld->iinr[nbld->nri] = nbls->iinr[i];
3054 nbld->gid[nbld->nri] = nbls->gid[i];
3055 nbld->shift[nbld->nri] = nbls->shift[i];
3057 for (int j = nbls->jindex[i]; j < nbls->jindex[i+1]; j++)
3059 nbld->jjnr[nbld->nrj] = nbls->jjnr[j];
3060 nbld->excl_fep[nbld->nrj] = nbls->excl_fep[j];
3064 nbld->jindex[nbld->nri] = nbld->nrj;
3068 /* Swap the list pointers */
3069 for (int th = 0; th < nnbl; th++)
3071 t_nblist *nbl_tmp = nbs->work[th].nbl_fep.release();
3072 nbs->work[th].nbl_fep.reset(nbl_lists->nbl_fep[th]);
3073 nbl_lists->nbl_fep[th] = nbl_tmp;
3077 fprintf(debug, "nbl_fep[%d] nri %4d nrj %4d\n",
3079 nbl_lists->nbl_fep[th]->nri,
3080 nbl_lists->nbl_fep[th]->nrj);
3085 /* Returns the next ci to be processes by our thread */
3086 static gmx_bool next_ci(const nbnxn_grid_t *grid,
3087 int nth, int ci_block,
3088 int *ci_x, int *ci_y,
3094 if (*ci_b == ci_block)
3096 /* Jump to the next block assigned to this task */
3097 *ci += (nth - 1)*ci_block;
3101 if (*ci >= grid->nc)
3106 while (*ci >= grid->cxy_ind[*ci_x*grid->numCells[YY] + *ci_y + 1])
3109 if (*ci_y == grid->numCells[YY])
3119 /* Returns the distance^2 for which we put cell pairs in the list
3120 * without checking atom pair distances. This is usually < rlist^2.
3122 static float boundingbox_only_distance2(const nbnxn_grid_t *gridi,
3123 const nbnxn_grid_t *gridj,
3127 /* If the distance between two sub-cell bounding boxes is less
3128 * than this distance, do not check the distance between
3129 * all particle pairs in the sub-cell, since then it is likely
3130 * that the box pair has atom pairs within the cut-off.
3131 * We use the nblist cut-off minus 0.5 times the average x/y diagonal
3132 * spacing of the sub-cells. Around 40% of the checked pairs are pruned.
3133 * Using more than 0.5 gains at most 0.5%.
3134 * If forces are calculated more than twice, the performance gain
3135 * in the force calculation outweighs the cost of checking.
3136 * Note that with subcell lists, the atom-pair distance check
3137 * is only performed when only 1 out of 8 sub-cells in within range,
3138 * this is because the GPU is much faster than the cpu.
3143 bbx = 0.5*(gridi->cellSize[XX] + gridj->cellSize[XX]);
3144 bby = 0.5*(gridi->cellSize[YY] + gridj->cellSize[YY]);
3147 bbx /= c_gpuNumClusterPerCellX;
3148 bby /= c_gpuNumClusterPerCellY;
3151 rbb2 = std::max(0.0, rlist - 0.5*std::sqrt(bbx*bbx + bby*bby));
3157 return (float)((1+GMX_FLOAT_EPS)*rbb2);
3161 static int get_ci_block_size(const nbnxn_grid_t *gridi,
3162 gmx_bool bDomDec, int nth)
3164 const int ci_block_enum = 5;
3165 const int ci_block_denom = 11;
3166 const int ci_block_min_atoms = 16;
3169 /* Here we decide how to distribute the blocks over the threads.
3170 * We use prime numbers to try to avoid that the grid size becomes
3171 * a multiple of the number of threads, which would lead to some
3172 * threads getting "inner" pairs and others getting boundary pairs,
3173 * which in turns will lead to load imbalance between threads.
3174 * Set the block size as 5/11/ntask times the average number of cells
3175 * in a y,z slab. This should ensure a quite uniform distribution
3176 * of the grid parts of the different thread along all three grid
3177 * zone boundaries with 3D domain decomposition. At the same time
3178 * the blocks will not become too small.
3180 ci_block = (gridi->nc*ci_block_enum)/(ci_block_denom*gridi->numCells[XX]*nth);
3182 /* Ensure the blocks are not too small: avoids cache invalidation */
3183 if (ci_block*gridi->na_sc < ci_block_min_atoms)
3185 ci_block = (ci_block_min_atoms + gridi->na_sc - 1)/gridi->na_sc;
3188 /* Without domain decomposition
3189 * or with less than 3 blocks per task, divide in nth blocks.
3191 if (!bDomDec || nth*3*ci_block > gridi->nc)
3193 ci_block = (gridi->nc + nth - 1)/nth;
3196 if (ci_block > 1 && (nth - 1)*ci_block >= gridi->nc)
3198 /* Some threads have no work. Although reducing the block size
3199 * does not decrease the block count on the first few threads,
3200 * with GPUs better mixing of "upper" cells that have more empty
3201 * clusters results in a somewhat lower max load over all threads.
3202 * Without GPUs the regime of so few atoms per thread is less
3203 * performance relevant, but with 8-wide SIMD the same reasoning
3204 * applies, since the pair list uses 4 i-atom "sub-clusters".
3212 /* Returns the number of bits to right-shift a cluster index to obtain
3213 * the corresponding force buffer flag index.
3215 static int getBufferFlagShift(int numAtomsPerCluster)
3217 int bufferFlagShift = 0;
3218 while ((numAtomsPerCluster << bufferFlagShift) < NBNXN_BUFFERFLAG_SIZE)
3223 return bufferFlagShift;
3226 /* Generates the part of pair-list nbl assigned to our thread */
3227 static void nbnxn_make_pairlist_part(const nbnxn_search *nbs,
3228 const nbnxn_grid_t *gridi,
3229 const nbnxn_grid_t *gridj,
3230 nbnxn_search_work_t *work,
3231 const nbnxn_atomdata_t *nbat,
3232 const t_blocka &exclusions,
3236 gmx_bool bFBufferFlag,
3239 float nsubpair_tot_est,
3241 nbnxn_pairlist_t *nbl,
3246 real rlist2, rl_fep2 = 0;
3248 int ci_b, ci, ci_x, ci_y, ci_xy, cj;
3252 real bx0, bx1, by0, by1, bz0, bz1;
3254 real d2cx, d2z, d2z_cx, d2z_cy, d2zx, d2zxy, d2xy;
3255 int cxf, cxl, cyf, cyf_x, cyl;
3256 int numDistanceChecks;
3257 int gridi_flag_shift = 0, gridj_flag_shift = 0;
3258 gmx_bitmask_t *gridj_flag = nullptr;
3259 int ncj_old_i, ncj_old_j;
3261 nbs_cycle_start(&work->cc[enbsCCsearch]);
3263 if (gridj->bSimple != nbl->bSimple || gridi->bSimple != nbl->bSimple)
3265 gmx_incons("Grid incompatible with pair-list");
3269 nbl->na_sc = gridj->na_sc;
3270 nbl->na_ci = gridj->na_c;
3271 nbl->na_cj = nbnxn_kernel_to_cluster_j_size(nb_kernel_type);
3272 na_cj_2log = get_2log(nbl->na_cj);
3278 /* Determine conversion of clusters to flag blocks */
3279 gridi_flag_shift = getBufferFlagShift(nbl->na_ci);
3280 gridj_flag_shift = getBufferFlagShift(nbl->na_cj);
3282 gridj_flag = work->buffer_flags.flag;
3285 copy_mat(nbs->box, box);
3287 rlist2 = nbl->rlist*nbl->rlist;
3289 if (nbs->bFEP && !nbl->bSimple)
3291 /* Determine an atom-pair list cut-off distance for FEP atom pairs.
3292 * We should not simply use rlist, since then we would not have
3293 * the small, effective buffering of the NxN lists.
3294 * The buffer is on overestimate, but the resulting cost for pairs
3295 * beyond rlist is neglible compared to the FEP pairs within rlist.
3297 rl_fep2 = nbl->rlist + effective_buffer_1x1_vs_MxN(gridi, gridj);
3301 fprintf(debug, "nbl_fep atom-pair rlist %f\n", rl_fep2);
3303 rl_fep2 = rl_fep2*rl_fep2;
3306 rbb2 = boundingbox_only_distance2(gridi, gridj, nbl->rlist, nbl->bSimple);
3310 fprintf(debug, "nbl bounding box only distance %f\n", std::sqrt(rbb2));
3313 /* Set the shift range */
3314 for (int d = 0; d < DIM; d++)
3316 /* Check if we need periodicity shifts.
3317 * Without PBC or with domain decomposition we don't need them.
3319 if (d >= ePBC2npbcdim(nbs->ePBC) || nbs->dd_dim[d])
3325 const real listRangeCellToCell = listRangeForGridCellToGridCell(rlist, *gridi, *gridj);
3327 box[XX][XX] - fabs(box[YY][XX]) - fabs(box[ZZ][XX]) < listRangeCellToCell)
3337 const bool bSimple = nbl->bSimple;
3338 gmx::ArrayRef<const nbnxn_bb_t> bb_i;
3340 gmx::ArrayRef<const float> pbb_i;
3350 /* We use the normal bounding box format for both grid types */
3353 gmx::ArrayRef<const float> bbcz_i = gridi->bbcz;
3354 gmx::ArrayRef<const int> flags_i = gridi->flags;
3355 gmx::ArrayRef<const float> bbcz_j = gridj->bbcz;
3356 int cell0_i = gridi->cell0;
3360 fprintf(debug, "nbl nc_i %d col.av. %.1f ci_block %d\n",
3361 gridi->nc, gridi->nc/static_cast<double>(gridi->numCells[XX]*gridi->numCells[YY]), ci_block);
3364 numDistanceChecks = 0;
3366 const real listRangeBBToJCell2 = gmx::square(listRangeForBoundingBoxToGridCell(rlist, *gridj));
3368 /* Initially ci_b and ci to 1 before where we want them to start,
3369 * as they will both be incremented in next_ci.
3372 ci = th*ci_block - 1;
3375 while (next_ci(gridi, nth, ci_block, &ci_x, &ci_y, &ci_b, &ci))
3377 if (bSimple && flags_i[ci] == 0)
3382 ncj_old_i = nbl->ncj;
3385 if (gridj != gridi && shp[XX] == 0)
3389 bx1 = bb_i[ci].upper[BB_X];
3393 bx1 = gridi->c0[XX] + (ci_x+1)*gridi->cellSize[XX];
3395 if (bx1 < gridj->c0[XX])
3397 d2cx = gmx::square(gridj->c0[XX] - bx1);
3399 if (d2cx >= listRangeBBToJCell2)
3406 ci_xy = ci_x*gridi->numCells[YY] + ci_y;
3408 /* Loop over shift vectors in three dimensions */
3409 for (int tz = -shp[ZZ]; tz <= shp[ZZ]; tz++)
3411 shz = tz*box[ZZ][ZZ];
3413 bz0 = bbcz_i[ci*NNBSBB_D ] + shz;
3414 bz1 = bbcz_i[ci*NNBSBB_D+1] + shz;
3422 d2z = gmx::square(bz1);
3426 d2z = gmx::square(bz0 - box[ZZ][ZZ]);
3429 d2z_cx = d2z + d2cx;
3431 if (d2z_cx >= rlist2)
3436 bz1_frac = bz1/(gridi->cxy_ind[ci_xy+1] - gridi->cxy_ind[ci_xy]);
3441 /* The check with bz1_frac close to or larger than 1 comes later */
3443 for (int ty = -shp[YY]; ty <= shp[YY]; ty++)
3445 shy = ty*box[YY][YY] + tz*box[ZZ][YY];
3449 by0 = bb_i[ci].lower[BB_Y] + shy;
3450 by1 = bb_i[ci].upper[BB_Y] + shy;
3454 by0 = gridi->c0[YY] + (ci_y )*gridi->cellSize[YY] + shy;
3455 by1 = gridi->c0[YY] + (ci_y+1)*gridi->cellSize[YY] + shy;
3458 get_cell_range<YY>(by0, by1,
3469 if (by1 < gridj->c0[YY])
3471 d2z_cy += gmx::square(gridj->c0[YY] - by1);
3473 else if (by0 > gridj->c1[YY])
3475 d2z_cy += gmx::square(by0 - gridj->c1[YY]);
3478 for (int tx = -shp[XX]; tx <= shp[XX]; tx++)
3480 shift = XYZ2IS(tx, ty, tz);
3482 if (c_pbcShiftBackward && gridi == gridj && shift > CENTRAL)
3487 shx = tx*box[XX][XX] + ty*box[YY][XX] + tz*box[ZZ][XX];
3491 bx0 = bb_i[ci].lower[BB_X] + shx;
3492 bx1 = bb_i[ci].upper[BB_X] + shx;
3496 bx0 = gridi->c0[XX] + (ci_x )*gridi->cellSize[XX] + shx;
3497 bx1 = gridi->c0[XX] + (ci_x+1)*gridi->cellSize[XX] + shx;
3500 get_cell_range<XX>(bx0, bx1,
3512 new_ci_entry(nbl, cell0_i+ci, shift, flags_i[ci]);
3516 new_sci_entry(nbl, cell0_i+ci, shift);
3519 if ((!c_pbcShiftBackward || (shift == CENTRAL &&
3523 /* Leave the pairs with i > j.
3524 * x is the major index, so skip half of it.
3531 set_icell_bb_simple(bb_i, ci, shx, shy, shz,
3537 set_icell_bbxxxx_supersub(pbb_i, ci, shx, shy, shz,
3540 set_icell_bb_supersub(bb_i, ci, shx, shy, shz,
3545 nbs->icell_set_x(cell0_i+ci, shx, shy, shz,
3546 nbat->xstride, nbat->x,
3549 for (int cx = cxf; cx <= cxl; cx++)
3552 if (gridj->c0[XX] + cx*gridj->cellSize[XX] > bx1)
3554 d2zx += gmx::square(gridj->c0[XX] + cx*gridj->cellSize[XX] - bx1);
3556 else if (gridj->c0[XX] + (cx+1)*gridj->cellSize[XX] < bx0)
3558 d2zx += gmx::square(gridj->c0[XX] + (cx+1)*gridj->cellSize[XX] - bx0);
3561 if (gridi == gridj &&
3563 (!c_pbcShiftBackward || shift == CENTRAL) &&
3566 /* Leave the pairs with i > j.
3567 * Skip half of y when i and j have the same x.
3576 for (int cy = cyf_x; cy <= cyl; cy++)
3578 const int columnStart = gridj->cxy_ind[cx*gridj->numCells[YY] + cy];
3579 const int columnEnd = gridj->cxy_ind[cx*gridj->numCells[YY] + cy + 1];
3582 if (gridj->c0[YY] + cy*gridj->cellSize[YY] > by1)
3584 d2zxy += gmx::square(gridj->c0[YY] + cy*gridj->cellSize[YY] - by1);
3586 else if (gridj->c0[YY] + (cy+1)*gridj->cellSize[YY] < by0)
3588 d2zxy += gmx::square(gridj->c0[YY] + (cy+1)*gridj->cellSize[YY] - by0);
3590 if (columnStart < columnEnd && d2zxy < listRangeBBToJCell2)
3592 /* To improve efficiency in the common case
3593 * of a homogeneous particle distribution,
3594 * we estimate the index of the middle cell
3595 * in range (midCell). We search down and up
3596 * starting from this index.
3598 * Note that the bbcz_j array contains bounds
3599 * for i-clusters, thus for clusters of 4 atoms.
3600 * For the common case where the j-cluster size
3601 * is 8, we could step with a stride of 2,
3602 * but we do not do this because it would
3603 * complicate this code even more.
3605 int midCell = columnStart + static_cast<int>(bz1_frac*(columnEnd - columnStart));
3606 if (midCell >= columnEnd)
3608 midCell = columnEnd - 1;
3613 /* Find the lowest cell that can possibly
3615 * Check if we hit the bottom of the grid,
3616 * if the j-cell is below the i-cell and if so,
3617 * if it is within range.
3619 int downTestCell = midCell;
3620 while (downTestCell >= columnStart &&
3621 (bbcz_j[downTestCell*NNBSBB_D + 1] >= bz0 ||
3622 d2xy + gmx::square(bbcz_j[downTestCell*NNBSBB_D + 1] - bz0) < rlist2))
3626 int firstCell = downTestCell + 1;
3628 /* Find the highest cell that can possibly
3630 * Check if we hit the top of the grid,
3631 * if the j-cell is above the i-cell and if so,
3632 * if it is within range.
3634 int upTestCell = midCell + 1;
3635 while (upTestCell < columnEnd &&
3636 (bbcz_j[upTestCell*NNBSBB_D] <= bz1 ||
3637 d2xy + gmx::square(bbcz_j[upTestCell*NNBSBB_D] - bz1) < rlist2))
3641 int lastCell = upTestCell - 1;
3643 #define NBNXN_REFCODE 0
3646 /* Simple reference code, for debugging,
3647 * overrides the more complex code above.
3649 firstCell = columnEnd;
3651 for (int k = columnStart; k < columnEnd; k++)
3653 if (d2xy + gmx::square(bbcz_j[k*NNBSBB_D + 1] - bz0) < rlist2 &&
3658 if (d2xy + gmx::square(bbcz_j[k*NNBSBB_D] - bz1) < rlist2 &&
3669 /* We want each atom/cell pair only once,
3670 * only use cj >= ci.
3672 if (!c_pbcShiftBackward || shift == CENTRAL)
3674 firstCell = std::max(firstCell, ci);
3678 if (firstCell <= lastCell)
3680 GMX_ASSERT(firstCell >= columnStart && lastCell < columnEnd, "The range should reside within the current grid column");
3682 /* For f buffer flags with simple lists */
3683 ncj_old_j = nbl->ncj;
3687 /* We have a maximum of 2 j-clusters
3688 * per i-cluster sized cell.
3690 check_cell_list_space_simple(nbl, 2*(lastCell - firstCell + 1));
3694 check_cell_list_space_supersub(nbl, lastCell - firstCell + 1);
3697 switch (nb_kernel_type)
3699 case nbnxnk4x4_PlainC:
3700 makeClusterListSimple(gridj,
3701 nbl, ci, firstCell, lastCell,
3702 (gridi == gridj && shift == CENTRAL),
3705 &numDistanceChecks);
3707 #ifdef GMX_NBNXN_SIMD_4XN
3708 case nbnxnk4xN_SIMD_4xN:
3709 makeClusterListSimd4xn(gridj,
3710 nbl, ci, firstCell, lastCell,
3711 (gridi == gridj && shift == CENTRAL),
3714 &numDistanceChecks);
3717 #ifdef GMX_NBNXN_SIMD_2XNN
3718 case nbnxnk4xN_SIMD_2xNN:
3719 makeClusterListSimd2xnn(gridj,
3720 nbl, ci, firstCell, lastCell,
3721 (gridi == gridj && shift == CENTRAL),
3724 &numDistanceChecks);
3727 case nbnxnk8x8x8_PlainC:
3728 case nbnxnk8x8x8_GPU:
3729 for (cj = firstCell; cj <= lastCell; cj++)
3731 make_cluster_list_supersub(gridi, gridj,
3733 (gridi == gridj && shift == CENTRAL && ci == cj),
3734 nbat->xstride, nbat->x,
3736 &numDistanceChecks);
3741 if (bFBufferFlag && nbl->ncj > ncj_old_j)
3743 int cbf = nbl->cj[ncj_old_j].cj >> gridj_flag_shift;
3744 int cbl = nbl->cj[nbl->ncj-1].cj >> gridj_flag_shift;
3745 for (int cb = cbf; cb <= cbl; cb++)
3747 bitmask_init_bit(&gridj_flag[cb], th);
3751 nbl->ncjInUse += nbl->ncj - ncj_old_j;
3757 /* Set the exclusions for this ci list */
3760 setExclusionsForSimpleIentry(nbs,
3762 shift == CENTRAL && gridi == gridj,
3769 make_fep_list(nbs, nbat, nbl,
3770 shift == CENTRAL && gridi == gridj,
3771 &(nbl->ci[nbl->nci]),
3772 gridi, gridj, nbl_fep);
3777 setExclusionsForGpuIentry(nbs,
3779 shift == CENTRAL && gridi == gridj,
3780 nbl->sci[nbl->nsci],
3785 make_fep_list_supersub(nbs, nbat, nbl,
3786 shift == CENTRAL && gridi == gridj,
3787 &(nbl->sci[nbl->nsci]),
3790 gridi, gridj, nbl_fep);
3794 /* Close this ci list */
3797 close_ci_entry_simple(nbl);
3801 close_ci_entry_supersub(nbl,
3803 progBal, nsubpair_tot_est,
3810 if (bFBufferFlag && nbl->ncj > ncj_old_i)
3812 bitmask_init_bit(&(work->buffer_flags.flag[(gridi->cell0+ci)>>gridi_flag_shift]), th);
3816 work->ndistc = numDistanceChecks;
3818 nbs_cycle_stop(&work->cc[enbsCCsearch]);
3820 GMX_ASSERT(nbl->ncjInUse == nbl->ncj || nbs->bFEP, "Without free-energy all cj pair-list entries should be in use. Note that subsequent code does not make use of the equality, this check is only here to catch bugs");
3824 fprintf(debug, "number of distance checks %d\n", numDistanceChecks);
3828 print_nblist_statistics_simple(debug, nbl, nbs, rlist);
3832 print_nblist_statistics_supersub(debug, nbl, nbs, rlist);
3837 fprintf(debug, "nbl FEP list pairs: %d\n", nbl_fep->nrj);
3842 static void reduce_buffer_flags(const nbnxn_search *nbs,
3844 const nbnxn_buffer_flags_t *dest)
3846 for (int s = 0; s < nsrc; s++)
3848 gmx_bitmask_t * flag = nbs->work[s].buffer_flags.flag;
3850 for (int b = 0; b < dest->nflag; b++)
3852 bitmask_union(&(dest->flag[b]), flag[b]);
3857 static void print_reduction_cost(const nbnxn_buffer_flags_t *flags, int nout)
3859 int nelem, nkeep, ncopy, nred, out;
3860 gmx_bitmask_t mask_0;
3866 bitmask_init_bit(&mask_0, 0);
3867 for (int b = 0; b < flags->nflag; b++)
3869 if (bitmask_is_equal(flags->flag[b], mask_0))
3871 /* Only flag 0 is set, no copy of reduction required */
3875 else if (!bitmask_is_zero(flags->flag[b]))
3878 for (out = 0; out < nout; out++)
3880 if (bitmask_is_set(flags->flag[b], out))
3897 fprintf(debug, "nbnxn reduction: #flag %d #list %d elem %4.2f, keep %4.2f copy %4.2f red %4.2f\n",
3899 nelem/static_cast<double>(flags->nflag),
3900 nkeep/static_cast<double>(flags->nflag),
3901 ncopy/static_cast<double>(flags->nflag),
3902 nred/static_cast<double>(flags->nflag));
3905 /* Copies the list entries from src to dest when cjStart <= *cjGlobal < cjEnd.
3906 * *cjGlobal is updated with the cj count in src.
3907 * When setFlags==true, flag bit t is set in flag for all i and j clusters.
3909 template<bool setFlags>
3910 static void copySelectedListRange(const nbnxn_ci_t * gmx_restrict srcCi,
3911 const nbnxn_pairlist_t * gmx_restrict src,
3912 nbnxn_pairlist_t * gmx_restrict dest,
3913 gmx_bitmask_t *flag,
3914 int iFlagShift, int jFlagShift, int t)
3916 int ncj = srcCi->cj_ind_end - srcCi->cj_ind_start;
3918 if (dest->nci + 1 >= dest->ci_nalloc)
3920 nb_realloc_ci(dest, dest->nci + 1);
3922 check_cell_list_space_simple(dest, ncj);
3924 dest->ci[dest->nci] = *srcCi;
3925 dest->ci[dest->nci].cj_ind_start = dest->ncj;
3926 dest->ci[dest->nci].cj_ind_end = dest->ncj + ncj;
3930 bitmask_init_bit(&flag[srcCi->ci >> iFlagShift], t);
3933 for (int j = srcCi->cj_ind_start; j < srcCi->cj_ind_end; j++)
3935 dest->cj[dest->ncj++] = src->cj[j];
3939 /* NOTE: This is relatively expensive, since this
3940 * operation is done for all elements in the list,
3941 * whereas at list generation this is done only
3942 * once for each flag entry.
3944 bitmask_init_bit(&flag[src->cj[j].cj >> jFlagShift], t);
3951 /* This routine re-balances the pairlists such that all are nearly equally
3952 * sized. Only whole i-entries are moved between lists. These are moved
3953 * between the ends of the lists, such that the buffer reduction cost should
3954 * not change significantly.
3955 * Note that all original reduction flags are currently kept. This can lead
3956 * to reduction of parts of the force buffer that could be avoided. But since
3957 * the original lists are quite balanced, this will only give minor overhead.
3959 static void rebalanceSimpleLists(int numLists,
3960 nbnxn_pairlist_t * const * const srcSet,
3961 nbnxn_pairlist_t **destSet,
3962 gmx::ArrayRef<nbnxn_search_work_t> searchWork)
3965 for (int s = 0; s < numLists; s++)
3967 ncjTotal += srcSet[s]->ncjInUse;
3969 int ncjTarget = (ncjTotal + numLists - 1)/numLists;
3971 #pragma omp parallel num_threads(numLists)
3973 int t = gmx_omp_get_thread_num();
3975 int cjStart = ncjTarget* t;
3976 int cjEnd = ncjTarget*(t + 1);
3978 /* The destination pair-list for task/thread t */
3979 nbnxn_pairlist_t *dest = destSet[t];
3981 clear_pairlist(dest);
3982 dest->bSimple = srcSet[0]->bSimple;
3983 dest->na_ci = srcSet[0]->na_ci;
3984 dest->na_cj = srcSet[0]->na_cj;
3986 /* Note that the flags in the work struct (still) contain flags
3987 * for all entries that are present in srcSet->nbl[t].
3989 gmx_bitmask_t *flag = searchWork[t].buffer_flags.flag;
3991 int iFlagShift = getBufferFlagShift(dest->na_ci);
3992 int jFlagShift = getBufferFlagShift(dest->na_cj);
3995 for (int s = 0; s < numLists && cjGlobal < cjEnd; s++)
3997 const nbnxn_pairlist_t *src = srcSet[s];
3999 if (cjGlobal + src->ncjInUse > cjStart)
4001 for (int i = 0; i < src->nci && cjGlobal < cjEnd; i++)
4003 const nbnxn_ci_t *srcCi = &src->ci[i];
4004 int ncj = srcCi->cj_ind_end - srcCi->cj_ind_start;
4005 if (cjGlobal >= cjStart)
4007 /* If the source list is not our own, we need to set
4008 * extra flags (the template bool parameter).
4012 copySelectedListRange
4015 flag, iFlagShift, jFlagShift, t);
4019 copySelectedListRange
4022 dest, flag, iFlagShift, jFlagShift, t);
4030 cjGlobal += src->ncjInUse;
4034 dest->ncjInUse = dest->ncj;
4038 int ncjTotalNew = 0;
4039 for (int s = 0; s < numLists; s++)
4041 ncjTotalNew += destSet[s]->ncjInUse;
4043 GMX_RELEASE_ASSERT(ncjTotalNew == ncjTotal, "The total size of the lists before and after rebalancing should match");
4047 /* Returns if the pairlists are so imbalanced that it is worth rebalancing. */
4048 static bool checkRebalanceSimpleLists(const nbnxn_pairlist_set_t *listSet)
4050 int numLists = listSet->nnbl;
4053 for (int s = 0; s < numLists; s++)
4055 ncjMax = std::max(ncjMax, listSet->nbl[s]->ncjInUse);
4056 ncjTotal += listSet->nbl[s]->ncjInUse;
4060 fprintf(debug, "Pair-list ncjMax %d ncjTotal %d\n", ncjMax, ncjTotal);
4062 /* The rebalancing adds 3% extra time to the search. Heuristically we
4063 * determined that under common conditions the non-bonded kernel balance
4064 * improvement will outweigh this when the imbalance is more than 3%.
4065 * But this will, obviously, depend on search vs kernel time and nstlist.
4067 const real rebalanceTolerance = 1.03;
4069 return numLists*ncjMax > ncjTotal*rebalanceTolerance;
4072 /* Perform a count (linear) sort to sort the smaller lists to the end.
4073 * This avoids load imbalance on the GPU, as large lists will be
4074 * scheduled and executed first and the smaller lists later.
4075 * Load balancing between multi-processors only happens at the end
4076 * and there smaller lists lead to more effective load balancing.
4077 * The sorting is done on the cj4 count, not on the actual pair counts.
4078 * Not only does this make the sort faster, but it also results in
4079 * better load balancing than using a list sorted on exact load.
4080 * This function swaps the pointer in the pair list to avoid a copy operation.
4082 static void sort_sci(nbnxn_pairlist_t *nbl)
4084 nbnxn_list_work_t *work;
4086 nbnxn_sci_t *sci_sort;
4088 if (nbl->ncj4 <= nbl->nsci)
4090 /* nsci = 0 or all sci have size 1, sorting won't change the order */
4096 /* We will distinguish differences up to double the average */
4097 m = (2*nbl->ncj4)/nbl->nsci;
4099 if (m + 1 > work->sort_nalloc)
4101 work->sort_nalloc = over_alloc_large(m + 1);
4102 srenew(work->sort, work->sort_nalloc);
4105 if (work->sci_sort_nalloc != nbl->sci_nalloc)
4107 work->sci_sort_nalloc = nbl->sci_nalloc;
4108 nbnxn_realloc_void(reinterpret_cast<void **>(&work->sci_sort),
4110 work->sci_sort_nalloc*sizeof(*work->sci_sort),
4111 nbl->alloc, nbl->free);
4114 /* Count the entries of each size */
4115 for (int i = 0; i <= m; i++)
4119 for (int s = 0; s < nbl->nsci; s++)
4121 int i = std::min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start);
4124 /* Calculate the offset for each count */
4127 for (int i = m - 1; i >= 0; i--)
4130 work->sort[i] = work->sort[i + 1] + s0;
4134 /* Sort entries directly into place */
4135 sci_sort = work->sci_sort;
4136 for (int s = 0; s < nbl->nsci; s++)
4138 int i = std::min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start);
4139 sci_sort[work->sort[i]++] = nbl->sci[s];
4142 /* Swap the sci pointers so we use the new, sorted list */
4143 work->sci_sort = nbl->sci;
4144 nbl->sci = sci_sort;
4147 /* Make a local or non-local pair-list, depending on iloc */
4148 void nbnxn_make_pairlist(nbnxn_search *nbs,
4149 nbnxn_atomdata_t *nbat,
4150 const t_blocka *excl,
4152 int min_ci_balanced,
4153 nbnxn_pairlist_set_t *nbl_list,
4158 nbnxn_grid_t *gridi, *gridj;
4160 int nsubpair_target;
4161 float nsubpair_tot_est;
4163 nbnxn_pairlist_t **nbl;
4165 gmx_bool CombineNBLists;
4167 int np_tot, np_noq, np_hlj, nap;
4169 nnbl = nbl_list->nnbl;
4170 nbl = nbl_list->nbl;
4171 CombineNBLists = nbl_list->bCombined;
4175 fprintf(debug, "ns making %d nblists\n", nnbl);
4178 nbat->bUseBufferFlags = (nbat->nout > 1);
4179 /* We should re-init the flags before making the first list */
4180 if (nbat->bUseBufferFlags && LOCAL_I(iloc))
4182 init_buffer_flags(&nbat->buffer_flags, nbat->natoms);
4185 if (nbl_list->bSimple)
4188 switch (nb_kernel_type)
4190 #ifdef GMX_NBNXN_SIMD_4XN
4191 case nbnxnk4xN_SIMD_4xN:
4192 nbs->icell_set_x = icell_set_x_simd_4xn;
4195 #ifdef GMX_NBNXN_SIMD_2XNN
4196 case nbnxnk4xN_SIMD_2xNN:
4197 nbs->icell_set_x = icell_set_x_simd_2xnn;
4201 nbs->icell_set_x = icell_set_x_simple;
4205 /* MSVC 2013 complains about switch statements without case */
4206 nbs->icell_set_x = icell_set_x_simple;
4211 nbs->icell_set_x = icell_set_x_supersub;
4216 /* Only zone (grid) 0 vs 0 */
4223 nzi = nbs->zones->nizone;
4226 if (!nbl_list->bSimple && min_ci_balanced > 0)
4228 get_nsubpair_target(nbs, iloc, rlist, min_ci_balanced,
4229 &nsubpair_target, &nsubpair_tot_est);
4233 nsubpair_target = 0;
4234 nsubpair_tot_est = 0;
4237 /* Clear all pair-lists */
4238 for (int th = 0; th < nnbl; th++)
4240 clear_pairlist(nbl[th]);
4244 clear_pairlist_fep(nbl_list->nbl_fep[th]);
4248 for (int zi = 0; zi < nzi; zi++)
4250 gridi = &nbs->grid[zi];
4252 if (NONLOCAL_I(iloc))
4254 zj0 = nbs->zones->izone[zi].j0;
4255 zj1 = nbs->zones->izone[zi].j1;
4261 for (int zj = zj0; zj < zj1; zj++)
4263 gridj = &nbs->grid[zj];
4267 fprintf(debug, "ns search grid %d vs %d\n", zi, zj);
4270 nbs_cycle_start(&nbs->cc[enbsCCsearch]);
4272 ci_block = get_ci_block_size(gridi, nbs->DomDec, nnbl);
4274 /* With GPU: generate progressively smaller lists for
4275 * load balancing for local only or non-local with 2 zones.
4277 progBal = (LOCAL_I(iloc) || nbs->zones->n <= 2);
4279 #pragma omp parallel for num_threads(nnbl) schedule(static)
4280 for (int th = 0; th < nnbl; th++)
4284 /* Re-init the thread-local work flag data before making
4285 * the first list (not an elegant conditional).
4287 if (nbat->bUseBufferFlags && ((zi == 0 && zj == 0)))
4289 init_buffer_flags(&nbs->work[th].buffer_flags, nbat->natoms);
4292 if (CombineNBLists && th > 0)
4294 clear_pairlist(nbl[th]);
4297 /* Divide the i super cell equally over the nblists */
4298 nbnxn_make_pairlist_part(nbs, gridi, gridj,
4299 &nbs->work[th], nbat, *excl,
4303 nbat->bUseBufferFlags,
4305 progBal, nsubpair_tot_est,
4308 nbl_list->nbl_fep[th]);
4310 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
4312 nbs_cycle_stop(&nbs->cc[enbsCCsearch]);
4317 for (int th = 0; th < nnbl; th++)
4319 inc_nrnb(nrnb, eNR_NBNXN_DIST2, nbs->work[th].ndistc);
4321 if (nbl_list->bSimple)
4323 np_tot += nbl[th]->ncj;
4324 np_noq += nbl[th]->work->ncj_noq;
4325 np_hlj += nbl[th]->work->ncj_hlj;
4329 /* This count ignores potential subsequent pair pruning */
4330 np_tot += nbl[th]->nci_tot;
4333 nap = nbl[0]->na_ci*nbl[0]->na_cj;
4334 nbl_list->natpair_ljq = (np_tot - np_noq)*nap - np_hlj*nap/2;
4335 nbl_list->natpair_lj = np_noq*nap;
4336 nbl_list->natpair_q = np_hlj*nap/2;
4338 if (CombineNBLists && nnbl > 1)
4340 nbs_cycle_start(&nbs->cc[enbsCCcombine]);
4342 combine_nblists(nnbl-1, nbl+1, nbl[0]);
4344 nbs_cycle_stop(&nbs->cc[enbsCCcombine]);
4349 if (nbl_list->bSimple)
4351 if (nnbl > 1 && checkRebalanceSimpleLists(nbl_list))
4353 rebalanceSimpleLists(nbl_list->nnbl, nbl_list->nbl, nbl_list->nbl_work, nbs->work);
4355 /* Swap the pointer of the sets of pair lists */
4356 nbnxn_pairlist_t **tmp = nbl_list->nbl;
4357 nbl_list->nbl = nbl_list->nbl_work;
4358 nbl_list->nbl_work = tmp;
4363 /* Sort the entries on size, large ones first */
4364 if (CombineNBLists || nnbl == 1)
4370 #pragma omp parallel for num_threads(nnbl) schedule(static)
4371 for (int th = 0; th < nnbl; th++)
4377 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
4382 if (nbat->bUseBufferFlags)
4384 reduce_buffer_flags(nbs, nbl_list->nnbl, &nbat->buffer_flags);
4389 /* Balance the free-energy lists over all the threads */
4390 balance_fep_lists(nbs, nbl_list);
4393 /* This is a fresh list, so not pruned, stored using ci and nci.
4394 * ciOuter and nciOuter are invalid at this point.
4396 GMX_ASSERT(nbl_list->nbl[0]->nciOuter == -1, "nciOuter should have been set to -1 to signal that it is invalid");
4398 /* Special performance logging stuff (env.var. GMX_NBNXN_CYCLE) */
4401 nbs->search_count++;
4403 if (nbs->print_cycles &&
4404 (!nbs->DomDec || !LOCAL_I(iloc)) &&
4405 nbs->search_count % 100 == 0)
4407 nbs_cycle_print(stderr, nbs);
4410 /* If we have more than one list, they either got rebalancing (CPU)
4411 * or combined (GPU), so we should dump the final result to debug.
4413 if (debug && nbl_list->nnbl > 1)
4415 if (nbl_list->bSimple)
4417 for (int t = 0; t < nbl_list->nnbl; t++)
4419 print_nblist_statistics_simple(debug, nbl_list->nbl[t], nbs, rlist);
4424 print_nblist_statistics_supersub(debug, nbl_list->nbl[0], nbs, rlist);
4432 if (nbl_list->bSimple)
4434 for (int t = 0; t < nbl_list->nnbl; t++)
4436 print_nblist_ci_cj(debug, nbl_list->nbl[t]);
4441 print_nblist_sci_cj(debug, nbl_list->nbl[0]);
4445 if (nbat->bUseBufferFlags)
4447 print_reduction_cost(&nbat->buffer_flags, nbl_list->nnbl);
4452 void nbnxnPrepareListForDynamicPruning(nbnxn_pairlist_set_t *listSet)
4454 /* TODO: Restructure the lists so we have actual outer and inner
4455 * list objects so we can set a single pointer instead of
4456 * swapping several pointers.
4459 for (int i = 0; i < listSet->nnbl; i++)
4461 /* The search produced a list in ci/cj.
4462 * Swap the list pointers so we get the outer list is ciOuter,cjOuter
4463 * and we can prune that to get an inner list in ci/cj.
4465 nbnxn_pairlist_t *list = listSet->nbl[i];
4466 list->nciOuter = list->nci;
4468 nbnxn_ci_t *ciTmp = list->ciOuter;
4469 list->ciOuter = list->ci;
4472 nbnxn_cj_t *cjTmp = list->cjOuter;
4473 list->cjOuter = list->cj;
4476 /* Signal that this inner list is currently invalid */