2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018,2019, 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.
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/ns.h"
55 #include "gromacs/mdtypes/group.h"
56 #include "gromacs/mdtypes/md_enums.h"
57 #include "gromacs/nbnxm/atomdata.h"
58 #include "gromacs/nbnxm/gpu_data_mgmt.h"
59 #include "gromacs/nbnxm/nbnxm.h"
60 #include "gromacs/nbnxm/nbnxm_geometry.h"
61 #include "gromacs/nbnxm/nbnxm_simd.h"
62 #include "gromacs/nbnxm/pairlistset.h"
63 #include "gromacs/pbcutil/ishift.h"
64 #include "gromacs/pbcutil/pbc.h"
65 #include "gromacs/simd/simd.h"
66 #include "gromacs/simd/vector_operations.h"
67 #include "gromacs/topology/block.h"
68 #include "gromacs/utility/exceptions.h"
69 #include "gromacs/utility/fatalerror.h"
70 #include "gromacs/utility/gmxomp.h"
71 #include "gromacs/utility/smalloc.h"
75 #include "pairlistwork.h"
77 using namespace gmx; // TODO: Remove when this file is moved into gmx namespace
79 using BoundingBox = Nbnxm::BoundingBox; // TODO: Remove when refactoring this file
80 using BoundingBox1D = Nbnxm::BoundingBox1D; // TODO: Remove when refactoring this file
82 using Grid = Nbnxm::Grid; // TODO: Remove when refactoring this file
84 // Convience alias for partial Nbnxn namespace usage
85 using InteractionLocality = Nbnxm::InteractionLocality;
87 /* We shift the i-particles backward for PBC.
88 * This leads to more conditionals than shifting forward.
89 * We do this to get more balanced pair lists.
91 constexpr bool c_pbcShiftBackward = true;
94 static void nbs_cycle_clear(nbnxn_cycle_t *cc)
96 for (int i = 0; i < enbsCCnr; i++)
103 static double Mcyc_av(const nbnxn_cycle_t *cc)
105 return static_cast<double>(cc->c)*1e-6/cc->count;
108 static void nbs_cycle_print(FILE *fp, const nbnxn_search *nbs)
111 fprintf(fp, "ns %4d grid %4.1f search %4.1f red.f %5.3f",
112 nbs->cc[enbsCCgrid].count,
113 Mcyc_av(&nbs->cc[enbsCCgrid]),
114 Mcyc_av(&nbs->cc[enbsCCsearch]),
115 Mcyc_av(&nbs->cc[enbsCCreducef]));
117 if (nbs->work.size() > 1)
119 if (nbs->cc[enbsCCcombine].count > 0)
121 fprintf(fp, " comb %5.2f",
122 Mcyc_av(&nbs->cc[enbsCCcombine]));
124 fprintf(fp, " s. th");
125 for (const nbnxn_search_work_t &work : nbs->work)
127 fprintf(fp, " %4.1f",
128 Mcyc_av(&work.cc[enbsCCsearch]));
134 /* Layout for the nonbonded NxN pair lists */
135 enum class NbnxnLayout
137 NoSimd4x4, // i-cluster size 4, j-cluster size 4
138 Simd4xN, // i-cluster size 4, j-cluster size SIMD width
139 Simd2xNN, // i-cluster size 4, j-cluster size half SIMD width
140 Gpu8x8x8 // i-cluster size 8, j-cluster size 8 + super-clustering
144 /* Returns the j-cluster size */
145 template <NbnxnLayout layout>
146 static constexpr int jClusterSize()
148 static_assert(layout == NbnxnLayout::NoSimd4x4 || layout == NbnxnLayout::Simd4xN || layout == NbnxnLayout::Simd2xNN, "Currently jClusterSize only supports CPU layouts");
150 return layout == NbnxnLayout::Simd4xN ? GMX_SIMD_REAL_WIDTH : (layout == NbnxnLayout::Simd2xNN ? GMX_SIMD_REAL_WIDTH/2 : c_nbnxnCpuIClusterSize);
153 /*! \brief Returns the j-cluster index given the i-cluster index.
155 * \tparam jClusterSize The number of atoms in a j-cluster
156 * \tparam jSubClusterIndex The j-sub-cluster index (0/1), used when size(j-cluster) < size(i-cluster)
157 * \param[in] ci The i-cluster index
159 template <int jClusterSize, int jSubClusterIndex>
160 static inline int cjFromCi(int ci)
162 static_assert(jClusterSize == c_nbnxnCpuIClusterSize/2 || jClusterSize == c_nbnxnCpuIClusterSize || jClusterSize == c_nbnxnCpuIClusterSize*2, "Only j-cluster sizes 2, 4 and 8 are currently implemented");
164 static_assert(jSubClusterIndex == 0 || jSubClusterIndex == 1,
165 "Only sub-cluster indices 0 and 1 are supported");
167 if (jClusterSize == c_nbnxnCpuIClusterSize/2)
169 if (jSubClusterIndex == 0)
175 return ((ci + 1) << 1) - 1;
178 else if (jClusterSize == c_nbnxnCpuIClusterSize)
188 /*! \brief Returns the j-cluster index given the i-cluster index.
190 * \tparam layout The pair-list layout
191 * \tparam jSubClusterIndex The j-sub-cluster index (0/1), used when size(j-cluster) < size(i-cluster)
192 * \param[in] ci The i-cluster index
194 template <NbnxnLayout layout, int jSubClusterIndex>
195 static inline int cjFromCi(int ci)
197 constexpr int clusterSize = jClusterSize<layout>();
199 return cjFromCi<clusterSize, jSubClusterIndex>(ci);
202 /* Returns the nbnxn coordinate data index given the i-cluster index */
203 template <NbnxnLayout layout>
204 static inline int xIndexFromCi(int ci)
206 constexpr int clusterSize = jClusterSize<layout>();
208 static_assert(clusterSize == c_nbnxnCpuIClusterSize/2 || clusterSize == c_nbnxnCpuIClusterSize || clusterSize == c_nbnxnCpuIClusterSize*2, "Only j-cluster sizes 2, 4 and 8 are currently implemented");
210 if (clusterSize <= c_nbnxnCpuIClusterSize)
212 /* Coordinates are stored packed in groups of 4 */
217 /* Coordinates packed in 8, i-cluster size is half the packing width */
218 return (ci >> 1)*STRIDE_P8 + (ci & 1)*(c_packX8 >> 1);
222 /* Returns the nbnxn coordinate data index given the j-cluster index */
223 template <NbnxnLayout layout>
224 static inline int xIndexFromCj(int cj)
226 constexpr int clusterSize = jClusterSize<layout>();
228 static_assert(clusterSize == c_nbnxnCpuIClusterSize/2 || clusterSize == c_nbnxnCpuIClusterSize || clusterSize == c_nbnxnCpuIClusterSize*2, "Only j-cluster sizes 2, 4 and 8 are currently implemented");
230 if (clusterSize == c_nbnxnCpuIClusterSize/2)
232 /* Coordinates are stored packed in groups of 4 */
233 return (cj >> 1)*STRIDE_P4 + (cj & 1)*(c_packX4 >> 1);
235 else if (clusterSize == c_nbnxnCpuIClusterSize)
237 /* Coordinates are stored packed in groups of 4 */
242 /* Coordinates are stored packed in groups of 8 */
248 /* Initializes a single nbnxn_pairlist_t data structure */
249 static void nbnxn_init_pairlist_fep(t_nblist *nl)
251 nl->type = GMX_NBLIST_INTERACTION_FREE_ENERGY;
252 nl->igeometry = GMX_NBLIST_GEOMETRY_PARTICLE_PARTICLE;
253 /* The interaction functions are set in the free energy kernel fuction */
266 nl->jindex = nullptr;
268 nl->excl_fep = nullptr;
272 static void free_nblist(t_nblist *nl)
282 nbnxn_search_work_t::nbnxn_search_work_t() :
285 buffer_flags({0, nullptr, 0}),
287 nbl_fep(new t_nblist),
290 nbnxn_init_pairlist_fep(nbl_fep.get());
295 nbnxn_search_work_t::~nbnxn_search_work_t()
297 sfree(buffer_flags.flag);
299 free_nblist(nbl_fep.get());
302 nbnxn_search::nbnxn_search(int ePBC,
303 const ivec *n_dd_cells,
304 const gmx_domdec_zones_t *zones,
305 const PairlistType pairlistType,
316 // The correct value will be set during the gridding
320 DomDec = n_dd_cells != nullptr;
323 for (int d = 0; d < DIM; d++)
325 if ((*n_dd_cells)[d] > 1)
328 /* Each grid matches a DD zone */
334 grid.resize(numGrids, pairlistType);
336 /* Initialize detailed nbsearch cycle counting */
337 print_cycles = (getenv("GMX_NBNXN_CYCLE") != nullptr);
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.
365 listRangeForBoundingBoxToGridCell(real rlist,
366 const Grid::Dimensions &gridDims)
368 return rlist + gridDims.maxAtomGroupRadius;
371 /* Returns the pair-list cutoff between a grid cells given an atom-to-atom pair-list cutoff
373 * Given a cutoff distance between atoms, this functions returns the cutoff
374 * distance2 between two grid cells.
375 * Since atoms can be geometrically outside of the cell they have been
376 * assigned to (when atom groups instead of individual atoms are assigned
377 * to cells), this distance returned can be larger than the input.
380 listRangeForGridCellToGridCell(real rlist,
381 const Grid::Dimensions &iGridDims,
382 const Grid::Dimensions &jGridDims)
384 return rlist + iGridDims.maxAtomGroupRadius + jGridDims.maxAtomGroupRadius;
387 /* Determines the cell range along one dimension that
388 * the bounding box b0 - b1 sees.
391 static void get_cell_range(real b0, real b1,
392 const Grid::Dimensions &jGridDims,
393 real d2, real rlist, int *cf, int *cl)
395 real listRangeBBToCell2 = gmx::square(listRangeForBoundingBoxToGridCell(rlist, jGridDims));
396 real distanceInCells = (b0 - jGridDims.lowerCorner[dim])*jGridDims.invCellSize[dim];
397 *cf = std::max(static_cast<int>(distanceInCells), 0);
400 d2 + gmx::square((b0 - jGridDims.lowerCorner[dim]) - (*cf - 1 + 1)*jGridDims.cellSize[dim]) < listRangeBBToCell2)
405 *cl = std::min(static_cast<int>((b1 - jGridDims.lowerCorner[dim])*jGridDims.invCellSize[dim]), jGridDims.numCells[dim] - 1);
406 while (*cl < jGridDims.numCells[dim] - 1 &&
407 d2 + gmx::square((*cl + 1)*jGridDims.cellSize[dim] - (b1 - jGridDims.lowerCorner[dim])) < listRangeBBToCell2)
413 /* Reference code calculating the distance^2 between two bounding boxes */
415 static float box_dist2(float bx0, float bx1, float by0,
416 float by1, float bz0, float bz1,
417 const BoundingBox *bb)
420 float dl, dh, dm, dm0;
424 dl = bx0 - bb->upper.x;
425 dh = bb->lower.x - bx1;
426 dm = std::max(dl, dh);
427 dm0 = std::max(dm, 0.0f);
430 dl = by0 - bb->upper.y;
431 dh = bb->lower.y - by1;
432 dm = std::max(dl, dh);
433 dm0 = std::max(dm, 0.0f);
436 dl = bz0 - bb->upper.z;
437 dh = bb->lower.z - bz1;
438 dm = std::max(dl, dh);
439 dm0 = std::max(dm, 0.0f);
446 #if !NBNXN_SEARCH_BB_SIMD4
448 /*! \brief Plain C code calculating the distance^2 between two bounding boxes in xyz0 format
450 * \param[in] bb_i First bounding box
451 * \param[in] bb_j Second bounding box
453 static float clusterBoundingBoxDistance2(const BoundingBox &bb_i,
454 const BoundingBox &bb_j)
456 float dl = bb_i.lower.x - bb_j.upper.x;
457 float dh = bb_j.lower.x - bb_i.upper.x;
458 float dm = std::max(dl, dh);
459 float dm0 = std::max(dm, 0.0f);
462 dl = bb_i.lower.y - bb_j.upper.y;
463 dh = bb_j.lower.y - bb_i.upper.y;
464 dm = std::max(dl, dh);
465 dm0 = std::max(dm, 0.0f);
468 dl = bb_i.lower.z - bb_j.upper.z;
469 dh = bb_j.lower.z - bb_i.upper.z;
470 dm = std::max(dl, dh);
471 dm0 = std::max(dm, 0.0f);
477 #else /* NBNXN_SEARCH_BB_SIMD4 */
479 /*! \brief 4-wide SIMD code calculating the distance^2 between two bounding boxes in xyz0 format
481 * \param[in] bb_i First bounding box, should be aligned for 4-wide SIMD
482 * \param[in] bb_j Second bounding box, should be aligned for 4-wide SIMD
484 static float clusterBoundingBoxDistance2(const BoundingBox &bb_i,
485 const BoundingBox &bb_j)
487 // TODO: During SIMDv2 transition only some archs use namespace (remove when done)
490 const Simd4Float bb_i_S0 = load4(bb_i.lower.ptr());
491 const Simd4Float bb_i_S1 = load4(bb_i.upper.ptr());
492 const Simd4Float bb_j_S0 = load4(bb_j.lower.ptr());
493 const Simd4Float bb_j_S1 = load4(bb_j.upper.ptr());
495 const Simd4Float dl_S = bb_i_S0 - bb_j_S1;
496 const Simd4Float dh_S = bb_j_S0 - bb_i_S1;
498 const Simd4Float dm_S = max(dl_S, dh_S);
499 const Simd4Float dm0_S = max(dm_S, simd4SetZeroF());
501 return dotProduct(dm0_S, dm0_S);
504 /* Calculate bb bounding distances of bb_i[si,...,si+3] and store them in d2 */
505 template <int boundingBoxStart>
506 static inline void gmx_simdcall
507 clusterBoundingBoxDistance2_xxxx_simd4_inner(const float *bb_i,
509 const Simd4Float xj_l,
510 const Simd4Float yj_l,
511 const Simd4Float zj_l,
512 const Simd4Float xj_h,
513 const Simd4Float yj_h,
514 const Simd4Float zj_h)
516 const int shi = boundingBoxStart*Nbnxm::c_numBoundingBoxBounds1D*DIM;
518 const Simd4Float zero = setZero();
520 const Simd4Float xi_l = load4(bb_i + shi + 0*STRIDE_PBB);
521 const Simd4Float yi_l = load4(bb_i + shi + 1*STRIDE_PBB);
522 const Simd4Float zi_l = load4(bb_i + shi + 2*STRIDE_PBB);
523 const Simd4Float xi_h = load4(bb_i + shi + 3*STRIDE_PBB);
524 const Simd4Float yi_h = load4(bb_i + shi + 4*STRIDE_PBB);
525 const Simd4Float zi_h = load4(bb_i + shi + 5*STRIDE_PBB);
527 const Simd4Float dx_0 = xi_l - xj_h;
528 const Simd4Float dy_0 = yi_l - yj_h;
529 const Simd4Float dz_0 = zi_l - zj_h;
531 const Simd4Float dx_1 = xj_l - xi_h;
532 const Simd4Float dy_1 = yj_l - yi_h;
533 const Simd4Float dz_1 = zj_l - zi_h;
535 const Simd4Float mx = max(dx_0, dx_1);
536 const Simd4Float my = max(dy_0, dy_1);
537 const Simd4Float mz = max(dz_0, dz_1);
539 const Simd4Float m0x = max(mx, zero);
540 const Simd4Float m0y = max(my, zero);
541 const Simd4Float m0z = max(mz, zero);
543 const Simd4Float d2x = m0x * m0x;
544 const Simd4Float d2y = m0y * m0y;
545 const Simd4Float d2z = m0z * m0z;
547 const Simd4Float d2s = d2x + d2y;
548 const Simd4Float d2t = d2s + d2z;
550 store4(d2 + boundingBoxStart, d2t);
553 /* 4-wide SIMD code for nsi bb distances for bb format xxxxyyyyzzzz */
555 clusterBoundingBoxDistance2_xxxx_simd4(const float *bb_j,
560 // TODO: During SIMDv2 transition only some archs use namespace (remove when done)
563 const Simd4Float xj_l = Simd4Float(bb_j[0*STRIDE_PBB]);
564 const Simd4Float yj_l = Simd4Float(bb_j[1*STRIDE_PBB]);
565 const Simd4Float zj_l = Simd4Float(bb_j[2*STRIDE_PBB]);
566 const Simd4Float xj_h = Simd4Float(bb_j[3*STRIDE_PBB]);
567 const Simd4Float yj_h = Simd4Float(bb_j[4*STRIDE_PBB]);
568 const Simd4Float zj_h = Simd4Float(bb_j[5*STRIDE_PBB]);
570 /* Here we "loop" over si (0,STRIDE_PBB) from 0 to nsi with step STRIDE_PBB.
571 * But as we know the number of iterations is 1 or 2, we unroll manually.
573 clusterBoundingBoxDistance2_xxxx_simd4_inner<0>(bb_i, d2,
576 if (STRIDE_PBB < nsi)
578 clusterBoundingBoxDistance2_xxxx_simd4_inner<STRIDE_PBB>(bb_i, d2,
584 #endif /* NBNXN_SEARCH_BB_SIMD4 */
587 /* Returns if any atom pair from two clusters is within distance sqrt(rlist2) */
588 static inline gmx_bool
589 clusterpair_in_range(const NbnxnPairlistGpuWork &work,
591 int csj, int stride, const real *x_j,
594 #if !GMX_SIMD4_HAVE_REAL
597 * All coordinates are stored as xyzxyz...
600 const real *x_i = work.iSuperClusterData.x.data();
602 for (int i = 0; i < c_nbnxnGpuClusterSize; i++)
604 int i0 = (si*c_nbnxnGpuClusterSize + i)*DIM;
605 for (int j = 0; j < c_nbnxnGpuClusterSize; j++)
607 int j0 = (csj*c_nbnxnGpuClusterSize + j)*stride;
609 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]);
620 #else /* !GMX_SIMD4_HAVE_REAL */
622 /* 4-wide SIMD version.
623 * The coordinates x_i are stored as xxxxyyyy..., x_j is stored xyzxyz...
624 * Using 8-wide AVX(2) is not faster on Intel Sandy Bridge and Haswell.
626 static_assert(c_nbnxnGpuClusterSize == 8 || c_nbnxnGpuClusterSize == 4,
627 "A cluster is hard-coded to 4/8 atoms.");
629 Simd4Real rc2_S = Simd4Real(rlist2);
631 const real *x_i = work.iSuperClusterData.xSimd.data();
633 int dim_stride = c_nbnxnGpuClusterSize*DIM;
634 Simd4Real ix_S0 = load4(x_i + si*dim_stride + 0*GMX_SIMD4_WIDTH);
635 Simd4Real iy_S0 = load4(x_i + si*dim_stride + 1*GMX_SIMD4_WIDTH);
636 Simd4Real iz_S0 = load4(x_i + si*dim_stride + 2*GMX_SIMD4_WIDTH);
638 Simd4Real ix_S1, iy_S1, iz_S1;
639 if (c_nbnxnGpuClusterSize == 8)
641 ix_S1 = load4(x_i + si*dim_stride + 3*GMX_SIMD4_WIDTH);
642 iy_S1 = load4(x_i + si*dim_stride + 4*GMX_SIMD4_WIDTH);
643 iz_S1 = load4(x_i + si*dim_stride + 5*GMX_SIMD4_WIDTH);
645 /* We loop from the outer to the inner particles to maximize
646 * the chance that we find a pair in range quickly and return.
648 int j0 = csj*c_nbnxnGpuClusterSize;
649 int j1 = j0 + c_nbnxnGpuClusterSize - 1;
652 Simd4Real jx0_S, jy0_S, jz0_S;
653 Simd4Real jx1_S, jy1_S, jz1_S;
655 Simd4Real dx_S0, dy_S0, dz_S0;
656 Simd4Real dx_S1, dy_S1, dz_S1;
657 Simd4Real dx_S2, dy_S2, dz_S2;
658 Simd4Real dx_S3, dy_S3, dz_S3;
669 Simd4Bool wco_any_S01, wco_any_S23, wco_any_S;
671 jx0_S = Simd4Real(x_j[j0*stride+0]);
672 jy0_S = Simd4Real(x_j[j0*stride+1]);
673 jz0_S = Simd4Real(x_j[j0*stride+2]);
675 jx1_S = Simd4Real(x_j[j1*stride+0]);
676 jy1_S = Simd4Real(x_j[j1*stride+1]);
677 jz1_S = Simd4Real(x_j[j1*stride+2]);
679 /* Calculate distance */
680 dx_S0 = ix_S0 - jx0_S;
681 dy_S0 = iy_S0 - jy0_S;
682 dz_S0 = iz_S0 - jz0_S;
683 dx_S2 = ix_S0 - jx1_S;
684 dy_S2 = iy_S0 - jy1_S;
685 dz_S2 = iz_S0 - jz1_S;
686 if (c_nbnxnGpuClusterSize == 8)
688 dx_S1 = ix_S1 - jx0_S;
689 dy_S1 = iy_S1 - jy0_S;
690 dz_S1 = iz_S1 - jz0_S;
691 dx_S3 = ix_S1 - jx1_S;
692 dy_S3 = iy_S1 - jy1_S;
693 dz_S3 = iz_S1 - jz1_S;
696 /* rsq = dx*dx+dy*dy+dz*dz */
697 rsq_S0 = norm2(dx_S0, dy_S0, dz_S0);
698 rsq_S2 = norm2(dx_S2, dy_S2, dz_S2);
699 if (c_nbnxnGpuClusterSize == 8)
701 rsq_S1 = norm2(dx_S1, dy_S1, dz_S1);
702 rsq_S3 = norm2(dx_S3, dy_S3, dz_S3);
705 wco_S0 = (rsq_S0 < rc2_S);
706 wco_S2 = (rsq_S2 < rc2_S);
707 if (c_nbnxnGpuClusterSize == 8)
709 wco_S1 = (rsq_S1 < rc2_S);
710 wco_S3 = (rsq_S3 < rc2_S);
712 if (c_nbnxnGpuClusterSize == 8)
714 wco_any_S01 = wco_S0 || wco_S1;
715 wco_any_S23 = wco_S2 || wco_S3;
716 wco_any_S = wco_any_S01 || wco_any_S23;
720 wco_any_S = wco_S0 || wco_S2;
723 if (anyTrue(wco_any_S))
734 #endif /* !GMX_SIMD4_HAVE_REAL */
737 /* Returns the j-cluster index for index cjIndex in a cj list */
738 static inline int nblCj(gmx::ArrayRef<const nbnxn_cj_t> cjList,
741 return cjList[cjIndex].cj;
744 /* Returns the j-cluster index for index cjIndex in a cj4 list */
745 static inline int nblCj(gmx::ArrayRef<const nbnxn_cj4_t> cj4List,
748 return cj4List[cjIndex/c_nbnxnGpuJgroupSize].cj[cjIndex & (c_nbnxnGpuJgroupSize - 1)];
751 /* Returns the i-interaction mask of the j sub-cell for index cj_ind */
752 static unsigned int nbl_imask0(const NbnxnPairlistGpu *nbl, int cj_ind)
754 return nbl->cj4[cj_ind/c_nbnxnGpuJgroupSize].imei[0].imask;
757 /* Initializes a single NbnxnPairlistCpu data structure */
758 static void nbnxn_init_pairlist(NbnxnPairlistCpu *nbl)
760 nbl->na_ci = c_nbnxnCpuIClusterSize;
763 nbl->ciOuter.clear();
766 nbl->cjOuter.clear();
769 nbl->work = new NbnxnPairlistCpuWork();
772 NbnxnPairlistGpu::NbnxnPairlistGpu(gmx::PinningPolicy pinningPolicy) :
773 na_ci(c_nbnxnGpuClusterSize),
774 na_cj(c_nbnxnGpuClusterSize),
775 na_sc(c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize),
777 sci({}, {pinningPolicy}),
778 cj4({}, {pinningPolicy}),
779 excl({}, {pinningPolicy}),
782 static_assert(c_nbnxnGpuNumClusterPerSupercluster == c_gpuNumClusterPerCell,
783 "The search code assumes that the a super-cluster matches a search grid cell");
785 static_assert(sizeof(cj4[0].imei[0].imask)*8 >= c_nbnxnGpuJgroupSize*c_gpuNumClusterPerCell,
786 "The i super-cluster cluster interaction mask does not contain a sufficient number of bits");
788 static_assert(sizeof(excl[0])*8 >= c_nbnxnGpuJgroupSize*c_gpuNumClusterPerCell, "The GPU exclusion mask does not contain a sufficient number of bits");
790 // We always want a first entry without any exclusions
793 work = new NbnxnPairlistGpuWork();
796 void nbnxn_init_pairlist_set(nbnxn_pairlist_set_t *nbl_list)
799 (nbl_list->params.pairlistType == PairlistType::Simple4x2 ||
800 nbl_list->params.pairlistType == PairlistType::Simple4x4 ||
801 nbl_list->params.pairlistType == PairlistType::Simple4x8);
802 // Currently GPU lists are always combined
803 nbl_list->bCombined = !nbl_list->bSimple;
805 nbl_list->nnbl = gmx_omp_nthreads_get(emntNonbonded);
807 if (!nbl_list->bCombined &&
808 nbl_list->nnbl > NBNXN_BUFFERFLAG_MAX_THREADS)
810 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.",
811 nbl_list->nnbl, NBNXN_BUFFERFLAG_MAX_THREADS, NBNXN_BUFFERFLAG_MAX_THREADS);
814 if (nbl_list->bSimple)
816 snew(nbl_list->nbl, nbl_list->nnbl);
817 if (nbl_list->nnbl > 1)
819 snew(nbl_list->nbl_work, nbl_list->nnbl);
824 snew(nbl_list->nblGpu, nbl_list->nnbl);
826 nbl_list->nbl_fep.resize(nbl_list->nnbl);
827 /* Execute in order to avoid memory interleaving between threads */
828 #pragma omp parallel for num_threads(nbl_list->nnbl) schedule(static)
829 for (int i = 0; i < nbl_list->nnbl; i++)
833 /* Allocate the nblist data structure locally on each thread
834 * to optimize memory access for NUMA architectures.
836 if (nbl_list->bSimple)
838 nbl_list->nbl[i] = new NbnxnPairlistCpu();
840 nbnxn_init_pairlist(nbl_list->nbl[i]);
841 if (nbl_list->nnbl > 1)
843 nbl_list->nbl_work[i] = new NbnxnPairlistCpu();
844 nbnxn_init_pairlist(nbl_list->nbl_work[i]);
849 /* Only list 0 is used on the GPU, use normal allocation for i>0 */
850 auto pinningPolicy = (i == 0 ? gmx::PinningPolicy::PinnedIfSupported : gmx::PinningPolicy::CannotBePinned);
852 nbl_list->nblGpu[i] = new NbnxnPairlistGpu(pinningPolicy);
855 snew(nbl_list->nbl_fep[i], 1);
856 nbnxn_init_pairlist_fep(nbl_list->nbl_fep[i]);
858 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
862 /* Print statistics of a pair list, used for debug output */
863 static void print_nblist_statistics(FILE *fp, const NbnxnPairlistCpu *nbl,
864 const nbnxn_search *nbs, real rl)
866 const Grid &grid = nbs->grid[0];
867 const Grid::Dimensions &dims = grid.dimensions();
869 fprintf(fp, "nbl nci %zu ncj %d\n",
870 nbl->ci.size(), nbl->ncjInUse);
871 const int numAtomsJCluster = grid.geometry().numAtomsJCluster;
872 const double numAtomsPerCell = nbl->ncjInUse/static_cast<double>(grid.numCells())*numAtomsJCluster;
873 fprintf(fp, "nbl na_cj %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
874 nbl->na_cj, rl, nbl->ncjInUse, nbl->ncjInUse/static_cast<double>(grid.numCells()),
876 numAtomsPerCell/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid.numCells()*numAtomsJCluster/(dims.gridSize[XX]*dims.gridSize[YY]*dims.gridSize[ZZ])));
878 fprintf(fp, "nbl average j cell list length %.1f\n",
879 0.25*nbl->ncjInUse/std::max(static_cast<double>(nbl->ci.size()), 1.0));
881 int cs[SHIFTS] = { 0 };
883 for (const nbnxn_ci_t &ciEntry : nbl->ci)
885 cs[ciEntry.shift & NBNXN_CI_SHIFT] +=
886 ciEntry.cj_ind_end - ciEntry.cj_ind_start;
888 int j = ciEntry.cj_ind_start;
889 while (j < ciEntry.cj_ind_end &&
890 nbl->cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
896 fprintf(fp, "nbl cell pairs, total: %zu excl: %d %.1f%%\n",
897 nbl->cj.size(), npexcl, 100*npexcl/std::max(static_cast<double>(nbl->cj.size()), 1.0));
898 for (int s = 0; s < SHIFTS; s++)
902 fprintf(fp, "nbl shift %2d ncj %3d\n", s, cs[s]);
907 /* Print statistics of a pair lists, used for debug output */
908 static void print_nblist_statistics(FILE *fp, const NbnxnPairlistGpu *nbl,
909 const nbnxn_search *nbs, real rl)
911 const Grid &grid = nbs->grid[0];
912 const Grid::Dimensions &dims = grid.dimensions();
914 fprintf(fp, "nbl nsci %zu ncj4 %zu nsi %d excl4 %zu\n",
915 nbl->sci.size(), nbl->cj4.size(), nbl->nci_tot, nbl->excl.size());
916 const int numAtomsCluster = grid.geometry().numAtomsICluster;
917 const double numAtomsPerCell = nbl->nci_tot/static_cast<double>(grid.numClusters())*numAtomsCluster;
918 fprintf(fp, "nbl na_c %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
919 nbl->na_ci, rl, nbl->nci_tot, nbl->nci_tot/static_cast<double>(grid.numClusters()),
921 numAtomsPerCell/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid.numClusters()*numAtomsCluster/(dims.gridSize[XX]*dims.gridSize[YY]*dims.gridSize[ZZ])));
926 int c[c_gpuNumClusterPerCell + 1] = { 0 };
927 for (const nbnxn_sci_t &sci : nbl->sci)
930 for (int j4 = sci.cj4_ind_start; j4 < sci.cj4_ind_end; j4++)
932 for (int j = 0; j < c_nbnxnGpuJgroupSize; j++)
935 for (int si = 0; si < c_gpuNumClusterPerCell; si++)
937 if (nbl->cj4[j4].imei[0].imask & (1U << (j*c_gpuNumClusterPerCell + si)))
948 nsp_max = std::max(nsp_max, nsp);
950 if (!nbl->sci.empty())
952 sum_nsp /= nbl->sci.size();
953 sum_nsp2 /= nbl->sci.size();
955 fprintf(fp, "nbl #cluster-pairs: av %.1f stddev %.1f max %d\n",
956 sum_nsp, std::sqrt(sum_nsp2 - sum_nsp*sum_nsp), nsp_max);
958 if (!nbl->cj4.empty())
960 for (int b = 0; b <= c_gpuNumClusterPerCell; b++)
962 fprintf(fp, "nbl j-list #i-subcell %d %7d %4.1f\n",
963 b, c[b], 100.0*c[b]/size_t {nbl->cj4.size()*c_nbnxnGpuJgroupSize});
968 /* Returns a pointer to the exclusion mask for j-cluster-group \p cj4 and warp \p warp
969 * Generates a new exclusion entry when the j-cluster-group uses
970 * the default all-interaction mask at call time, so the returned mask
971 * can be modified when needed.
973 static nbnxn_excl_t *get_exclusion_mask(NbnxnPairlistGpu *nbl,
977 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
979 /* No exclusions set, make a new list entry */
980 const size_t oldSize = nbl->excl.size();
981 GMX_ASSERT(oldSize >= 1, "We should always have entry [0]");
982 /* Add entry with default values: no exclusions */
983 nbl->excl.resize(oldSize + 1);
984 nbl->cj4[cj4].imei[warp].excl_ind = oldSize;
987 return &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
990 static void set_self_and_newton_excls_supersub(NbnxnPairlistGpu *nbl,
991 int cj4_ind, int sj_offset,
992 int i_cluster_in_cell)
994 nbnxn_excl_t *excl[c_nbnxnGpuClusterpairSplit];
996 /* Here we only set the set self and double pair exclusions */
998 /* Reserve extra elements, so the resize() in get_exclusion_mask()
999 * will not invalidate excl entries in the loop below
1001 nbl->excl.reserve(nbl->excl.size() + c_nbnxnGpuClusterpairSplit);
1002 for (int w = 0; w < c_nbnxnGpuClusterpairSplit; w++)
1004 excl[w] = get_exclusion_mask(nbl, cj4_ind, w);
1007 /* Only minor < major bits set */
1008 for (int ej = 0; ej < nbl->na_ci; ej++)
1011 for (int ei = ej; ei < nbl->na_ci; ei++)
1013 excl[w]->pair[(ej & (c_nbnxnGpuJgroupSize-1))*nbl->na_ci + ei] &=
1014 ~(1U << (sj_offset*c_gpuNumClusterPerCell + i_cluster_in_cell));
1019 /* Returns a diagonal or off-diagonal interaction mask for plain C lists */
1020 static unsigned int get_imask(gmx_bool rdiag, int ci, int cj)
1022 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
1025 /* Returns a diagonal or off-diagonal interaction mask for cj-size=2 */
1026 gmx_unused static unsigned int get_imask_simd_j2(gmx_bool rdiag, int ci, int cj)
1028 return (rdiag && ci*2 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_0 :
1029 (rdiag && ci*2+1 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_1 :
1030 NBNXN_INTERACTION_MASK_ALL));
1033 /* Returns a diagonal or off-diagonal interaction mask for cj-size=4 */
1034 gmx_unused static unsigned int get_imask_simd_j4(gmx_bool rdiag, int ci, int cj)
1036 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
1039 /* Returns a diagonal or off-diagonal interaction mask for cj-size=8 */
1040 gmx_unused static unsigned int get_imask_simd_j8(gmx_bool rdiag, int ci, int cj)
1042 return (rdiag && ci == cj*2 ? NBNXN_INTERACTION_MASK_DIAG_J8_0 :
1043 (rdiag && ci == cj*2+1 ? NBNXN_INTERACTION_MASK_DIAG_J8_1 :
1044 NBNXN_INTERACTION_MASK_ALL));
1048 #if GMX_SIMD_REAL_WIDTH == 2
1049 #define get_imask_simd_4xn get_imask_simd_j2
1051 #if GMX_SIMD_REAL_WIDTH == 4
1052 #define get_imask_simd_4xn get_imask_simd_j4
1054 #if GMX_SIMD_REAL_WIDTH == 8
1055 #define get_imask_simd_4xn get_imask_simd_j8
1056 #define get_imask_simd_2xnn get_imask_simd_j4
1058 #if GMX_SIMD_REAL_WIDTH == 16
1059 #define get_imask_simd_2xnn get_imask_simd_j8
1063 /* Plain C code for checking and adding cluster-pairs to the list.
1065 * \param[in] gridj The j-grid
1066 * \param[in,out] nbl The pair-list to store the cluster pairs in
1067 * \param[in] icluster The index of the i-cluster
1068 * \param[in] jclusterFirst The first cluster in the j-range
1069 * \param[in] jclusterLast The last cluster in the j-range
1070 * \param[in] excludeSubDiagonal Exclude atom pairs with i-index > j-index
1071 * \param[in] x_j Coordinates for the j-atom, in xyz format
1072 * \param[in] rlist2 The squared list cut-off
1073 * \param[in] rbb2 The squared cut-off for putting cluster-pairs in the list based on bounding box distance only
1074 * \param[in,out] numDistanceChecks The number of distance checks performed
1077 makeClusterListSimple(const Grid &jGrid,
1078 NbnxnPairlistCpu * nbl,
1082 bool excludeSubDiagonal,
1083 const real * gmx_restrict x_j,
1086 int * gmx_restrict numDistanceChecks)
1088 const BoundingBox * gmx_restrict bb_ci = nbl->work->iClusterData.bb.data();
1089 const real * gmx_restrict x_ci = nbl->work->iClusterData.x.data();
1094 while (!InRange && jclusterFirst <= jclusterLast)
1096 real d2 = clusterBoundingBoxDistance2(bb_ci[0], jGrid.jBoundingBoxes()[jclusterFirst]);
1097 *numDistanceChecks += 2;
1099 /* Check if the distance is within the distance where
1100 * we use only the bounding box distance rbb,
1101 * or within the cut-off and there is at least one atom pair
1102 * within the cut-off.
1108 else if (d2 < rlist2)
1110 int cjf_gl = jGrid.cellOffset() + jclusterFirst;
1111 for (int i = 0; i < c_nbnxnCpuIClusterSize && !InRange; i++)
1113 for (int j = 0; j < c_nbnxnCpuIClusterSize; j++)
1115 InRange = InRange ||
1116 (gmx::square(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjf_gl*c_nbnxnCpuIClusterSize+j)*STRIDE_XYZ+XX]) +
1117 gmx::square(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjf_gl*c_nbnxnCpuIClusterSize+j)*STRIDE_XYZ+YY]) +
1118 gmx::square(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjf_gl*c_nbnxnCpuIClusterSize+j)*STRIDE_XYZ+ZZ]) < rlist2);
1121 *numDistanceChecks += c_nbnxnCpuIClusterSize*c_nbnxnCpuIClusterSize;
1134 while (!InRange && jclusterLast > jclusterFirst)
1136 real d2 = clusterBoundingBoxDistance2(bb_ci[0], jGrid.jBoundingBoxes()[jclusterLast]);
1137 *numDistanceChecks += 2;
1139 /* Check if the distance is within the distance where
1140 * we use only the bounding box distance rbb,
1141 * or within the cut-off and there is at least one atom pair
1142 * within the cut-off.
1148 else if (d2 < rlist2)
1150 int cjl_gl = jGrid.cellOffset() + jclusterLast;
1151 for (int i = 0; i < c_nbnxnCpuIClusterSize && !InRange; i++)
1153 for (int j = 0; j < c_nbnxnCpuIClusterSize; j++)
1155 InRange = InRange ||
1156 (gmx::square(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjl_gl*c_nbnxnCpuIClusterSize+j)*STRIDE_XYZ+XX]) +
1157 gmx::square(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjl_gl*c_nbnxnCpuIClusterSize+j)*STRIDE_XYZ+YY]) +
1158 gmx::square(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjl_gl*c_nbnxnCpuIClusterSize+j)*STRIDE_XYZ+ZZ]) < rlist2);
1161 *numDistanceChecks += c_nbnxnCpuIClusterSize*c_nbnxnCpuIClusterSize;
1169 if (jclusterFirst <= jclusterLast)
1171 for (int jcluster = jclusterFirst; jcluster <= jclusterLast; jcluster++)
1173 /* Store cj and the interaction mask */
1175 cjEntry.cj = jGrid.cellOffset() + jcluster;
1176 cjEntry.excl = get_imask(excludeSubDiagonal, icluster, jcluster);
1177 nbl->cj.push_back(cjEntry);
1179 /* Increase the closing index in the i list */
1180 nbl->ci.back().cj_ind_end = nbl->cj.size();
1184 #ifdef GMX_NBNXN_SIMD_4XN
1185 #include "gromacs/nbnxm/pairlist_simd_4xm.h"
1187 #ifdef GMX_NBNXN_SIMD_2XNN
1188 #include "gromacs/nbnxm/pairlist_simd_2xmm.h"
1191 /* Plain C or SIMD4 code for making a pair list of super-cell sci vs scj.
1192 * Checks bounding box distances and possibly atom pair distances.
1194 static void make_cluster_list_supersub(const Grid &iGrid,
1196 NbnxnPairlistGpu *nbl,
1199 const bool excludeSubDiagonal,
1204 int *numDistanceChecks)
1206 NbnxnPairlistGpuWork &work = *nbl->work;
1209 const float *pbb_ci = work.iSuperClusterData.bbPacked.data();
1211 const BoundingBox *bb_ci = work.iSuperClusterData.bb.data();
1214 assert(c_nbnxnGpuClusterSize == iGrid.geometry().numAtomsICluster);
1215 assert(c_nbnxnGpuClusterSize == jGrid.geometry().numAtomsICluster);
1217 /* We generate the pairlist mainly based on bounding-box distances
1218 * and do atom pair distance based pruning on the GPU.
1219 * Only if a j-group contains a single cluster-pair, we try to prune
1220 * that pair based on atom distances on the CPU to avoid empty j-groups.
1222 #define PRUNE_LIST_CPU_ONE 1
1223 #define PRUNE_LIST_CPU_ALL 0
1225 #if PRUNE_LIST_CPU_ONE
1229 float *d2l = work.distanceBuffer.data();
1231 for (int subc = 0; subc < jGrid.numClustersPerCell()[scj]; subc++)
1233 const int cj4_ind = work.cj_ind/c_nbnxnGpuJgroupSize;
1234 const int cj_offset = work.cj_ind - cj4_ind*c_nbnxnGpuJgroupSize;
1235 const int cj = scj*c_gpuNumClusterPerCell + subc;
1237 const int cj_gl = jGrid.cellOffset()*c_gpuNumClusterPerCell + cj;
1240 if (excludeSubDiagonal && sci == scj)
1246 ci1 = iGrid.numClustersPerCell()[sci];
1250 /* Determine all ci1 bb distances in one call with SIMD4 */
1251 clusterBoundingBoxDistance2_xxxx_simd4(jGrid.packedBoundingBoxes().data() + (cj >> STRIDE_PBB_2LOG)*NNBSBB_XXXX + (cj & (STRIDE_PBB-1)),
1253 *numDistanceChecks += c_nbnxnGpuClusterSize*2;
1257 unsigned int imask = 0;
1258 /* We use a fixed upper-bound instead of ci1 to help optimization */
1259 for (int ci = 0; ci < c_gpuNumClusterPerCell; ci++)
1267 /* Determine the bb distance between ci and cj */
1268 d2l[ci] = clusterBoundingBoxDistance2(bb_ci[ci], jGrid.jBoundingBoxes()[cj]);
1269 *numDistanceChecks += 2;
1273 #if PRUNE_LIST_CPU_ALL
1274 /* Check if the distance is within the distance where
1275 * we use only the bounding box distance rbb,
1276 * or within the cut-off and there is at least one atom pair
1277 * within the cut-off. This check is very costly.
1279 *numDistanceChecks += c_nbnxnGpuClusterSize*c_nbnxnGpuClusterSize;
1282 clusterpair_in_range(work, ci, cj_gl, stride, x, rlist2)))
1284 /* Check if the distance between the two bounding boxes
1285 * in within the pair-list cut-off.
1290 /* Flag this i-subcell to be taken into account */
1291 imask |= (1U << (cj_offset*c_gpuNumClusterPerCell + ci));
1293 #if PRUNE_LIST_CPU_ONE
1301 #if PRUNE_LIST_CPU_ONE
1302 /* If we only found 1 pair, check if any atoms are actually
1303 * within the cut-off, so we could get rid of it.
1305 if (npair == 1 && d2l[ci_last] >= rbb2 &&
1306 !clusterpair_in_range(work, ci_last, cj_gl, stride, x, rlist2))
1308 imask &= ~(1U << (cj_offset*c_gpuNumClusterPerCell + ci_last));
1315 /* We have at least one cluster pair: add a j-entry */
1316 if (static_cast<size_t>(cj4_ind) == nbl->cj4.size())
1318 nbl->cj4.resize(nbl->cj4.size() + 1);
1320 nbnxn_cj4_t *cj4 = &nbl->cj4[cj4_ind];
1322 cj4->cj[cj_offset] = cj_gl;
1324 /* Set the exclusions for the ci==sj entry.
1325 * Here we don't bother to check if this entry is actually flagged,
1326 * as it will nearly always be in the list.
1328 if (excludeSubDiagonal && sci == scj)
1330 set_self_and_newton_excls_supersub(nbl, cj4_ind, cj_offset, subc);
1333 /* Copy the cluster interaction mask to the list */
1334 for (int w = 0; w < c_nbnxnGpuClusterpairSplit; w++)
1336 cj4->imei[w].imask |= imask;
1339 nbl->work->cj_ind++;
1341 /* Keep the count */
1342 nbl->nci_tot += npair;
1344 /* Increase the closing index in i super-cell list */
1345 nbl->sci.back().cj4_ind_end =
1346 (nbl->work->cj_ind + c_nbnxnGpuJgroupSize - 1)/c_nbnxnGpuJgroupSize;
1351 /* Returns how many contiguous j-clusters we have starting in the i-list */
1352 template <typename CjListType>
1353 static int numContiguousJClusters(const int cjIndexStart,
1354 const int cjIndexEnd,
1355 gmx::ArrayRef<const CjListType> cjList)
1357 const int firstJCluster = nblCj(cjList, cjIndexStart);
1359 int numContiguous = 0;
1361 while (cjIndexStart + numContiguous < cjIndexEnd &&
1362 nblCj(cjList, cjIndexStart + numContiguous) == firstJCluster + numContiguous)
1367 return numContiguous;
1371 * \brief Helper struct for efficient searching for excluded atoms in a j-list
1375 /*! \brief Constructs a j-list range from \p cjList with the given index range */
1376 template <typename CjListType>
1377 JListRanges(int cjIndexStart,
1379 gmx::ArrayRef<const CjListType> cjList);
1381 int cjIndexStart; //!< The start index in the j-list
1382 int cjIndexEnd; //!< The end index in the j-list
1383 int cjFirst; //!< The j-cluster with index cjIndexStart
1384 int cjLast; //!< The j-cluster with index cjIndexEnd-1
1385 int numDirect; //!< Up to cjIndexStart+numDirect the j-clusters are cjFirst + the index offset
1389 template <typename CjListType>
1390 JListRanges::JListRanges(int cjIndexStart,
1392 gmx::ArrayRef<const CjListType> cjList) :
1393 cjIndexStart(cjIndexStart),
1394 cjIndexEnd(cjIndexEnd)
1396 GMX_ASSERT(cjIndexEnd > cjIndexStart, "JListRanges should only be called with non-empty lists");
1398 cjFirst = nblCj(cjList, cjIndexStart);
1399 cjLast = nblCj(cjList, cjIndexEnd - 1);
1401 /* Determine how many contiguous j-cells we have starting
1402 * from the first i-cell. This number can be used to directly
1403 * calculate j-cell indices for excluded atoms.
1405 numDirect = numContiguousJClusters(cjIndexStart, cjIndexEnd, cjList);
1409 /* Return the index of \p jCluster in the given range or -1 when not present
1411 * Note: This code is executed very often and therefore performance is
1412 * important. It should be inlined and fully optimized.
1414 template <typename CjListType>
1416 findJClusterInJList(int jCluster,
1417 const JListRanges &ranges,
1418 gmx::ArrayRef<const CjListType> cjList)
1422 if (jCluster < ranges.cjFirst + ranges.numDirect)
1424 /* We can calculate the index directly using the offset */
1425 index = ranges.cjIndexStart + jCluster - ranges.cjFirst;
1429 /* Search for jCluster using bisection */
1431 int rangeStart = ranges.cjIndexStart + ranges.numDirect;
1432 int rangeEnd = ranges.cjIndexEnd;
1434 while (index == -1 && rangeStart < rangeEnd)
1436 rangeMiddle = (rangeStart + rangeEnd) >> 1;
1438 const int clusterMiddle = nblCj(cjList, rangeMiddle);
1440 if (jCluster == clusterMiddle)
1442 index = rangeMiddle;
1444 else if (jCluster < clusterMiddle)
1446 rangeEnd = rangeMiddle;
1450 rangeStart = rangeMiddle + 1;
1458 // TODO: Get rid of the two functions below by renaming sci to ci (or something better)
1460 /* Return the i-entry in the list we are currently operating on */
1461 static nbnxn_ci_t *getOpenIEntry(NbnxnPairlistCpu *nbl)
1463 return &nbl->ci.back();
1466 /* Return the i-entry in the list we are currently operating on */
1467 static nbnxn_sci_t *getOpenIEntry(NbnxnPairlistGpu *nbl)
1469 return &nbl->sci.back();
1472 /* Set all atom-pair exclusions for a simple type list i-entry
1474 * Set all atom-pair exclusions from the topology stored in exclusions
1475 * as masks in the pair-list for simple list entry iEntry.
1478 setExclusionsForIEntry(const nbnxn_search *nbs,
1479 NbnxnPairlistCpu *nbl,
1480 gmx_bool diagRemoved,
1482 const nbnxn_ci_t &iEntry,
1483 const t_blocka &exclusions)
1485 if (iEntry.cj_ind_end == iEntry.cj_ind_start)
1487 /* Empty list: no exclusions */
1491 const JListRanges ranges(iEntry.cj_ind_start, iEntry.cj_ind_end, gmx::makeConstArrayRef(nbl->cj));
1493 const int iCluster = iEntry.ci;
1495 gmx::ArrayRef<const int> cell = nbs->cell;
1497 /* Loop over the atoms in the i-cluster */
1498 for (int i = 0; i < nbl->na_ci; i++)
1500 const int iIndex = iCluster*nbl->na_ci + i;
1501 const int iAtom = nbs->a[iIndex];
1504 /* Loop over the topology-based exclusions for this i-atom */
1505 for (int exclIndex = exclusions.index[iAtom]; exclIndex < exclusions.index[iAtom + 1]; exclIndex++)
1507 const int jAtom = exclusions.a[exclIndex];
1511 /* The self exclusion are already set, save some time */
1515 /* Get the index of the j-atom in the nbnxn atom data */
1516 const int jIndex = cell[jAtom];
1518 /* Without shifts we only calculate interactions j>i
1519 * for one-way pair-lists.
1521 if (diagRemoved && jIndex <= iIndex)
1526 const int jCluster = (jIndex >> na_cj_2log);
1528 /* Could the cluster se be in our list? */
1529 if (jCluster >= ranges.cjFirst && jCluster <= ranges.cjLast)
1532 findJClusterInJList(jCluster, ranges,
1533 gmx::makeConstArrayRef(nbl->cj));
1537 /* We found an exclusion, clear the corresponding
1540 const int innerJ = jIndex - (jCluster << na_cj_2log);
1542 nbl->cj[index].excl &= ~(1U << ((i << na_cj_2log) + innerJ));
1550 /* Add a new i-entry to the FEP list and copy the i-properties */
1551 static inline void fep_list_new_nri_copy(t_nblist *nlist)
1553 /* Add a new i-entry */
1556 assert(nlist->nri < nlist->maxnri);
1558 /* Duplicate the last i-entry, except for jindex, which continues */
1559 nlist->iinr[nlist->nri] = nlist->iinr[nlist->nri-1];
1560 nlist->shift[nlist->nri] = nlist->shift[nlist->nri-1];
1561 nlist->gid[nlist->nri] = nlist->gid[nlist->nri-1];
1562 nlist->jindex[nlist->nri] = nlist->nrj;
1565 /* For load balancing of the free-energy lists over threads, we set
1566 * the maximum nrj size of an i-entry to 40. This leads to good
1567 * load balancing in the worst case scenario of a single perturbed
1568 * particle on 16 threads, while not introducing significant overhead.
1569 * Note that half of the perturbed pairs will anyhow end up in very small lists,
1570 * since non perturbed i-particles will see few perturbed j-particles).
1572 const int max_nrj_fep = 40;
1574 /* Exclude the perturbed pairs from the Verlet list. This is only done to avoid
1575 * singularities for overlapping particles (0/0), since the charges and
1576 * LJ parameters have been zeroed in the nbnxn data structure.
1577 * Simultaneously make a group pair list for the perturbed pairs.
1579 static void make_fep_list(const nbnxn_search *nbs,
1580 const nbnxn_atomdata_t *nbat,
1581 NbnxnPairlistCpu *nbl,
1582 gmx_bool bDiagRemoved,
1584 real gmx_unused shx,
1585 real gmx_unused shy,
1586 real gmx_unused shz,
1587 real gmx_unused rlist_fep2,
1592 int ci, cj_ind_start, cj_ind_end, cja, cjr;
1594 int gid_i = 0, gid_j, gid;
1595 int egp_shift, egp_mask;
1597 int ind_i, ind_j, ai, aj;
1599 gmx_bool bFEP_i, bFEP_i_all;
1601 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
1609 cj_ind_start = nbl_ci->cj_ind_start;
1610 cj_ind_end = nbl_ci->cj_ind_end;
1612 /* In worst case we have alternating energy groups
1613 * and create #atom-pair lists, which means we need the size
1614 * of a cluster pair (na_ci*na_cj) times the number of cj's.
1616 nri_max = nbl->na_ci*nbl->na_cj*(cj_ind_end - cj_ind_start);
1617 if (nlist->nri + nri_max > nlist->maxnri)
1619 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
1620 reallocate_nblist(nlist);
1623 const int numAtomsJCluster = jGrid.geometry().numAtomsJCluster;
1625 const nbnxn_atomdata_t::Params &nbatParams = nbat->params();
1627 const int ngid = nbatParams.nenergrp;
1629 /* TODO: Consider adding a check in grompp and changing this to an assert */
1630 const int numBitsInEnergyGroupIdsForAtomsInJCluster = sizeof(gid_cj)*8;
1631 if (ngid*numAtomsJCluster > numBitsInEnergyGroupIdsForAtomsInJCluster)
1633 gmx_fatal(FARGS, "The Verlet scheme with %dx%d kernels and free-energy only supports up to %zu energy groups",
1634 iGrid.geometry().numAtomsICluster, numAtomsJCluster,
1635 (sizeof(gid_cj)*8)/numAtomsJCluster);
1638 egp_shift = nbatParams.neg_2log;
1639 egp_mask = (1 << egp_shift) - 1;
1641 /* Loop over the atoms in the i sub-cell */
1643 for (int i = 0; i < nbl->na_ci; i++)
1645 ind_i = ci*nbl->na_ci + i;
1650 nlist->jindex[nri+1] = nlist->jindex[nri];
1651 nlist->iinr[nri] = ai;
1652 /* The actual energy group pair index is set later */
1653 nlist->gid[nri] = 0;
1654 nlist->shift[nri] = nbl_ci->shift & NBNXN_CI_SHIFT;
1656 bFEP_i = iGrid.atomIsPerturbed(ci - iGrid.cellOffset(), i);
1658 bFEP_i_all = bFEP_i_all && bFEP_i;
1660 if (nlist->nrj + (cj_ind_end - cj_ind_start)*nbl->na_cj > nlist->maxnrj)
1662 nlist->maxnrj = over_alloc_small(nlist->nrj + (cj_ind_end - cj_ind_start)*nbl->na_cj);
1663 srenew(nlist->jjnr, nlist->maxnrj);
1664 srenew(nlist->excl_fep, nlist->maxnrj);
1669 gid_i = (nbatParams.energrp[ci] >> (egp_shift*i)) & egp_mask;
1672 for (int cj_ind = cj_ind_start; cj_ind < cj_ind_end; cj_ind++)
1674 unsigned int fep_cj;
1676 cja = nbl->cj[cj_ind].cj;
1678 if (numAtomsJCluster == jGrid.geometry().numAtomsICluster)
1680 cjr = cja - jGrid.cellOffset();
1681 fep_cj = jGrid.fepBits(cjr);
1684 gid_cj = nbatParams.energrp[cja];
1687 else if (2*numAtomsJCluster == jGrid.geometry().numAtomsICluster)
1689 cjr = cja - jGrid.cellOffset()*2;
1690 /* Extract half of the ci fep/energrp mask */
1691 fep_cj = (jGrid.fepBits(cjr >> 1) >> ((cjr & 1)*numAtomsJCluster)) & ((1 << numAtomsJCluster) - 1);
1694 gid_cj = nbatParams.energrp[cja >> 1] >> ((cja & 1)*numAtomsJCluster*egp_shift) & ((1 << (numAtomsJCluster*egp_shift)) - 1);
1699 cjr = cja - (jGrid.cellOffset() >> 1);
1700 /* Combine two ci fep masks/energrp */
1701 fep_cj = jGrid.fepBits(cjr*2) + (jGrid.fepBits(cjr*2 + 1) << jGrid.geometry().numAtomsICluster);
1704 gid_cj = nbatParams.energrp[cja*2] + (nbatParams.energrp[cja*2+1] << (jGrid.geometry().numAtomsICluster*egp_shift));
1708 if (bFEP_i || fep_cj != 0)
1710 for (int j = 0; j < nbl->na_cj; j++)
1712 /* Is this interaction perturbed and not excluded? */
1713 ind_j = cja*nbl->na_cj + j;
1716 (bFEP_i || (fep_cj & (1 << j))) &&
1717 (!bDiagRemoved || ind_j >= ind_i))
1721 gid_j = (gid_cj >> (j*egp_shift)) & egp_mask;
1722 gid = GID(gid_i, gid_j, ngid);
1724 if (nlist->nrj > nlist->jindex[nri] &&
1725 nlist->gid[nri] != gid)
1727 /* Energy group pair changed: new list */
1728 fep_list_new_nri_copy(nlist);
1731 nlist->gid[nri] = gid;
1734 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
1736 fep_list_new_nri_copy(nlist);
1740 /* Add it to the FEP list */
1741 nlist->jjnr[nlist->nrj] = aj;
1742 nlist->excl_fep[nlist->nrj] = (nbl->cj[cj_ind].excl >> (i*nbl->na_cj + j)) & 1;
1745 /* Exclude it from the normal list.
1746 * Note that the charge has been set to zero,
1747 * but we need to avoid 0/0, as perturbed atoms
1748 * can be on top of each other.
1750 nbl->cj[cj_ind].excl &= ~(1U << (i*nbl->na_cj + j));
1756 if (nlist->nrj > nlist->jindex[nri])
1758 /* Actually add this new, non-empty, list */
1760 nlist->jindex[nlist->nri] = nlist->nrj;
1767 /* All interactions are perturbed, we can skip this entry */
1768 nbl_ci->cj_ind_end = cj_ind_start;
1769 nbl->ncjInUse -= cj_ind_end - cj_ind_start;
1773 /* Return the index of atom a within a cluster */
1774 static inline int cj_mod_cj4(int cj)
1776 return cj & (c_nbnxnGpuJgroupSize - 1);
1779 /* Convert a j-cluster to a cj4 group */
1780 static inline int cj_to_cj4(int cj)
1782 return cj/c_nbnxnGpuJgroupSize;
1785 /* Return the index of an j-atom within a warp */
1786 static inline int a_mod_wj(int a)
1788 return a & (c_nbnxnGpuClusterSize/c_nbnxnGpuClusterpairSplit - 1);
1791 /* As make_fep_list above, but for super/sub lists. */
1792 static void make_fep_list(const nbnxn_search *nbs,
1793 const nbnxn_atomdata_t *nbat,
1794 NbnxnPairlistGpu *nbl,
1795 gmx_bool bDiagRemoved,
1796 const nbnxn_sci_t *nbl_sci,
1807 int ind_i, ind_j, ai, aj;
1811 const nbnxn_cj4_t *cj4;
1813 const int numJClusterGroups = nbl_sci->numJClusterGroups();
1814 if (numJClusterGroups == 0)
1820 const int sci = nbl_sci->sci;
1822 const int cj4_ind_start = nbl_sci->cj4_ind_start;
1823 const int cj4_ind_end = nbl_sci->cj4_ind_end;
1825 /* Here we process one super-cell, max #atoms na_sc, versus a list
1826 * cj4 entries, each with max c_nbnxnGpuJgroupSize cj's, each
1827 * of size na_cj atoms.
1828 * On the GPU we don't support energy groups (yet).
1829 * So for each of the na_sc i-atoms, we need max one FEP list
1830 * for each max_nrj_fep j-atoms.
1832 nri_max = nbl->na_sc*nbl->na_cj*(1 + (numJClusterGroups*c_nbnxnGpuJgroupSize)/max_nrj_fep);
1833 if (nlist->nri + nri_max > nlist->maxnri)
1835 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
1836 reallocate_nblist(nlist);
1839 /* Loop over the atoms in the i super-cluster */
1840 for (int c = 0; c < c_gpuNumClusterPerCell; c++)
1842 c_abs = sci*c_gpuNumClusterPerCell + c;
1844 for (int i = 0; i < nbl->na_ci; i++)
1846 ind_i = c_abs*nbl->na_ci + i;
1851 nlist->jindex[nri+1] = nlist->jindex[nri];
1852 nlist->iinr[nri] = ai;
1853 /* With GPUs, energy groups are not supported */
1854 nlist->gid[nri] = 0;
1855 nlist->shift[nri] = nbl_sci->shift & NBNXN_CI_SHIFT;
1857 bFEP_i = iGrid.atomIsPerturbed(c_abs - iGrid.cellOffset()*c_gpuNumClusterPerCell, i);
1859 xi = nbat->x()[ind_i*nbat->xstride+XX] + shx;
1860 yi = nbat->x()[ind_i*nbat->xstride+YY] + shy;
1861 zi = nbat->x()[ind_i*nbat->xstride+ZZ] + shz;
1863 const int nrjMax = nlist->nrj + numJClusterGroups*c_nbnxnGpuJgroupSize*nbl->na_cj;
1864 if (nrjMax > nlist->maxnrj)
1866 nlist->maxnrj = over_alloc_small(nrjMax);
1867 srenew(nlist->jjnr, nlist->maxnrj);
1868 srenew(nlist->excl_fep, nlist->maxnrj);
1871 for (int cj4_ind = cj4_ind_start; cj4_ind < cj4_ind_end; cj4_ind++)
1873 cj4 = &nbl->cj4[cj4_ind];
1875 for (int gcj = 0; gcj < c_nbnxnGpuJgroupSize; gcj++)
1877 if ((cj4->imei[0].imask & (1U << (gcj*c_gpuNumClusterPerCell + c))) == 0)
1879 /* Skip this ci for this cj */
1884 cj4->cj[gcj] - jGrid.cellOffset()*c_gpuNumClusterPerCell;
1886 if (bFEP_i || jGrid.clusterIsPerturbed(cjr))
1888 for (int j = 0; j < nbl->na_cj; j++)
1890 /* Is this interaction perturbed and not excluded? */
1891 ind_j = (jGrid.cellOffset()*c_gpuNumClusterPerCell + cjr)*nbl->na_cj + j;
1894 (bFEP_i || jGrid.atomIsPerturbed(cjr, j)) &&
1895 (!bDiagRemoved || ind_j >= ind_i))
1898 unsigned int excl_bit;
1901 const int jHalf = j/(c_nbnxnGpuClusterSize/c_nbnxnGpuClusterpairSplit);
1902 nbnxn_excl_t *excl =
1903 get_exclusion_mask(nbl, cj4_ind, jHalf);
1905 excl_pair = a_mod_wj(j)*nbl->na_ci + i;
1906 excl_bit = (1U << (gcj*c_gpuNumClusterPerCell + c));
1908 dx = nbat->x()[ind_j*nbat->xstride+XX] - xi;
1909 dy = nbat->x()[ind_j*nbat->xstride+YY] - yi;
1910 dz = nbat->x()[ind_j*nbat->xstride+ZZ] - zi;
1912 /* The unpruned GPU list has more than 2/3
1913 * of the atom pairs beyond rlist. Using
1914 * this list will cause a lot of overhead
1915 * in the CPU FEP kernels, especially
1916 * relative to the fast GPU kernels.
1917 * So we prune the FEP list here.
1919 if (dx*dx + dy*dy + dz*dz < rlist_fep2)
1921 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
1923 fep_list_new_nri_copy(nlist);
1927 /* Add it to the FEP list */
1928 nlist->jjnr[nlist->nrj] = aj;
1929 nlist->excl_fep[nlist->nrj] = (excl->pair[excl_pair] & excl_bit) ? 1 : 0;
1933 /* Exclude it from the normal list.
1934 * Note that the charge and LJ parameters have
1935 * been set to zero, but we need to avoid 0/0,
1936 * as perturbed atoms can be on top of each other.
1938 excl->pair[excl_pair] &= ~excl_bit;
1942 /* Note that we could mask out this pair in imask
1943 * if all i- and/or all j-particles are perturbed.
1944 * But since the perturbed pairs on the CPU will
1945 * take an order of magnitude more time, the GPU
1946 * will finish before the CPU and there is no gain.
1952 if (nlist->nrj > nlist->jindex[nri])
1954 /* Actually add this new, non-empty, list */
1956 nlist->jindex[nlist->nri] = nlist->nrj;
1963 /* Set all atom-pair exclusions for a GPU type list i-entry
1965 * Sets all atom-pair exclusions from the topology stored in exclusions
1966 * as masks in the pair-list for i-super-cluster list entry iEntry.
1969 setExclusionsForIEntry(const nbnxn_search *nbs,
1970 NbnxnPairlistGpu *nbl,
1971 gmx_bool diagRemoved,
1972 int gmx_unused na_cj_2log,
1973 const nbnxn_sci_t &iEntry,
1974 const t_blocka &exclusions)
1976 if (iEntry.numJClusterGroups() == 0)
1982 /* Set the search ranges using start and end j-cluster indices.
1983 * Note that here we can not use cj4_ind_end, since the last cj4
1984 * can be only partially filled, so we use cj_ind.
1986 const JListRanges ranges(iEntry.cj4_ind_start*c_nbnxnGpuJgroupSize,
1988 gmx::makeConstArrayRef(nbl->cj4));
1990 GMX_ASSERT(nbl->na_ci == c_nbnxnGpuClusterSize, "na_ci should match the GPU cluster size");
1991 constexpr int c_clusterSize = c_nbnxnGpuClusterSize;
1992 constexpr int c_superClusterSize = c_nbnxnGpuNumClusterPerSupercluster*c_nbnxnGpuClusterSize;
1994 const int iSuperCluster = iEntry.sci;
1996 gmx::ArrayRef<const int> cell = nbs->cell;
1998 /* Loop over the atoms in the i super-cluster */
1999 for (int i = 0; i < c_superClusterSize; i++)
2001 const int iIndex = iSuperCluster*c_superClusterSize + i;
2002 const int iAtom = nbs->a[iIndex];
2005 const int iCluster = i/c_clusterSize;
2007 /* Loop over the topology-based exclusions for this i-atom */
2008 for (int exclIndex = exclusions.index[iAtom]; exclIndex < exclusions.index[iAtom + 1]; exclIndex++)
2010 const int jAtom = exclusions.a[exclIndex];
2014 /* The self exclusions are already set, save some time */
2018 /* Get the index of the j-atom in the nbnxn atom data */
2019 const int jIndex = cell[jAtom];
2021 /* Without shifts we only calculate interactions j>i
2022 * for one-way pair-lists.
2024 /* NOTE: We would like to use iIndex on the right hand side,
2025 * but that makes this routine 25% slower with gcc6/7.
2026 * Even using c_superClusterSize makes it slower.
2027 * Either of these changes triggers peeling of the exclIndex
2028 * loop, which apparently leads to far less efficient code.
2030 if (diagRemoved && jIndex <= iSuperCluster*nbl->na_sc + i)
2035 const int jCluster = jIndex/c_clusterSize;
2037 /* Check whether the cluster is in our list? */
2038 if (jCluster >= ranges.cjFirst && jCluster <= ranges.cjLast)
2041 findJClusterInJList(jCluster, ranges,
2042 gmx::makeConstArrayRef(nbl->cj4));
2046 /* We found an exclusion, clear the corresponding
2049 const unsigned int pairMask = (1U << (cj_mod_cj4(index)*c_gpuNumClusterPerCell + iCluster));
2050 /* Check if the i-cluster interacts with the j-cluster */
2051 if (nbl_imask0(nbl, index) & pairMask)
2053 const int innerI = (i & (c_clusterSize - 1));
2054 const int innerJ = (jIndex & (c_clusterSize - 1));
2056 /* Determine which j-half (CUDA warp) we are in */
2057 const int jHalf = innerJ/(c_clusterSize/c_nbnxnGpuClusterpairSplit);
2059 nbnxn_excl_t *interactionMask =
2060 get_exclusion_mask(nbl, cj_to_cj4(index), jHalf);
2062 interactionMask->pair[a_mod_wj(innerJ)*c_clusterSize + innerI] &= ~pairMask;
2071 /* Make a new ci entry at the back of nbl->ci */
2072 static void addNewIEntry(NbnxnPairlistCpu *nbl, int ci, int shift, int flags)
2076 ciEntry.shift = shift;
2077 /* Store the interaction flags along with the shift */
2078 ciEntry.shift |= flags;
2079 ciEntry.cj_ind_start = nbl->cj.size();
2080 ciEntry.cj_ind_end = nbl->cj.size();
2081 nbl->ci.push_back(ciEntry);
2084 /* Make a new sci entry at index nbl->nsci */
2085 static void addNewIEntry(NbnxnPairlistGpu *nbl, int sci, int shift, int gmx_unused flags)
2087 nbnxn_sci_t sciEntry;
2089 sciEntry.shift = shift;
2090 sciEntry.cj4_ind_start = nbl->cj4.size();
2091 sciEntry.cj4_ind_end = nbl->cj4.size();
2093 nbl->sci.push_back(sciEntry);
2096 /* Sort the simple j-list cj on exclusions.
2097 * Entries with exclusions will all be sorted to the beginning of the list.
2099 static void sort_cj_excl(nbnxn_cj_t *cj, int ncj,
2100 NbnxnPairlistCpuWork *work)
2102 work->cj.resize(ncj);
2104 /* Make a list of the j-cells involving exclusions */
2106 for (int j = 0; j < ncj; j++)
2108 if (cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
2110 work->cj[jnew++] = cj[j];
2113 /* Check if there are exclusions at all or not just the first entry */
2114 if (!((jnew == 0) ||
2115 (jnew == 1 && cj[0].excl != NBNXN_INTERACTION_MASK_ALL)))
2117 for (int j = 0; j < ncj; j++)
2119 if (cj[j].excl == NBNXN_INTERACTION_MASK_ALL)
2121 work->cj[jnew++] = cj[j];
2124 for (int j = 0; j < ncj; j++)
2126 cj[j] = work->cj[j];
2131 /* Close this simple list i entry */
2132 static void closeIEntry(NbnxnPairlistCpu *nbl,
2133 int gmx_unused sp_max_av,
2134 gmx_bool gmx_unused progBal,
2135 float gmx_unused nsp_tot_est,
2136 int gmx_unused thread,
2137 int gmx_unused nthread)
2139 nbnxn_ci_t &ciEntry = nbl->ci.back();
2141 /* All content of the new ci entry have already been filled correctly,
2142 * we only need to sort and increase counts or remove the entry when empty.
2144 const int jlen = ciEntry.cj_ind_end - ciEntry.cj_ind_start;
2147 sort_cj_excl(nbl->cj.data() + ciEntry.cj_ind_start, jlen, nbl->work);
2149 /* The counts below are used for non-bonded pair/flop counts
2150 * and should therefore match the available kernel setups.
2152 if (!(ciEntry.shift & NBNXN_CI_DO_COUL(0)))
2154 nbl->work->ncj_noq += jlen;
2156 else if ((ciEntry.shift & NBNXN_CI_HALF_LJ(0)) ||
2157 !(ciEntry.shift & NBNXN_CI_DO_LJ(0)))
2159 nbl->work->ncj_hlj += jlen;
2164 /* Entry is empty: remove it */
2169 /* Split sci entry for load balancing on the GPU.
2170 * Splitting ensures we have enough lists to fully utilize the whole GPU.
2171 * With progBal we generate progressively smaller lists, which improves
2172 * load balancing. As we only know the current count on our own thread,
2173 * we will need to estimate the current total amount of i-entries.
2174 * As the lists get concatenated later, this estimate depends
2175 * both on nthread and our own thread index.
2177 static void split_sci_entry(NbnxnPairlistGpu *nbl,
2179 gmx_bool progBal, float nsp_tot_est,
2180 int thread, int nthread)
2188 /* Estimate the total numbers of ci's of the nblist combined
2189 * over all threads using the target number of ci's.
2191 nsp_est = (nsp_tot_est*thread)/nthread + nbl->nci_tot;
2193 /* The first ci blocks should be larger, to avoid overhead.
2194 * The last ci blocks should be smaller, to improve load balancing.
2195 * The factor 3/2 makes the first block 3/2 times the target average
2196 * and ensures that the total number of blocks end up equal to
2197 * that of equally sized blocks of size nsp_target_av.
2199 nsp_max = static_cast<int>(nsp_target_av*(nsp_tot_est*1.5/(nsp_est + nsp_tot_est)));
2203 nsp_max = nsp_target_av;
2206 const int cj4_start = nbl->sci.back().cj4_ind_start;
2207 const int cj4_end = nbl->sci.back().cj4_ind_end;
2208 const int j4len = cj4_end - cj4_start;
2210 if (j4len > 1 && j4len*c_gpuNumClusterPerCell*c_nbnxnGpuJgroupSize > nsp_max)
2212 /* Modify the last ci entry and process the cj4's again */
2218 for (int cj4 = cj4_start; cj4 < cj4_end; cj4++)
2220 int nsp_cj4_p = nsp_cj4;
2221 /* Count the number of cluster pairs in this cj4 group */
2223 for (int p = 0; p < c_gpuNumClusterPerCell*c_nbnxnGpuJgroupSize; p++)
2225 nsp_cj4 += (nbl->cj4[cj4].imei[0].imask >> p) & 1;
2228 /* If adding the current cj4 with nsp_cj4 pairs get us further
2229 * away from our target nsp_max, split the list before this cj4.
2231 if (nsp > 0 && nsp_max - nsp < nsp + nsp_cj4 - nsp_max)
2233 /* Split the list at cj4 */
2234 nbl->sci.back().cj4_ind_end = cj4;
2235 /* Create a new sci entry */
2237 sciNew.sci = nbl->sci.back().sci;
2238 sciNew.shift = nbl->sci.back().shift;
2239 sciNew.cj4_ind_start = cj4;
2240 nbl->sci.push_back(sciNew);
2243 nsp_cj4_e = nsp_cj4_p;
2249 /* Put the remaining cj4's in the last sci entry */
2250 nbl->sci.back().cj4_ind_end = cj4_end;
2252 /* Possibly balance out the last two sci's
2253 * by moving the last cj4 of the second last sci.
2255 if (nsp_sci - nsp_cj4_e >= nsp + nsp_cj4_e)
2257 GMX_ASSERT(nbl->sci.size() >= 2, "We expect at least two elements");
2258 nbl->sci[nbl->sci.size() - 2].cj4_ind_end--;
2259 nbl->sci[nbl->sci.size() - 1].cj4_ind_start--;
2264 /* Clost this super/sub list i entry */
2265 static void closeIEntry(NbnxnPairlistGpu *nbl,
2267 gmx_bool progBal, float nsp_tot_est,
2268 int thread, int nthread)
2270 nbnxn_sci_t &sciEntry = *getOpenIEntry(nbl);
2272 /* All content of the new ci entry have already been filled correctly,
2273 * we only need to, potentially, split or remove the entry when empty.
2275 int j4len = sciEntry.numJClusterGroups();
2278 /* We can only have complete blocks of 4 j-entries in a list,
2279 * so round the count up before closing.
2281 int ncj4 = (nbl->work->cj_ind + c_nbnxnGpuJgroupSize - 1)/c_nbnxnGpuJgroupSize;
2282 nbl->work->cj_ind = ncj4*c_nbnxnGpuJgroupSize;
2286 /* Measure the size of the new entry and potentially split it */
2287 split_sci_entry(nbl, nsp_max_av, progBal, nsp_tot_est,
2293 /* Entry is empty: remove it */
2294 nbl->sci.pop_back();
2298 /* Syncs the working array before adding another grid pair to the GPU list */
2299 static void sync_work(NbnxnPairlistCpu gmx_unused *nbl)
2303 /* Syncs the working array before adding another grid pair to the GPU list */
2304 static void sync_work(NbnxnPairlistGpu *nbl)
2306 nbl->work->cj_ind = nbl->cj4.size()*c_nbnxnGpuJgroupSize;
2309 /* Clears an NbnxnPairlistCpu data structure */
2310 static void clear_pairlist(NbnxnPairlistCpu *nbl)
2316 nbl->ciOuter.clear();
2317 nbl->cjOuter.clear();
2319 nbl->work->ncj_noq = 0;
2320 nbl->work->ncj_hlj = 0;
2323 /* Clears an NbnxnPairlistGpu data structure */
2324 static void clear_pairlist(NbnxnPairlistGpu *nbl)
2328 nbl->excl.resize(1);
2332 /* Clears a group scheme pair list */
2333 static void clear_pairlist_fep(t_nblist *nl)
2337 if (nl->jindex == nullptr)
2339 snew(nl->jindex, 1);
2344 /* Sets a simple list i-cell bounding box, including PBC shift */
2345 static inline void set_icell_bb_simple(gmx::ArrayRef<const BoundingBox> bb,
2347 real shx, real shy, real shz,
2350 bb_ci->lower.x = bb[ci].lower.x + shx;
2351 bb_ci->lower.y = bb[ci].lower.y + shy;
2352 bb_ci->lower.z = bb[ci].lower.z + shz;
2353 bb_ci->upper.x = bb[ci].upper.x + shx;
2354 bb_ci->upper.y = bb[ci].upper.y + shy;
2355 bb_ci->upper.z = bb[ci].upper.z + shz;
2358 /* Sets a simple list i-cell bounding box, including PBC shift */
2359 static inline void set_icell_bb(const Grid &iGrid,
2361 real shx, real shy, real shz,
2362 NbnxnPairlistCpuWork *work)
2364 set_icell_bb_simple(iGrid.iBoundingBoxes(), ci, shx, shy, shz,
2365 &work->iClusterData.bb[0]);
2369 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
2370 static void set_icell_bbxxxx_supersub(gmx::ArrayRef<const float> bb,
2372 real shx, real shy, real shz,
2375 int ia = ci*(c_gpuNumClusterPerCell >> STRIDE_PBB_2LOG)*NNBSBB_XXXX;
2376 for (int m = 0; m < (c_gpuNumClusterPerCell >> STRIDE_PBB_2LOG)*NNBSBB_XXXX; m += NNBSBB_XXXX)
2378 for (int i = 0; i < STRIDE_PBB; i++)
2380 bb_ci[m+0*STRIDE_PBB+i] = bb[ia+m+0*STRIDE_PBB+i] + shx;
2381 bb_ci[m+1*STRIDE_PBB+i] = bb[ia+m+1*STRIDE_PBB+i] + shy;
2382 bb_ci[m+2*STRIDE_PBB+i] = bb[ia+m+2*STRIDE_PBB+i] + shz;
2383 bb_ci[m+3*STRIDE_PBB+i] = bb[ia+m+3*STRIDE_PBB+i] + shx;
2384 bb_ci[m+4*STRIDE_PBB+i] = bb[ia+m+4*STRIDE_PBB+i] + shy;
2385 bb_ci[m+5*STRIDE_PBB+i] = bb[ia+m+5*STRIDE_PBB+i] + shz;
2391 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
2392 gmx_unused static void set_icell_bb_supersub(gmx::ArrayRef<const BoundingBox> bb,
2394 real shx, real shy, real shz,
2397 for (int i = 0; i < c_gpuNumClusterPerCell; i++)
2399 set_icell_bb_simple(bb, ci*c_gpuNumClusterPerCell+i,
2405 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
2406 gmx_unused static void set_icell_bb(const Grid &iGrid,
2408 real shx, real shy, real shz,
2409 NbnxnPairlistGpuWork *work)
2412 set_icell_bbxxxx_supersub(iGrid.packedBoundingBoxes(), ci, shx, shy, shz,
2413 work->iSuperClusterData.bbPacked.data());
2415 set_icell_bb_supersub(iGrid.iBoundingBoxes(), ci, shx, shy, shz,
2416 work->iSuperClusterData.bb.data());
2420 /* Copies PBC shifted i-cell atom coordinates x,y,z to working array */
2421 static void icell_set_x_simple(int ci,
2422 real shx, real shy, real shz,
2423 int stride, const real *x,
2424 NbnxnPairlistCpuWork::IClusterData *iClusterData)
2426 const int ia = ci*c_nbnxnCpuIClusterSize;
2428 for (int i = 0; i < c_nbnxnCpuIClusterSize; i++)
2430 iClusterData->x[i*STRIDE_XYZ+XX] = x[(ia+i)*stride+XX] + shx;
2431 iClusterData->x[i*STRIDE_XYZ+YY] = x[(ia+i)*stride+YY] + shy;
2432 iClusterData->x[i*STRIDE_XYZ+ZZ] = x[(ia+i)*stride+ZZ] + shz;
2436 static void icell_set_x(int ci,
2437 real shx, real shy, real shz,
2438 int stride, const real *x,
2439 const Nbnxm::KernelType kernelType,
2440 NbnxnPairlistCpuWork *work)
2445 #ifdef GMX_NBNXN_SIMD_4XN
2446 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
2447 icell_set_x_simd_4xn(ci, shx, shy, shz, stride, x, work);
2450 #ifdef GMX_NBNXN_SIMD_2XNN
2451 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
2452 icell_set_x_simd_2xnn(ci, shx, shy, shz, stride, x, work);
2456 case Nbnxm::KernelType::Cpu4x4_PlainC:
2457 icell_set_x_simple(ci, shx, shy, shz, stride, x, &work->iClusterData);
2460 GMX_ASSERT(false, "Unhandled case");
2465 /* Copies PBC shifted super-cell atom coordinates x,y,z to working array */
2466 static void icell_set_x(int ci,
2467 real shx, real shy, real shz,
2468 int stride, const real *x,
2469 Nbnxm::KernelType gmx_unused kernelType,
2470 NbnxnPairlistGpuWork *work)
2472 #if !GMX_SIMD4_HAVE_REAL
2474 real * x_ci = work->iSuperClusterData.x.data();
2476 int ia = ci*c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize;
2477 for (int i = 0; i < c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize; i++)
2479 x_ci[i*DIM + XX] = x[(ia+i)*stride + XX] + shx;
2480 x_ci[i*DIM + YY] = x[(ia+i)*stride + YY] + shy;
2481 x_ci[i*DIM + ZZ] = x[(ia+i)*stride + ZZ] + shz;
2484 #else /* !GMX_SIMD4_HAVE_REAL */
2486 real * x_ci = work->iSuperClusterData.xSimd.data();
2488 for (int si = 0; si < c_gpuNumClusterPerCell; si++)
2490 for (int i = 0; i < c_nbnxnGpuClusterSize; i += GMX_SIMD4_WIDTH)
2492 int io = si*c_nbnxnGpuClusterSize + i;
2493 int ia = ci*c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize + io;
2494 for (int j = 0; j < GMX_SIMD4_WIDTH; j++)
2496 x_ci[io*DIM + j + XX*GMX_SIMD4_WIDTH] = x[(ia + j)*stride + XX] + shx;
2497 x_ci[io*DIM + j + YY*GMX_SIMD4_WIDTH] = x[(ia + j)*stride + YY] + shy;
2498 x_ci[io*DIM + j + ZZ*GMX_SIMD4_WIDTH] = x[(ia + j)*stride + ZZ] + shz;
2503 #endif /* !GMX_SIMD4_HAVE_REAL */
2506 static real minimum_subgrid_size_xy(const Grid &grid)
2508 const Grid::Dimensions &dims = grid.dimensions();
2510 if (grid.geometry().isSimple)
2512 return std::min(dims.cellSize[XX], dims.cellSize[YY]);
2516 return std::min(dims.cellSize[XX]/c_gpuNumClusterPerCellX,
2517 dims.cellSize[YY]/c_gpuNumClusterPerCellY);
2521 static real effective_buffer_1x1_vs_MxN(const Grid &iGrid,
2524 const real eff_1x1_buffer_fac_overest = 0.1;
2526 /* Determine an atom-pair list cut-off buffer size for atom pairs,
2527 * to be added to rlist (including buffer) used for MxN.
2528 * This is for converting an MxN list to a 1x1 list. This means we can't
2529 * use the normal buffer estimate, as we have an MxN list in which
2530 * some atom pairs beyond rlist are missing. We want to capture
2531 * the beneficial effect of buffering by extra pairs just outside rlist,
2532 * while removing the useless pairs that are further away from rlist.
2533 * (Also the buffer could have been set manually not using the estimate.)
2534 * This buffer size is an overestimate.
2535 * We add 10% of the smallest grid sub-cell dimensions.
2536 * Note that the z-size differs per cell and we don't use this,
2537 * so we overestimate.
2538 * With PME, the 10% value gives a buffer that is somewhat larger
2539 * than the effective buffer with a tolerance of 0.005 kJ/mol/ps.
2540 * Smaller tolerances or using RF lead to a smaller effective buffer,
2541 * so 10% gives a safe overestimate.
2543 return eff_1x1_buffer_fac_overest*(minimum_subgrid_size_xy(iGrid) +
2544 minimum_subgrid_size_xy(jGrid));
2547 /* Estimates the interaction volume^2 for non-local interactions */
2548 static real nonlocal_vol2(const struct gmx_domdec_zones_t *zones, const rvec ls, real r)
2556 /* Here we simply add up the volumes of 1, 2 or 3 1D decomposition
2557 * not home interaction volume^2. As these volumes are not additive,
2558 * this is an overestimate, but it would only be significant in the limit
2559 * of small cells, where we anyhow need to split the lists into
2560 * as small parts as possible.
2563 for (int z = 0; z < zones->n; z++)
2565 if (zones->shift[z][XX] + zones->shift[z][YY] + zones->shift[z][ZZ] == 1)
2570 for (int d = 0; d < DIM; d++)
2572 if (zones->shift[z][d] == 0)
2576 za *= zones->size[z].x1[d] - zones->size[z].x0[d];
2580 /* 4 octants of a sphere */
2581 vold_est = 0.25*M_PI*r*r*r*r;
2582 /* 4 quarter pie slices on the edges */
2583 vold_est += 4*cl*M_PI/6.0*r*r*r;
2584 /* One rectangular volume on a face */
2585 vold_est += ca*0.5*r*r;
2587 vol2_est_tot += vold_est*za;
2591 return vol2_est_tot;
2594 /* Estimates the average size of a full j-list for super/sub setup */
2595 static void get_nsubpair_target(const nbnxn_search *nbs,
2596 const InteractionLocality iloc,
2598 const int min_ci_balanced,
2599 int *nsubpair_target,
2600 float *nsubpair_tot_est)
2602 /* The target value of 36 seems to be the optimum for Kepler.
2603 * Maxwell is less sensitive to the exact value.
2605 const int nsubpair_target_min = 36;
2606 real r_eff_sup, vol_est, nsp_est, nsp_est_nl;
2608 const Grid &grid = nbs->grid[0];
2610 /* We don't need to balance list sizes if:
2611 * - We didn't request balancing.
2612 * - The number of grid cells >= the number of lists requested,
2613 * since we will always generate at least #cells lists.
2614 * - We don't have any cells, since then there won't be any lists.
2616 if (min_ci_balanced <= 0 || grid.numCells() >= min_ci_balanced || grid.numCells() == 0)
2618 /* nsubpair_target==0 signals no balancing */
2619 *nsubpair_target = 0;
2620 *nsubpair_tot_est = 0;
2626 const int numAtomsCluster = grid.geometry().numAtomsICluster;
2627 const Grid::Dimensions &dims = grid.dimensions();
2629 ls[XX] = dims.cellSize[XX]/c_gpuNumClusterPerCellX;
2630 ls[YY] = dims.cellSize[YY]/c_gpuNumClusterPerCellY;
2631 ls[ZZ] = numAtomsCluster/(dims.atomDensity*ls[XX]*ls[YY]);
2633 /* The formulas below are a heuristic estimate of the average nsj per si*/
2634 r_eff_sup = rlist + nbnxn_get_rlist_effective_inc(numAtomsCluster, ls);
2636 if (!nbs->DomDec || nbs->zones->n == 1)
2643 gmx::square(dims.atomDensity/numAtomsCluster)*
2644 nonlocal_vol2(nbs->zones, ls, r_eff_sup);
2647 if (iloc == InteractionLocality::Local)
2649 /* Sub-cell interacts with itself */
2650 vol_est = ls[XX]*ls[YY]*ls[ZZ];
2651 /* 6/2 rectangular volume on the faces */
2652 vol_est += (ls[XX]*ls[YY] + ls[XX]*ls[ZZ] + ls[YY]*ls[ZZ])*r_eff_sup;
2653 /* 12/2 quarter pie slices on the edges */
2654 vol_est += 2*(ls[XX] + ls[YY] + ls[ZZ])*0.25*M_PI*gmx::square(r_eff_sup);
2655 /* 4 octants of a sphere */
2656 vol_est += 0.5*4.0/3.0*M_PI*gmx::power3(r_eff_sup);
2658 /* Estimate the number of cluster pairs as the local number of
2659 * clusters times the volume they interact with times the density.
2661 nsp_est = grid.numClusters()*vol_est*dims.atomDensity/numAtomsCluster;
2663 /* Subtract the non-local pair count */
2664 nsp_est -= nsp_est_nl;
2666 /* For small cut-offs nsp_est will be an underesimate.
2667 * With DD nsp_est_nl is an overestimate so nsp_est can get negative.
2668 * So to avoid too small or negative nsp_est we set a minimum of
2669 * all cells interacting with all 3^3 direct neighbors (3^3-1)/2+1=14.
2670 * This might be a slight overestimate for small non-periodic groups of
2671 * atoms as will occur for a local domain with DD, but for small
2672 * groups of atoms we'll anyhow be limited by nsubpair_target_min,
2673 * so this overestimation will not matter.
2675 nsp_est = std::max(nsp_est, grid.numClusters()*14._real);
2679 fprintf(debug, "nsp_est local %5.1f non-local %5.1f\n",
2680 nsp_est, nsp_est_nl);
2685 nsp_est = nsp_est_nl;
2688 /* Thus the (average) maximum j-list size should be as follows.
2689 * Since there is overhead, we shouldn't make the lists too small
2690 * (and we can't chop up j-groups) so we use a minimum target size of 36.
2692 *nsubpair_target = std::max(nsubpair_target_min,
2693 roundToInt(nsp_est/min_ci_balanced));
2694 *nsubpair_tot_est = static_cast<int>(nsp_est);
2698 fprintf(debug, "nbl nsp estimate %.1f, nsubpair_target %d\n",
2699 nsp_est, *nsubpair_target);
2703 /* Debug list print function */
2704 static void print_nblist_ci_cj(FILE *fp, const NbnxnPairlistCpu *nbl)
2706 for (const nbnxn_ci_t &ciEntry : nbl->ci)
2708 fprintf(fp, "ci %4d shift %2d ncj %3d\n",
2709 ciEntry.ci, ciEntry.shift,
2710 ciEntry.cj_ind_end - ciEntry.cj_ind_start);
2712 for (int j = ciEntry.cj_ind_start; j < ciEntry.cj_ind_end; j++)
2714 fprintf(fp, " cj %5d imask %x\n",
2721 /* Debug list print function */
2722 static void print_nblist_sci_cj(FILE *fp, const NbnxnPairlistGpu *nbl)
2724 for (const nbnxn_sci_t &sci : nbl->sci)
2726 fprintf(fp, "ci %4d shift %2d ncj4 %2d\n",
2728 sci.numJClusterGroups());
2731 for (int j4 = sci.cj4_ind_start; j4 < sci.cj4_ind_end; j4++)
2733 for (int j = 0; j < c_nbnxnGpuJgroupSize; j++)
2735 fprintf(fp, " sj %5d imask %x\n",
2737 nbl->cj4[j4].imei[0].imask);
2738 for (int si = 0; si < c_gpuNumClusterPerCell; si++)
2740 if (nbl->cj4[j4].imei[0].imask & (1U << (j*c_gpuNumClusterPerCell + si)))
2747 fprintf(fp, "ci %4d shift %2d ncj4 %2d ncp %3d\n",
2749 sci.numJClusterGroups(),
2754 /* Combine pair lists *nbl generated on multiple threads nblc */
2755 static void combine_nblists(int nnbl, NbnxnPairlistGpu **nbl,
2756 NbnxnPairlistGpu *nblc)
2758 int nsci = nblc->sci.size();
2759 int ncj4 = nblc->cj4.size();
2760 int nexcl = nblc->excl.size();
2761 for (int i = 0; i < nnbl; i++)
2763 nsci += nbl[i]->sci.size();
2764 ncj4 += nbl[i]->cj4.size();
2765 nexcl += nbl[i]->excl.size();
2768 /* Resize with the final, combined size, so we can fill in parallel */
2769 /* NOTE: For better performance we should use default initialization */
2770 nblc->sci.resize(nsci);
2771 nblc->cj4.resize(ncj4);
2772 nblc->excl.resize(nexcl);
2774 /* Each thread should copy its own data to the combined arrays,
2775 * as otherwise data will go back and forth between different caches.
2777 #if GMX_OPENMP && !(defined __clang_analyzer__)
2778 int nthreads = gmx_omp_nthreads_get(emntPairsearch);
2781 #pragma omp parallel for num_threads(nthreads) schedule(static)
2782 for (int n = 0; n < nnbl; n++)
2786 /* Determine the offset in the combined data for our thread.
2787 * Note that the original sizes in nblc are lost.
2789 int sci_offset = nsci;
2790 int cj4_offset = ncj4;
2791 int excl_offset = nexcl;
2793 for (int i = n; i < nnbl; i++)
2795 sci_offset -= nbl[i]->sci.size();
2796 cj4_offset -= nbl[i]->cj4.size();
2797 excl_offset -= nbl[i]->excl.size();
2800 const NbnxnPairlistGpu &nbli = *nbl[n];
2802 for (size_t i = 0; i < nbli.sci.size(); i++)
2804 nblc->sci[sci_offset + i] = nbli.sci[i];
2805 nblc->sci[sci_offset + i].cj4_ind_start += cj4_offset;
2806 nblc->sci[sci_offset + i].cj4_ind_end += cj4_offset;
2809 for (size_t j4 = 0; j4 < nbli.cj4.size(); j4++)
2811 nblc->cj4[cj4_offset + j4] = nbli.cj4[j4];
2812 nblc->cj4[cj4_offset + j4].imei[0].excl_ind += excl_offset;
2813 nblc->cj4[cj4_offset + j4].imei[1].excl_ind += excl_offset;
2816 for (size_t j4 = 0; j4 < nbli.excl.size(); j4++)
2818 nblc->excl[excl_offset + j4] = nbli.excl[j4];
2821 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
2824 for (int n = 0; n < nnbl; n++)
2826 nblc->nci_tot += nbl[n]->nci_tot;
2830 static void balance_fep_lists(const nbnxn_search *nbs,
2831 nbnxn_pairlist_set_t *nbl_lists)
2834 int nri_tot, nrj_tot, nrj_target;
2838 nnbl = nbl_lists->nnbl;
2842 /* Nothing to balance */
2846 /* Count the total i-lists and pairs */
2849 for (int th = 0; th < nnbl; th++)
2851 nri_tot += nbl_lists->nbl_fep[th]->nri;
2852 nrj_tot += nbl_lists->nbl_fep[th]->nrj;
2855 nrj_target = (nrj_tot + nnbl - 1)/nnbl;
2857 assert(gmx_omp_nthreads_get(emntNonbonded) == nnbl);
2859 #pragma omp parallel for schedule(static) num_threads(nnbl)
2860 for (int th = 0; th < nnbl; th++)
2864 t_nblist *nbl = nbs->work[th].nbl_fep.get();
2866 /* Note that here we allocate for the total size, instead of
2867 * a per-thread esimate (which is hard to obtain).
2869 if (nri_tot > nbl->maxnri)
2871 nbl->maxnri = over_alloc_large(nri_tot);
2872 reallocate_nblist(nbl);
2874 if (nri_tot > nbl->maxnri || nrj_tot > nbl->maxnrj)
2876 nbl->maxnrj = over_alloc_small(nrj_tot);
2877 srenew(nbl->jjnr, nbl->maxnrj);
2878 srenew(nbl->excl_fep, nbl->maxnrj);
2881 clear_pairlist_fep(nbl);
2883 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
2886 /* Loop over the source lists and assign and copy i-entries */
2888 nbld = nbs->work[th_dest].nbl_fep.get();
2889 for (int th = 0; th < nnbl; th++)
2893 nbls = nbl_lists->nbl_fep[th];
2895 for (int i = 0; i < nbls->nri; i++)
2899 /* The number of pairs in this i-entry */
2900 nrj = nbls->jindex[i+1] - nbls->jindex[i];
2902 /* Decide if list th_dest is too large and we should procede
2903 * to the next destination list.
2905 if (th_dest+1 < nnbl && nbld->nrj > 0 &&
2906 nbld->nrj + nrj - nrj_target > nrj_target - nbld->nrj)
2909 nbld = nbs->work[th_dest].nbl_fep.get();
2912 nbld->iinr[nbld->nri] = nbls->iinr[i];
2913 nbld->gid[nbld->nri] = nbls->gid[i];
2914 nbld->shift[nbld->nri] = nbls->shift[i];
2916 for (int j = nbls->jindex[i]; j < nbls->jindex[i+1]; j++)
2918 nbld->jjnr[nbld->nrj] = nbls->jjnr[j];
2919 nbld->excl_fep[nbld->nrj] = nbls->excl_fep[j];
2923 nbld->jindex[nbld->nri] = nbld->nrj;
2927 /* Swap the list pointers */
2928 for (int th = 0; th < nnbl; th++)
2930 t_nblist *nbl_tmp = nbs->work[th].nbl_fep.release();
2931 nbs->work[th].nbl_fep.reset(nbl_lists->nbl_fep[th]);
2932 nbl_lists->nbl_fep[th] = nbl_tmp;
2936 fprintf(debug, "nbl_fep[%d] nri %4d nrj %4d\n",
2938 nbl_lists->nbl_fep[th]->nri,
2939 nbl_lists->nbl_fep[th]->nrj);
2944 /* Returns the next ci to be processes by our thread */
2945 static gmx_bool next_ci(const Grid &grid,
2946 int nth, int ci_block,
2947 int *ci_x, int *ci_y,
2953 if (*ci_b == ci_block)
2955 /* Jump to the next block assigned to this task */
2956 *ci += (nth - 1)*ci_block;
2960 if (*ci >= grid.numCells())
2965 while (*ci >= grid.firstCellInColumn(*ci_x*grid.dimensions().numCells[YY] + *ci_y + 1))
2968 if (*ci_y == grid.dimensions().numCells[YY])
2978 /* Returns the distance^2 for which we put cell pairs in the list
2979 * without checking atom pair distances. This is usually < rlist^2.
2981 static float boundingbox_only_distance2(const Grid::Dimensions &iGridDims,
2982 const Grid::Dimensions &jGridDims,
2986 /* If the distance between two sub-cell bounding boxes is less
2987 * than this distance, do not check the distance between
2988 * all particle pairs in the sub-cell, since then it is likely
2989 * that the box pair has atom pairs within the cut-off.
2990 * We use the nblist cut-off minus 0.5 times the average x/y diagonal
2991 * spacing of the sub-cells. Around 40% of the checked pairs are pruned.
2992 * Using more than 0.5 gains at most 0.5%.
2993 * If forces are calculated more than twice, the performance gain
2994 * in the force calculation outweighs the cost of checking.
2995 * Note that with subcell lists, the atom-pair distance check
2996 * is only performed when only 1 out of 8 sub-cells in within range,
2997 * this is because the GPU is much faster than the cpu.
3002 bbx = 0.5*(iGridDims.cellSize[XX] + jGridDims.cellSize[XX]);
3003 bby = 0.5*(iGridDims.cellSize[YY] + jGridDims.cellSize[YY]);
3006 bbx /= c_gpuNumClusterPerCellX;
3007 bby /= c_gpuNumClusterPerCellY;
3010 rbb2 = std::max(0.0, rlist - 0.5*std::sqrt(bbx*bbx + bby*bby));
3016 return (float)((1+GMX_FLOAT_EPS)*rbb2);
3020 static int get_ci_block_size(const Grid &iGrid,
3021 gmx_bool bDomDec, int nth)
3023 const int ci_block_enum = 5;
3024 const int ci_block_denom = 11;
3025 const int ci_block_min_atoms = 16;
3028 /* Here we decide how to distribute the blocks over the threads.
3029 * We use prime numbers to try to avoid that the grid size becomes
3030 * a multiple of the number of threads, which would lead to some
3031 * threads getting "inner" pairs and others getting boundary pairs,
3032 * which in turns will lead to load imbalance between threads.
3033 * Set the block size as 5/11/ntask times the average number of cells
3034 * in a y,z slab. This should ensure a quite uniform distribution
3035 * of the grid parts of the different thread along all three grid
3036 * zone boundaries with 3D domain decomposition. At the same time
3037 * the blocks will not become too small.
3039 ci_block = (iGrid.numCells()*ci_block_enum)/(ci_block_denom*iGrid.dimensions().numCells[XX]*nth);
3041 const int numAtomsPerCell = iGrid.geometry().numAtomsPerCell;
3043 /* Ensure the blocks are not too small: avoids cache invalidation */
3044 if (ci_block*numAtomsPerCell < ci_block_min_atoms)
3046 ci_block = (ci_block_min_atoms + numAtomsPerCell - 1)/numAtomsPerCell;
3049 /* Without domain decomposition
3050 * or with less than 3 blocks per task, divide in nth blocks.
3052 if (!bDomDec || nth*3*ci_block > iGrid.numCells())
3054 ci_block = (iGrid.numCells() + nth - 1)/nth;
3057 if (ci_block > 1 && (nth - 1)*ci_block >= iGrid.numCells())
3059 /* Some threads have no work. Although reducing the block size
3060 * does not decrease the block count on the first few threads,
3061 * with GPUs better mixing of "upper" cells that have more empty
3062 * clusters results in a somewhat lower max load over all threads.
3063 * Without GPUs the regime of so few atoms per thread is less
3064 * performance relevant, but with 8-wide SIMD the same reasoning
3065 * applies, since the pair list uses 4 i-atom "sub-clusters".
3073 /* Returns the number of bits to right-shift a cluster index to obtain
3074 * the corresponding force buffer flag index.
3076 static int getBufferFlagShift(int numAtomsPerCluster)
3078 int bufferFlagShift = 0;
3079 while ((numAtomsPerCluster << bufferFlagShift) < NBNXN_BUFFERFLAG_SIZE)
3084 return bufferFlagShift;
3087 static bool pairlistIsSimple(const NbnxnPairlistCpu gmx_unused &pairlist)
3092 static bool pairlistIsSimple(const NbnxnPairlistGpu gmx_unused &pairlist)
3097 static void makeClusterListWrapper(NbnxnPairlistCpu *nbl,
3098 const Grid gmx_unused &iGrid,
3101 const int firstCell,
3103 const bool excludeSubDiagonal,
3104 const nbnxn_atomdata_t *nbat,
3107 const Nbnxm::KernelType kernelType,
3108 int *numDistanceChecks)
3112 case Nbnxm::KernelType::Cpu4x4_PlainC:
3113 makeClusterListSimple(jGrid,
3114 nbl, ci, firstCell, lastCell,
3120 #ifdef GMX_NBNXN_SIMD_4XN
3121 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
3122 makeClusterListSimd4xn(jGrid,
3123 nbl, ci, firstCell, lastCell,
3130 #ifdef GMX_NBNXN_SIMD_2XNN
3131 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
3132 makeClusterListSimd2xnn(jGrid,
3133 nbl, ci, firstCell, lastCell,
3141 GMX_ASSERT(false, "Unhandled kernel type");
3145 static void makeClusterListWrapper(NbnxnPairlistGpu *nbl,
3146 const Grid &gmx_unused iGrid,
3149 const int firstCell,
3151 const bool excludeSubDiagonal,
3152 const nbnxn_atomdata_t *nbat,
3155 Nbnxm::KernelType gmx_unused kernelType,
3156 int *numDistanceChecks)
3158 for (int cj = firstCell; cj <= lastCell; cj++)
3160 make_cluster_list_supersub(iGrid, jGrid,
3163 nbat->xstride, nbat->x().data(),
3169 static int getNumSimpleJClustersInList(const NbnxnPairlistCpu &nbl)
3171 return nbl.cj.size();
3174 static int getNumSimpleJClustersInList(const gmx_unused NbnxnPairlistGpu &nbl)
3179 static void incrementNumSimpleJClustersInList(NbnxnPairlistCpu *nbl,
3182 nbl->ncjInUse += nbl->cj.size() - ncj_old_j;
3185 static void incrementNumSimpleJClustersInList(NbnxnPairlistGpu gmx_unused *nbl,
3186 int gmx_unused ncj_old_j)
3190 static void checkListSizeConsistency(const NbnxnPairlistCpu &nbl,
3191 const bool haveFreeEnergy)
3193 GMX_RELEASE_ASSERT(static_cast<size_t>(nbl.ncjInUse) == nbl.cj.size() || haveFreeEnergy,
3194 "Without free-energy all cj pair-list entries should be in use. "
3195 "Note that subsequent code does not make use of the equality, "
3196 "this check is only here to catch bugs");
3199 static void checkListSizeConsistency(const NbnxnPairlistGpu gmx_unused &nbl,
3200 bool gmx_unused haveFreeEnergy)
3202 /* We currently can not check consistency here */
3205 /* Set the buffer flags for newly added entries in the list */
3206 static void setBufferFlags(const NbnxnPairlistCpu &nbl,
3207 const int ncj_old_j,
3208 const int gridj_flag_shift,
3209 gmx_bitmask_t *gridj_flag,
3212 if (gmx::ssize(nbl.cj) > ncj_old_j)
3214 int cbFirst = nbl.cj[ncj_old_j].cj >> gridj_flag_shift;
3215 int cbLast = nbl.cj.back().cj >> gridj_flag_shift;
3216 for (int cb = cbFirst; cb <= cbLast; cb++)
3218 bitmask_init_bit(&gridj_flag[cb], th);
3223 static void setBufferFlags(const NbnxnPairlistGpu gmx_unused &nbl,
3224 int gmx_unused ncj_old_j,
3225 int gmx_unused gridj_flag_shift,
3226 gmx_bitmask_t gmx_unused *gridj_flag,
3229 GMX_ASSERT(false, "This function should never be called");
3232 /* Generates the part of pair-list nbl assigned to our thread */
3233 template <typename T>
3234 static void nbnxn_make_pairlist_part(const nbnxn_search *nbs,
3237 nbnxn_search_work_t *work,
3238 const nbnxn_atomdata_t *nbat,
3239 const t_blocka &exclusions,
3241 const Nbnxm::KernelType kernelType,
3243 gmx_bool bFBufferFlag,
3246 float nsubpair_tot_est,
3253 real rlist2, rl_fep2 = 0;
3255 int ci_b, ci, ci_x, ci_y, ci_xy;
3257 real bx0, bx1, by0, by1, bz0, bz1;
3259 real d2cx, d2z, d2z_cx, d2z_cy, d2zx, d2zxy, d2xy;
3260 int cxf, cxl, cyf, cyf_x, cyl;
3261 int numDistanceChecks;
3262 int gridi_flag_shift = 0, gridj_flag_shift = 0;
3263 gmx_bitmask_t *gridj_flag = nullptr;
3264 int ncj_old_i, ncj_old_j;
3266 nbs_cycle_start(&work->cc[enbsCCsearch]);
3268 if (jGrid.geometry().isSimple != pairlistIsSimple(*nbl) ||
3269 iGrid.geometry().isSimple != pairlistIsSimple(*nbl))
3271 gmx_incons("Grid incompatible with pair-list");
3275 GMX_ASSERT(nbl->na_ci == jGrid.geometry().numAtomsICluster,
3276 "The cluster sizes in the list and grid should match");
3277 nbl->na_cj = Nbnxm::JClusterSizePerKernelType[kernelType];
3278 na_cj_2log = get_2log(nbl->na_cj);
3284 /* Determine conversion of clusters to flag blocks */
3285 gridi_flag_shift = getBufferFlagShift(nbl->na_ci);
3286 gridj_flag_shift = getBufferFlagShift(nbl->na_cj);
3288 gridj_flag = work->buffer_flags.flag;
3291 copy_mat(nbs->box, box);
3293 rlist2 = nbl->rlist*nbl->rlist;
3295 if (nbs->bFEP && !pairlistIsSimple(*nbl))
3297 /* Determine an atom-pair list cut-off distance for FEP atom pairs.
3298 * We should not simply use rlist, since then we would not have
3299 * the small, effective buffering of the NxN lists.
3300 * The buffer is on overestimate, but the resulting cost for pairs
3301 * beyond rlist is neglible compared to the FEP pairs within rlist.
3303 rl_fep2 = nbl->rlist + effective_buffer_1x1_vs_MxN(iGrid, jGrid);
3307 fprintf(debug, "nbl_fep atom-pair rlist %f\n", rl_fep2);
3309 rl_fep2 = rl_fep2*rl_fep2;
3312 const Grid::Dimensions &iGridDims = iGrid.dimensions();
3313 const Grid::Dimensions &jGridDims = jGrid.dimensions();
3315 rbb2 = boundingbox_only_distance2(iGridDims, jGridDims, nbl->rlist, pairlistIsSimple(*nbl));
3319 fprintf(debug, "nbl bounding box only distance %f\n", std::sqrt(rbb2));
3322 const bool isIntraGridList = (&iGrid == &jGrid);
3324 /* Set the shift range */
3325 for (int d = 0; d < DIM; d++)
3327 /* Check if we need periodicity shifts.
3328 * Without PBC or with domain decomposition we don't need them.
3330 if (d >= ePBC2npbcdim(nbs->ePBC) || nbs->dd_dim[d])
3336 const real listRangeCellToCell =
3337 listRangeForGridCellToGridCell(rlist, iGrid.dimensions(), jGrid.dimensions());
3339 box[XX][XX] - fabs(box[YY][XX]) - fabs(box[ZZ][XX]) < listRangeCellToCell)
3349 const bool bSimple = pairlistIsSimple(*nbl);
3350 gmx::ArrayRef<const BoundingBox> bb_i;
3352 gmx::ArrayRef<const float> pbb_i;
3355 bb_i = iGrid.iBoundingBoxes();
3359 pbb_i = iGrid.packedBoundingBoxes();
3362 /* We use the normal bounding box format for both grid types */
3363 bb_i = iGrid.iBoundingBoxes();
3365 gmx::ArrayRef<const BoundingBox1D> bbcz_i = iGrid.zBoundingBoxes();
3366 gmx::ArrayRef<const int> flags_i = iGrid.clusterFlags();
3367 gmx::ArrayRef<const BoundingBox1D> bbcz_j = jGrid.zBoundingBoxes();
3368 int cell0_i = iGrid.cellOffset();
3372 fprintf(debug, "nbl nc_i %d col.av. %.1f ci_block %d\n",
3373 iGrid.numCells(), iGrid.numCells()/static_cast<double>(iGrid.numColumns()), ci_block);
3376 numDistanceChecks = 0;
3378 const real listRangeBBToJCell2 = gmx::square(listRangeForBoundingBoxToGridCell(rlist, jGrid.dimensions()));
3380 /* Initially ci_b and ci to 1 before where we want them to start,
3381 * as they will both be incremented in next_ci.
3384 ci = th*ci_block - 1;
3387 while (next_ci(iGrid, nth, ci_block, &ci_x, &ci_y, &ci_b, &ci))
3389 if (bSimple && flags_i[ci] == 0)
3394 ncj_old_i = getNumSimpleJClustersInList(*nbl);
3397 if (!isIntraGridList && shp[XX] == 0)
3401 bx1 = bb_i[ci].upper.x;
3405 bx1 = iGridDims.lowerCorner[XX] + (ci_x+1)*iGridDims.cellSize[XX];
3407 if (bx1 < jGridDims.lowerCorner[XX])
3409 d2cx = gmx::square(jGridDims.lowerCorner[XX] - bx1);
3411 if (d2cx >= listRangeBBToJCell2)
3418 ci_xy = ci_x*iGridDims.numCells[YY] + ci_y;
3420 /* Loop over shift vectors in three dimensions */
3421 for (int tz = -shp[ZZ]; tz <= shp[ZZ]; tz++)
3423 const real shz = tz*box[ZZ][ZZ];
3425 bz0 = bbcz_i[ci].lower + shz;
3426 bz1 = bbcz_i[ci].upper + shz;
3434 d2z = gmx::square(bz1);
3438 d2z = gmx::square(bz0 - box[ZZ][ZZ]);
3441 d2z_cx = d2z + d2cx;
3443 if (d2z_cx >= rlist2)
3448 bz1_frac = bz1/iGrid.numCellsInColumn(ci_xy);
3453 /* The check with bz1_frac close to or larger than 1 comes later */
3455 for (int ty = -shp[YY]; ty <= shp[YY]; ty++)
3457 const real shy = ty*box[YY][YY] + tz*box[ZZ][YY];
3461 by0 = bb_i[ci].lower.y + shy;
3462 by1 = bb_i[ci].upper.y + shy;
3466 by0 = iGridDims.lowerCorner[YY] + (ci_y )*iGridDims.cellSize[YY] + shy;
3467 by1 = iGridDims.lowerCorner[YY] + (ci_y + 1)*iGridDims.cellSize[YY] + shy;
3470 get_cell_range<YY>(by0, by1,
3481 if (by1 < jGridDims.lowerCorner[YY])
3483 d2z_cy += gmx::square(jGridDims.lowerCorner[YY] - by1);
3485 else if (by0 > jGridDims.upperCorner[YY])
3487 d2z_cy += gmx::square(by0 - jGridDims.upperCorner[YY]);
3490 for (int tx = -shp[XX]; tx <= shp[XX]; tx++)
3492 const int shift = XYZ2IS(tx, ty, tz);
3494 const bool excludeSubDiagonal = (isIntraGridList && shift == CENTRAL);
3496 if (c_pbcShiftBackward && isIntraGridList && shift > CENTRAL)
3501 const real shx = tx*box[XX][XX] + ty*box[YY][XX] + tz*box[ZZ][XX];
3505 bx0 = bb_i[ci].lower.x + shx;
3506 bx1 = bb_i[ci].upper.x + shx;
3510 bx0 = iGridDims.lowerCorner[XX] + (ci_x )*iGridDims.cellSize[XX] + shx;
3511 bx1 = iGridDims.lowerCorner[XX] + (ci_x+1)*iGridDims.cellSize[XX] + shx;
3514 get_cell_range<XX>(bx0, bx1,
3524 addNewIEntry(nbl, cell0_i+ci, shift, flags_i[ci]);
3526 if ((!c_pbcShiftBackward || excludeSubDiagonal) &&
3529 /* Leave the pairs with i > j.
3530 * x is the major index, so skip half of it.
3535 set_icell_bb(iGrid, ci, shx, shy, shz,
3538 icell_set_x(cell0_i+ci, shx, shy, shz,
3539 nbat->xstride, nbat->x().data(),
3543 for (int cx = cxf; cx <= cxl; cx++)
3546 if (jGridDims.lowerCorner[XX] + cx*jGridDims.cellSize[XX] > bx1)
3548 d2zx += gmx::square(jGridDims.lowerCorner[XX] + cx*jGridDims.cellSize[XX] - bx1);
3550 else if (jGridDims.lowerCorner[XX] + (cx+1)*jGridDims.cellSize[XX] < bx0)
3552 d2zx += gmx::square(jGridDims.lowerCorner[XX] + (cx+1)*jGridDims.cellSize[XX] - bx0);
3555 if (isIntraGridList &&
3557 (!c_pbcShiftBackward || shift == CENTRAL) &&
3560 /* Leave the pairs with i > j.
3561 * Skip half of y when i and j have the same x.
3570 for (int cy = cyf_x; cy <= cyl; cy++)
3572 const int columnStart = jGrid.firstCellInColumn(cx*jGridDims.numCells[YY] + cy);
3573 const int columnEnd = jGrid.firstCellInColumn(cx*jGridDims.numCells[YY] + cy + 1);
3576 if (jGridDims.lowerCorner[YY] + cy*jGridDims.cellSize[YY] > by1)
3578 d2zxy += gmx::square(jGridDims.lowerCorner[YY] + cy*jGridDims.cellSize[YY] - by1);
3580 else if (jGridDims.lowerCorner[YY] + (cy + 1)*jGridDims.cellSize[YY] < by0)
3582 d2zxy += gmx::square(jGridDims.lowerCorner[YY] + (cy + 1)*jGridDims.cellSize[YY] - by0);
3584 if (columnStart < columnEnd && d2zxy < listRangeBBToJCell2)
3586 /* To improve efficiency in the common case
3587 * of a homogeneous particle distribution,
3588 * we estimate the index of the middle cell
3589 * in range (midCell). We search down and up
3590 * starting from this index.
3592 * Note that the bbcz_j array contains bounds
3593 * for i-clusters, thus for clusters of 4 atoms.
3594 * For the common case where the j-cluster size
3595 * is 8, we could step with a stride of 2,
3596 * but we do not do this because it would
3597 * complicate this code even more.
3599 int midCell = columnStart + static_cast<int>(bz1_frac*(columnEnd - columnStart));
3600 if (midCell >= columnEnd)
3602 midCell = columnEnd - 1;
3607 /* Find the lowest cell that can possibly
3609 * Check if we hit the bottom of the grid,
3610 * if the j-cell is below the i-cell and if so,
3611 * if it is within range.
3613 int downTestCell = midCell;
3614 while (downTestCell >= columnStart &&
3615 (bbcz_j[downTestCell].upper >= bz0 ||
3616 d2xy + gmx::square(bbcz_j[downTestCell].upper - bz0) < rlist2))
3620 int firstCell = downTestCell + 1;
3622 /* Find the highest cell that can possibly
3624 * Check if we hit the top of the grid,
3625 * if the j-cell is above the i-cell and if so,
3626 * if it is within range.
3628 int upTestCell = midCell + 1;
3629 while (upTestCell < columnEnd &&
3630 (bbcz_j[upTestCell].lower <= bz1 ||
3631 d2xy + gmx::square(bbcz_j[upTestCell].lower - bz1) < rlist2))
3635 int lastCell = upTestCell - 1;
3637 #define NBNXN_REFCODE 0
3640 /* Simple reference code, for debugging,
3641 * overrides the more complex code above.
3643 firstCell = columnEnd;
3645 for (int k = columnStart; k < columnEnd; k++)
3647 if (d2xy + gmx::square(bbcz_j[k*NNBSBB_D + 1] - bz0) < rlist2 &&
3652 if (d2xy + gmx::square(bbcz_j[k*NNBSBB_D] - bz1) < rlist2 &&
3661 if (isIntraGridList)
3663 /* We want each atom/cell pair only once,
3664 * only use cj >= ci.
3666 if (!c_pbcShiftBackward || shift == CENTRAL)
3668 firstCell = std::max(firstCell, ci);
3672 if (firstCell <= lastCell)
3674 GMX_ASSERT(firstCell >= columnStart && lastCell < columnEnd, "The range should reside within the current grid column");
3676 /* For f buffer flags with simple lists */
3677 ncj_old_j = getNumSimpleJClustersInList(*nbl);
3679 makeClusterListWrapper(nbl,
3681 jGrid, firstCell, lastCell,
3686 &numDistanceChecks);
3690 setBufferFlags(*nbl, ncj_old_j, gridj_flag_shift,
3694 incrementNumSimpleJClustersInList(nbl, ncj_old_j);
3700 /* Set the exclusions for this ci list */
3701 setExclusionsForIEntry(nbs,
3705 *getOpenIEntry(nbl),
3710 make_fep_list(nbs, nbat, nbl,
3715 iGrid, jGrid, nbl_fep);
3718 /* Close this ci list */
3721 progBal, nsubpair_tot_est,
3727 if (bFBufferFlag && getNumSimpleJClustersInList(*nbl) > ncj_old_i)
3729 bitmask_init_bit(&(work->buffer_flags.flag[(iGrid.cellOffset() + ci) >> gridi_flag_shift]), th);
3733 work->ndistc = numDistanceChecks;
3735 nbs_cycle_stop(&work->cc[enbsCCsearch]);
3737 checkListSizeConsistency(*nbl, nbs->bFEP);
3741 fprintf(debug, "number of distance checks %d\n", numDistanceChecks);
3743 print_nblist_statistics(debug, nbl, nbs, rlist);
3747 fprintf(debug, "nbl FEP list pairs: %d\n", nbl_fep->nrj);
3752 static void reduce_buffer_flags(const nbnxn_search *nbs,
3754 const nbnxn_buffer_flags_t *dest)
3756 for (int s = 0; s < nsrc; s++)
3758 gmx_bitmask_t * flag = nbs->work[s].buffer_flags.flag;
3760 for (int b = 0; b < dest->nflag; b++)
3762 bitmask_union(&(dest->flag[b]), flag[b]);
3767 static void print_reduction_cost(const nbnxn_buffer_flags_t *flags, int nout)
3769 int nelem, nkeep, ncopy, nred, out;
3770 gmx_bitmask_t mask_0;
3776 bitmask_init_bit(&mask_0, 0);
3777 for (int b = 0; b < flags->nflag; b++)
3779 if (bitmask_is_equal(flags->flag[b], mask_0))
3781 /* Only flag 0 is set, no copy of reduction required */
3785 else if (!bitmask_is_zero(flags->flag[b]))
3788 for (out = 0; out < nout; out++)
3790 if (bitmask_is_set(flags->flag[b], out))
3807 fprintf(debug, "nbnxn reduction: #flag %d #list %d elem %4.2f, keep %4.2f copy %4.2f red %4.2f\n",
3809 nelem/static_cast<double>(flags->nflag),
3810 nkeep/static_cast<double>(flags->nflag),
3811 ncopy/static_cast<double>(flags->nflag),
3812 nred/static_cast<double>(flags->nflag));
3815 /* Copies the list entries from src to dest when cjStart <= *cjGlobal < cjEnd.
3816 * *cjGlobal is updated with the cj count in src.
3817 * When setFlags==true, flag bit t is set in flag for all i and j clusters.
3819 template<bool setFlags>
3820 static void copySelectedListRange(const nbnxn_ci_t * gmx_restrict srcCi,
3821 const NbnxnPairlistCpu * gmx_restrict src,
3822 NbnxnPairlistCpu * gmx_restrict dest,
3823 gmx_bitmask_t *flag,
3824 int iFlagShift, int jFlagShift, int t)
3826 const int ncj = srcCi->cj_ind_end - srcCi->cj_ind_start;
3828 dest->ci.push_back(*srcCi);
3829 dest->ci.back().cj_ind_start = dest->cj.size();
3830 dest->ci.back().cj_ind_end = dest->cj.size() + ncj;
3834 bitmask_init_bit(&flag[srcCi->ci >> iFlagShift], t);
3837 for (int j = srcCi->cj_ind_start; j < srcCi->cj_ind_end; j++)
3839 dest->cj.push_back(src->cj[j]);
3843 /* NOTE: This is relatively expensive, since this
3844 * operation is done for all elements in the list,
3845 * whereas at list generation this is done only
3846 * once for each flag entry.
3848 bitmask_init_bit(&flag[src->cj[j].cj >> jFlagShift], t);
3853 /* This routine re-balances the pairlists such that all are nearly equally
3854 * sized. Only whole i-entries are moved between lists. These are moved
3855 * between the ends of the lists, such that the buffer reduction cost should
3856 * not change significantly.
3857 * Note that all original reduction flags are currently kept. This can lead
3858 * to reduction of parts of the force buffer that could be avoided. But since
3859 * the original lists are quite balanced, this will only give minor overhead.
3861 static void rebalanceSimpleLists(int numLists,
3862 NbnxnPairlistCpu * const * const srcSet,
3863 NbnxnPairlistCpu **destSet,
3864 gmx::ArrayRef<nbnxn_search_work_t> searchWork)
3867 for (int s = 0; s < numLists; s++)
3869 ncjTotal += srcSet[s]->ncjInUse;
3871 int ncjTarget = (ncjTotal + numLists - 1)/numLists;
3873 #pragma omp parallel num_threads(numLists)
3875 int t = gmx_omp_get_thread_num();
3877 int cjStart = ncjTarget* t;
3878 int cjEnd = ncjTarget*(t + 1);
3880 /* The destination pair-list for task/thread t */
3881 NbnxnPairlistCpu *dest = destSet[t];
3883 clear_pairlist(dest);
3884 dest->na_cj = srcSet[0]->na_cj;
3886 /* Note that the flags in the work struct (still) contain flags
3887 * for all entries that are present in srcSet->nbl[t].
3889 gmx_bitmask_t *flag = searchWork[t].buffer_flags.flag;
3891 int iFlagShift = getBufferFlagShift(dest->na_ci);
3892 int jFlagShift = getBufferFlagShift(dest->na_cj);
3895 for (int s = 0; s < numLists && cjGlobal < cjEnd; s++)
3897 const NbnxnPairlistCpu *src = srcSet[s];
3899 if (cjGlobal + src->ncjInUse > cjStart)
3901 for (gmx::index i = 0; i < gmx::ssize(src->ci) && cjGlobal < cjEnd; i++)
3903 const nbnxn_ci_t *srcCi = &src->ci[i];
3904 int ncj = srcCi->cj_ind_end - srcCi->cj_ind_start;
3905 if (cjGlobal >= cjStart)
3907 /* If the source list is not our own, we need to set
3908 * extra flags (the template bool parameter).
3912 copySelectedListRange
3915 flag, iFlagShift, jFlagShift, t);
3919 copySelectedListRange
3922 dest, flag, iFlagShift, jFlagShift, t);
3930 cjGlobal += src->ncjInUse;
3934 dest->ncjInUse = dest->cj.size();
3938 int ncjTotalNew = 0;
3939 for (int s = 0; s < numLists; s++)
3941 ncjTotalNew += destSet[s]->ncjInUse;
3943 GMX_RELEASE_ASSERT(ncjTotalNew == ncjTotal, "The total size of the lists before and after rebalancing should match");
3947 /* Returns if the pairlists are so imbalanced that it is worth rebalancing. */
3948 static bool checkRebalanceSimpleLists(const nbnxn_pairlist_set_t *listSet)
3950 int numLists = listSet->nnbl;
3953 for (int s = 0; s < numLists; s++)
3955 ncjMax = std::max(ncjMax, listSet->nbl[s]->ncjInUse);
3956 ncjTotal += listSet->nbl[s]->ncjInUse;
3960 fprintf(debug, "Pair-list ncjMax %d ncjTotal %d\n", ncjMax, ncjTotal);
3962 /* The rebalancing adds 3% extra time to the search. Heuristically we
3963 * determined that under common conditions the non-bonded kernel balance
3964 * improvement will outweigh this when the imbalance is more than 3%.
3965 * But this will, obviously, depend on search vs kernel time and nstlist.
3967 const real rebalanceTolerance = 1.03;
3969 return numLists*ncjMax > ncjTotal*rebalanceTolerance;
3972 /* Perform a count (linear) sort to sort the smaller lists to the end.
3973 * This avoids load imbalance on the GPU, as large lists will be
3974 * scheduled and executed first and the smaller lists later.
3975 * Load balancing between multi-processors only happens at the end
3976 * and there smaller lists lead to more effective load balancing.
3977 * The sorting is done on the cj4 count, not on the actual pair counts.
3978 * Not only does this make the sort faster, but it also results in
3979 * better load balancing than using a list sorted on exact load.
3980 * This function swaps the pointer in the pair list to avoid a copy operation.
3982 static void sort_sci(NbnxnPairlistGpu *nbl)
3984 if (nbl->cj4.size() <= nbl->sci.size())
3986 /* nsci = 0 or all sci have size 1, sorting won't change the order */
3990 NbnxnPairlistGpuWork &work = *nbl->work;
3992 /* We will distinguish differences up to double the average */
3993 const int m = (2*nbl->cj4.size())/nbl->sci.size();
3995 /* Resize work.sci_sort so we can sort into it */
3996 work.sci_sort.resize(nbl->sci.size());
3998 std::vector<int> &sort = work.sortBuffer;
3999 /* Set up m + 1 entries in sort, initialized at 0 */
4001 sort.resize(m + 1, 0);
4002 /* Count the entries of each size */
4003 for (const nbnxn_sci_t &sci : nbl->sci)
4005 int i = std::min(m, sci.numJClusterGroups());
4008 /* Calculate the offset for each count */
4011 for (int i = m - 1; i >= 0; i--)
4014 sort[i] = sort[i + 1] + s0;
4018 /* Sort entries directly into place */
4019 gmx::ArrayRef<nbnxn_sci_t> sci_sort = work.sci_sort;
4020 for (const nbnxn_sci_t &sci : nbl->sci)
4022 int i = std::min(m, sci.numJClusterGroups());
4023 sci_sort[sort[i]++] = sci;
4026 /* Swap the sci pointers so we use the new, sorted list */
4027 std::swap(nbl->sci, work.sci_sort);
4031 nonbonded_verlet_t::PairlistSets::construct(const InteractionLocality iLocality,
4033 nbnxn_atomdata_t *nbat,
4034 const t_blocka *excl,
4035 const Nbnxm::KernelType kernelType,
4039 nbnxn_pairlist_set_t *nbl_list = &pairlistSet(iLocality);
4041 const real rlist = nbl_list->params.rlistOuter;
4043 int nsubpair_target;
4044 float nsubpair_tot_est;
4047 gmx_bool CombineNBLists;
4049 int np_tot, np_noq, np_hlj, nap;
4051 nnbl = nbl_list->nnbl;
4052 CombineNBLists = nbl_list->bCombined;
4056 fprintf(debug, "ns making %d nblists\n", nnbl);
4059 nbat->bUseBufferFlags = (nbat->out.size() > 1);
4060 /* We should re-init the flags before making the first list */
4061 if (nbat->bUseBufferFlags && iLocality == InteractionLocality::Local)
4063 init_buffer_flags(&nbat->buffer_flags, nbat->numAtoms());
4067 if (iLocality == InteractionLocality::Local)
4069 /* Only zone (grid) 0 vs 0 */
4074 nzi = nbs->zones->nizone;
4077 if (!nbl_list->bSimple && minimumIlistCountForGpuBalancing_ > 0)
4079 get_nsubpair_target(nbs, iLocality, rlist, minimumIlistCountForGpuBalancing_,
4080 &nsubpair_target, &nsubpair_tot_est);
4084 nsubpair_target = 0;
4085 nsubpair_tot_est = 0;
4088 /* Clear all pair-lists */
4089 for (int th = 0; th < nnbl; th++)
4091 if (nbl_list->bSimple)
4093 clear_pairlist(nbl_list->nbl[th]);
4097 clear_pairlist(nbl_list->nblGpu[th]);
4102 clear_pairlist_fep(nbl_list->nbl_fep[th]);
4106 for (int zi = 0; zi < nzi; zi++)
4108 const Grid &iGrid = nbs->grid[zi];
4112 if (iLocality == InteractionLocality::Local)
4119 zj0 = nbs->zones->izone[zi].j0;
4120 zj1 = nbs->zones->izone[zi].j1;
4126 for (int zj = zj0; zj < zj1; zj++)
4128 const Grid &jGrid = nbs->grid[zj];
4132 fprintf(debug, "ns search grid %d vs %d\n", zi, zj);
4135 nbs_cycle_start(&nbs->cc[enbsCCsearch]);
4137 ci_block = get_ci_block_size(iGrid, nbs->DomDec, nnbl);
4139 /* With GPU: generate progressively smaller lists for
4140 * load balancing for local only or non-local with 2 zones.
4142 progBal = (iLocality == InteractionLocality::Local || nbs->zones->n <= 2);
4144 #pragma omp parallel for num_threads(nnbl) schedule(static)
4145 for (int th = 0; th < nnbl; th++)
4149 /* Re-init the thread-local work flag data before making
4150 * the first list (not an elegant conditional).
4152 if (nbat->bUseBufferFlags && ((zi == 0 && zj == 0)))
4154 init_buffer_flags(&nbs->work[th].buffer_flags, nbat->numAtoms());
4157 if (CombineNBLists && th > 0)
4159 GMX_ASSERT(!nbl_list->bSimple, "Can only combine GPU lists");
4161 clear_pairlist(nbl_list->nblGpu[th]);
4164 /* Divide the i super cell equally over the nblists */
4165 if (nbl_list->bSimple)
4167 nbnxn_make_pairlist_part(nbs, iGrid, jGrid,
4168 &nbs->work[th], nbat, *excl,
4172 nbat->bUseBufferFlags,
4174 progBal, nsubpair_tot_est,
4177 nbl_list->nbl_fep[th]);
4181 nbnxn_make_pairlist_part(nbs, iGrid, jGrid,
4182 &nbs->work[th], nbat, *excl,
4186 nbat->bUseBufferFlags,
4188 progBal, nsubpair_tot_est,
4190 nbl_list->nblGpu[th],
4191 nbl_list->nbl_fep[th]);
4194 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
4196 nbs_cycle_stop(&nbs->cc[enbsCCsearch]);
4201 for (int th = 0; th < nnbl; th++)
4203 inc_nrnb(nrnb, eNR_NBNXN_DIST2, nbs->work[th].ndistc);
4205 if (nbl_list->bSimple)
4207 NbnxnPairlistCpu *nbl = nbl_list->nbl[th];
4208 np_tot += nbl->cj.size();
4209 np_noq += nbl->work->ncj_noq;
4210 np_hlj += nbl->work->ncj_hlj;
4214 NbnxnPairlistGpu *nbl = nbl_list->nblGpu[th];
4215 /* This count ignores potential subsequent pair pruning */
4216 np_tot += nbl->nci_tot;
4219 if (nbl_list->bSimple)
4221 nap = nbl_list->nbl[0]->na_ci*nbl_list->nbl[0]->na_cj;
4225 nap = gmx::square(nbl_list->nblGpu[0]->na_ci);
4227 nbl_list->natpair_ljq = (np_tot - np_noq)*nap - np_hlj*nap/2;
4228 nbl_list->natpair_lj = np_noq*nap;
4229 nbl_list->natpair_q = np_hlj*nap/2;
4231 if (CombineNBLists && nnbl > 1)
4233 GMX_ASSERT(!nbl_list->bSimple, "Can only combine GPU lists");
4234 NbnxnPairlistGpu **nbl = nbl_list->nblGpu;
4236 nbs_cycle_start(&nbs->cc[enbsCCcombine]);
4238 combine_nblists(nnbl-1, nbl+1, nbl[0]);
4240 nbs_cycle_stop(&nbs->cc[enbsCCcombine]);
4245 if (nbl_list->bSimple)
4247 if (nnbl > 1 && checkRebalanceSimpleLists(nbl_list))
4249 rebalanceSimpleLists(nbl_list->nnbl, nbl_list->nbl, nbl_list->nbl_work, nbs->work);
4251 /* Swap the pointer of the sets of pair lists */
4252 NbnxnPairlistCpu **tmp = nbl_list->nbl;
4253 nbl_list->nbl = nbl_list->nbl_work;
4254 nbl_list->nbl_work = tmp;
4259 /* Sort the entries on size, large ones first */
4260 if (CombineNBLists || nnbl == 1)
4262 sort_sci(nbl_list->nblGpu[0]);
4266 #pragma omp parallel for num_threads(nnbl) schedule(static)
4267 for (int th = 0; th < nnbl; th++)
4271 sort_sci(nbl_list->nblGpu[th]);
4273 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
4278 if (nbat->bUseBufferFlags)
4280 reduce_buffer_flags(nbs, nbl_list->nnbl, &nbat->buffer_flags);
4285 /* Balance the free-energy lists over all the threads */
4286 balance_fep_lists(nbs, nbl_list);
4289 if (nbl_list->bSimple)
4291 /* This is a fresh list, so not pruned, stored using ci.
4292 * ciOuter is invalid at this point.
4294 GMX_ASSERT(nbl_list->nbl[0]->ciOuter.empty(), "ciOuter is invalid so it should be empty");
4297 if (iLocality == Nbnxm::InteractionLocality::Local)
4299 outerListCreationStep_ = step;
4303 GMX_RELEASE_ASSERT(outerListCreationStep_ == step,
4304 "Outer list should be created at the same step as the inner list");
4307 /* Special performance logging stuff (env.var. GMX_NBNXN_CYCLE) */
4308 if (iLocality == InteractionLocality::Local)
4310 nbs->search_count++;
4312 if (nbs->print_cycles &&
4313 (!nbs->DomDec || iLocality == InteractionLocality::NonLocal) &&
4314 nbs->search_count % 100 == 0)
4316 nbs_cycle_print(stderr, nbs);
4319 /* If we have more than one list, they either got rebalancing (CPU)
4320 * or combined (GPU), so we should dump the final result to debug.
4322 if (debug && nbl_list->nnbl > 1)
4324 if (nbl_list->bSimple)
4326 for (int t = 0; t < nbl_list->nnbl; t++)
4328 print_nblist_statistics(debug, nbl_list->nbl[t], nbs, rlist);
4333 print_nblist_statistics(debug, nbl_list->nblGpu[0], nbs, rlist);
4341 if (nbl_list->bSimple)
4343 for (int t = 0; t < nbl_list->nnbl; t++)
4345 print_nblist_ci_cj(debug, nbl_list->nbl[t]);
4350 print_nblist_sci_cj(debug, nbl_list->nblGpu[0]);
4354 if (nbat->bUseBufferFlags)
4356 print_reduction_cost(&nbat->buffer_flags, nbl_list->nnbl);
4360 if (params_.useDynamicPruning && nbl_list->bSimple)
4362 nbnxnPrepareListForDynamicPruning(nbl_list);
4367 nonbonded_verlet_t::constructPairlist(const Nbnxm::InteractionLocality iLocality,
4368 const t_blocka *excl,
4372 pairlistSets_->construct(iLocality, nbs.get(), nbat.get(), excl,
4373 kernelSetup_.kernelType,
4378 /* Launch the transfer of the pairlist to the GPU.
4380 * NOTE: The launch overhead is currently not timed separately
4382 Nbnxm::gpu_init_pairlist(gpu_nbv,
4383 pairlistSets().pairlistSet(iLocality).nblGpu[0],
4388 void nbnxnPrepareListForDynamicPruning(nbnxn_pairlist_set_t *listSet)
4390 GMX_RELEASE_ASSERT(listSet->bSimple, "Should only be called for simple lists");
4392 /* TODO: Restructure the lists so we have actual outer and inner
4393 * list objects so we can set a single pointer instead of
4394 * swapping several pointers.
4397 for (int i = 0; i < listSet->nnbl; i++)
4399 NbnxnPairlistCpu &list = *listSet->nbl[i];
4401 /* The search produced a list in ci/cj.
4402 * Swap the list pointers so we get the outer list is ciOuter,cjOuter
4403 * and we can prune that to get an inner list in ci/cj.
4405 GMX_RELEASE_ASSERT(list.ciOuter.empty() && list.cjOuter.empty(),
4406 "The outer lists should be empty before preparation");
4408 std::swap(list.ci, list.ciOuter);
4409 std::swap(list.cj, list.cjOuter);