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 nbnxn_bb_t = 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 nbnxn_bb_t *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 /* Plain C code calculating the distance^2 between two bounding boxes */
447 static float subc_bb_dist2(int si,
448 const nbnxn_bb_t *bb_i_ci,
450 gmx::ArrayRef<const nbnxn_bb_t> bb_j_all)
452 const nbnxn_bb_t &bb_i = bb_i_ci[si];
453 const nbnxn_bb_t &bb_j = bb_j_all[csj];
456 float dl, dh, dm, dm0;
458 dl = bb_i.lower.x - bb_j.upper.x;
459 dh = bb_j.lower.x - bb_i.upper.x;
460 dm = std::max(dl, dh);
461 dm0 = std::max(dm, 0.0f);
464 dl = bb_i.lower.y - bb_j.upper.y;
465 dh = bb_j.lower.y - bb_i.upper.y;
466 dm = std::max(dl, dh);
467 dm0 = std::max(dm, 0.0f);
470 dl = bb_i.lower.z - bb_j.upper.z;
471 dh = bb_j.lower.z - bb_i.upper.z;
472 dm = std::max(dl, dh);
473 dm0 = std::max(dm, 0.0f);
479 #if NBNXN_SEARCH_BB_SIMD4
481 /* 4-wide SIMD code for bb distance for bb format xyz0 */
482 static float subc_bb_dist2_simd4(int si,
483 const nbnxn_bb_t *bb_i_ci,
485 gmx::ArrayRef<const nbnxn_bb_t> bb_j_all)
487 // TODO: During SIMDv2 transition only some archs use namespace (remove when done)
490 Simd4Float bb_i_S0, bb_i_S1;
491 Simd4Float bb_j_S0, bb_j_S1;
497 bb_i_S0 = load4(bb_i_ci[si].lower.ptr());
498 bb_i_S1 = load4(bb_i_ci[si].upper.ptr());
499 bb_j_S0 = load4(bb_j_all[csj].lower.ptr());
500 bb_j_S1 = load4(bb_j_all[csj].upper.ptr());
502 dl_S = bb_i_S0 - bb_j_S1;
503 dh_S = bb_j_S0 - bb_i_S1;
505 dm_S = max(dl_S, dh_S);
506 dm0_S = max(dm_S, simd4SetZeroF());
508 return dotProduct(dm0_S, dm0_S);
511 /* Calculate bb bounding distances of bb_i[si,...,si+3] and store them in d2 */
512 #define SUBC_BB_DIST2_SIMD4_XXXX_INNER(si, bb_i, d2) \
516 Simd4Float dx_0, dy_0, dz_0; \
517 Simd4Float dx_1, dy_1, dz_1; \
519 Simd4Float mx, my, mz; \
520 Simd4Float m0x, m0y, m0z; \
522 Simd4Float d2x, d2y, d2z; \
523 Simd4Float d2s, d2t; \
525 shi = (si)*Nbnxm::c_numBoundingBoxBounds1D*DIM; \
527 xi_l = load4((bb_i)+shi+0*STRIDE_PBB); \
528 yi_l = load4((bb_i)+shi+1*STRIDE_PBB); \
529 zi_l = load4((bb_i)+shi+2*STRIDE_PBB); \
530 xi_h = load4((bb_i)+shi+3*STRIDE_PBB); \
531 yi_h = load4((bb_i)+shi+4*STRIDE_PBB); \
532 zi_h = load4((bb_i)+shi+5*STRIDE_PBB); \
534 dx_0 = xi_l - xj_h; \
535 dy_0 = yi_l - yj_h; \
536 dz_0 = zi_l - zj_h; \
538 dx_1 = xj_l - xi_h; \
539 dy_1 = yj_l - yi_h; \
540 dz_1 = zj_l - zi_h; \
542 mx = max(dx_0, dx_1); \
543 my = max(dy_0, dy_1); \
544 mz = max(dz_0, dz_1); \
546 m0x = max(mx, zero); \
547 m0y = max(my, zero); \
548 m0z = max(mz, zero); \
557 store4((d2)+(si), d2t); \
560 /* 4-wide SIMD code for nsi bb distances for bb format xxxxyyyyzzzz */
561 static void subc_bb_dist2_simd4_xxxx(const float *bb_j,
562 int nsi, const float *bb_i,
565 // TODO: During SIMDv2 transition only some archs use namespace (remove when done)
568 Simd4Float xj_l, yj_l, zj_l;
569 Simd4Float xj_h, yj_h, zj_h;
570 Simd4Float xi_l, yi_l, zi_l;
571 Simd4Float xi_h, yi_h, zi_h;
577 xj_l = Simd4Float(bb_j[0*STRIDE_PBB]);
578 yj_l = Simd4Float(bb_j[1*STRIDE_PBB]);
579 zj_l = Simd4Float(bb_j[2*STRIDE_PBB]);
580 xj_h = Simd4Float(bb_j[3*STRIDE_PBB]);
581 yj_h = Simd4Float(bb_j[4*STRIDE_PBB]);
582 zj_h = Simd4Float(bb_j[5*STRIDE_PBB]);
584 /* Here we "loop" over si (0,STRIDE_PBB) from 0 to nsi with step STRIDE_PBB.
585 * But as we know the number of iterations is 1 or 2, we unroll manually.
587 SUBC_BB_DIST2_SIMD4_XXXX_INNER(0, bb_i, d2);
588 if (STRIDE_PBB < nsi)
590 SUBC_BB_DIST2_SIMD4_XXXX_INNER(STRIDE_PBB, bb_i, d2);
594 #endif /* NBNXN_SEARCH_BB_SIMD4 */
597 /* Returns if any atom pair from two clusters is within distance sqrt(rlist2) */
598 static inline gmx_bool
599 clusterpair_in_range(const NbnxnPairlistGpuWork &work,
601 int csj, int stride, const real *x_j,
604 #if !GMX_SIMD4_HAVE_REAL
607 * All coordinates are stored as xyzxyz...
610 const real *x_i = work.iSuperClusterData.x.data();
612 for (int i = 0; i < c_nbnxnGpuClusterSize; i++)
614 int i0 = (si*c_nbnxnGpuClusterSize + i)*DIM;
615 for (int j = 0; j < c_nbnxnGpuClusterSize; j++)
617 int j0 = (csj*c_nbnxnGpuClusterSize + j)*stride;
619 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]);
630 #else /* !GMX_SIMD4_HAVE_REAL */
632 /* 4-wide SIMD version.
633 * The coordinates x_i are stored as xxxxyyyy..., x_j is stored xyzxyz...
634 * Using 8-wide AVX(2) is not faster on Intel Sandy Bridge and Haswell.
636 static_assert(c_nbnxnGpuClusterSize == 8 || c_nbnxnGpuClusterSize == 4,
637 "A cluster is hard-coded to 4/8 atoms.");
639 Simd4Real rc2_S = Simd4Real(rlist2);
641 const real *x_i = work.iSuperClusterData.xSimd.data();
643 int dim_stride = c_nbnxnGpuClusterSize*DIM;
644 Simd4Real ix_S0 = load4(x_i + si*dim_stride + 0*GMX_SIMD4_WIDTH);
645 Simd4Real iy_S0 = load4(x_i + si*dim_stride + 1*GMX_SIMD4_WIDTH);
646 Simd4Real iz_S0 = load4(x_i + si*dim_stride + 2*GMX_SIMD4_WIDTH);
648 Simd4Real ix_S1, iy_S1, iz_S1;
649 if (c_nbnxnGpuClusterSize == 8)
651 ix_S1 = load4(x_i + si*dim_stride + 3*GMX_SIMD4_WIDTH);
652 iy_S1 = load4(x_i + si*dim_stride + 4*GMX_SIMD4_WIDTH);
653 iz_S1 = load4(x_i + si*dim_stride + 5*GMX_SIMD4_WIDTH);
655 /* We loop from the outer to the inner particles to maximize
656 * the chance that we find a pair in range quickly and return.
658 int j0 = csj*c_nbnxnGpuClusterSize;
659 int j1 = j0 + c_nbnxnGpuClusterSize - 1;
662 Simd4Real jx0_S, jy0_S, jz0_S;
663 Simd4Real jx1_S, jy1_S, jz1_S;
665 Simd4Real dx_S0, dy_S0, dz_S0;
666 Simd4Real dx_S1, dy_S1, dz_S1;
667 Simd4Real dx_S2, dy_S2, dz_S2;
668 Simd4Real dx_S3, dy_S3, dz_S3;
679 Simd4Bool wco_any_S01, wco_any_S23, wco_any_S;
681 jx0_S = Simd4Real(x_j[j0*stride+0]);
682 jy0_S = Simd4Real(x_j[j0*stride+1]);
683 jz0_S = Simd4Real(x_j[j0*stride+2]);
685 jx1_S = Simd4Real(x_j[j1*stride+0]);
686 jy1_S = Simd4Real(x_j[j1*stride+1]);
687 jz1_S = Simd4Real(x_j[j1*stride+2]);
689 /* Calculate distance */
690 dx_S0 = ix_S0 - jx0_S;
691 dy_S0 = iy_S0 - jy0_S;
692 dz_S0 = iz_S0 - jz0_S;
693 dx_S2 = ix_S0 - jx1_S;
694 dy_S2 = iy_S0 - jy1_S;
695 dz_S2 = iz_S0 - jz1_S;
696 if (c_nbnxnGpuClusterSize == 8)
698 dx_S1 = ix_S1 - jx0_S;
699 dy_S1 = iy_S1 - jy0_S;
700 dz_S1 = iz_S1 - jz0_S;
701 dx_S3 = ix_S1 - jx1_S;
702 dy_S3 = iy_S1 - jy1_S;
703 dz_S3 = iz_S1 - jz1_S;
706 /* rsq = dx*dx+dy*dy+dz*dz */
707 rsq_S0 = norm2(dx_S0, dy_S0, dz_S0);
708 rsq_S2 = norm2(dx_S2, dy_S2, dz_S2);
709 if (c_nbnxnGpuClusterSize == 8)
711 rsq_S1 = norm2(dx_S1, dy_S1, dz_S1);
712 rsq_S3 = norm2(dx_S3, dy_S3, dz_S3);
715 wco_S0 = (rsq_S0 < rc2_S);
716 wco_S2 = (rsq_S2 < rc2_S);
717 if (c_nbnxnGpuClusterSize == 8)
719 wco_S1 = (rsq_S1 < rc2_S);
720 wco_S3 = (rsq_S3 < rc2_S);
722 if (c_nbnxnGpuClusterSize == 8)
724 wco_any_S01 = wco_S0 || wco_S1;
725 wco_any_S23 = wco_S2 || wco_S3;
726 wco_any_S = wco_any_S01 || wco_any_S23;
730 wco_any_S = wco_S0 || wco_S2;
733 if (anyTrue(wco_any_S))
744 #endif /* !GMX_SIMD4_HAVE_REAL */
747 /* Returns the j-cluster index for index cjIndex in a cj list */
748 static inline int nblCj(gmx::ArrayRef<const nbnxn_cj_t> cjList,
751 return cjList[cjIndex].cj;
754 /* Returns the j-cluster index for index cjIndex in a cj4 list */
755 static inline int nblCj(gmx::ArrayRef<const nbnxn_cj4_t> cj4List,
758 return cj4List[cjIndex/c_nbnxnGpuJgroupSize].cj[cjIndex & (c_nbnxnGpuJgroupSize - 1)];
761 /* Returns the i-interaction mask of the j sub-cell for index cj_ind */
762 static unsigned int nbl_imask0(const NbnxnPairlistGpu *nbl, int cj_ind)
764 return nbl->cj4[cj_ind/c_nbnxnGpuJgroupSize].imei[0].imask;
767 /* Initializes a single NbnxnPairlistCpu data structure */
768 static void nbnxn_init_pairlist(NbnxnPairlistCpu *nbl)
770 nbl->na_ci = c_nbnxnCpuIClusterSize;
773 nbl->ciOuter.clear();
776 nbl->cjOuter.clear();
779 nbl->work = new NbnxnPairlistCpuWork();
782 NbnxnPairlistGpu::NbnxnPairlistGpu(gmx::PinningPolicy pinningPolicy) :
783 na_ci(c_nbnxnGpuClusterSize),
784 na_cj(c_nbnxnGpuClusterSize),
785 na_sc(c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize),
787 sci({}, {pinningPolicy}),
788 cj4({}, {pinningPolicy}),
789 excl({}, {pinningPolicy}),
792 static_assert(c_nbnxnGpuNumClusterPerSupercluster == c_gpuNumClusterPerCell,
793 "The search code assumes that the a super-cluster matches a search grid cell");
795 static_assert(sizeof(cj4[0].imei[0].imask)*8 >= c_nbnxnGpuJgroupSize*c_gpuNumClusterPerCell,
796 "The i super-cluster cluster interaction mask does not contain a sufficient number of bits");
798 static_assert(sizeof(excl[0])*8 >= c_nbnxnGpuJgroupSize*c_gpuNumClusterPerCell, "The GPU exclusion mask does not contain a sufficient number of bits");
800 // We always want a first entry without any exclusions
803 work = new NbnxnPairlistGpuWork();
806 void nbnxn_init_pairlist_set(nbnxn_pairlist_set_t *nbl_list)
809 (nbl_list->params.pairlistType == PairlistType::Simple4x2 ||
810 nbl_list->params.pairlistType == PairlistType::Simple4x4 ||
811 nbl_list->params.pairlistType == PairlistType::Simple4x8);
812 // Currently GPU lists are always combined
813 nbl_list->bCombined = !nbl_list->bSimple;
815 nbl_list->nnbl = gmx_omp_nthreads_get(emntNonbonded);
817 if (!nbl_list->bCombined &&
818 nbl_list->nnbl > NBNXN_BUFFERFLAG_MAX_THREADS)
820 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.",
821 nbl_list->nnbl, NBNXN_BUFFERFLAG_MAX_THREADS, NBNXN_BUFFERFLAG_MAX_THREADS);
824 if (nbl_list->bSimple)
826 snew(nbl_list->nbl, nbl_list->nnbl);
827 if (nbl_list->nnbl > 1)
829 snew(nbl_list->nbl_work, nbl_list->nnbl);
834 snew(nbl_list->nblGpu, nbl_list->nnbl);
836 nbl_list->nbl_fep.resize(nbl_list->nnbl);
837 /* Execute in order to avoid memory interleaving between threads */
838 #pragma omp parallel for num_threads(nbl_list->nnbl) schedule(static)
839 for (int i = 0; i < nbl_list->nnbl; i++)
843 /* Allocate the nblist data structure locally on each thread
844 * to optimize memory access for NUMA architectures.
846 if (nbl_list->bSimple)
848 nbl_list->nbl[i] = new NbnxnPairlistCpu();
850 nbnxn_init_pairlist(nbl_list->nbl[i]);
851 if (nbl_list->nnbl > 1)
853 nbl_list->nbl_work[i] = new NbnxnPairlistCpu();
854 nbnxn_init_pairlist(nbl_list->nbl_work[i]);
859 /* Only list 0 is used on the GPU, use normal allocation for i>0 */
860 auto pinningPolicy = (i == 0 ? gmx::PinningPolicy::PinnedIfSupported : gmx::PinningPolicy::CannotBePinned);
862 nbl_list->nblGpu[i] = new NbnxnPairlistGpu(pinningPolicy);
865 snew(nbl_list->nbl_fep[i], 1);
866 nbnxn_init_pairlist_fep(nbl_list->nbl_fep[i]);
868 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
872 /* Print statistics of a pair list, used for debug output */
873 static void print_nblist_statistics(FILE *fp, const NbnxnPairlistCpu *nbl,
874 const nbnxn_search *nbs, real rl)
876 const Grid &grid = nbs->grid[0];
877 const Grid::Dimensions &dims = grid.dimensions();
879 fprintf(fp, "nbl nci %zu ncj %d\n",
880 nbl->ci.size(), nbl->ncjInUse);
881 const int numAtomsJCluster = grid.geometry().numAtomsJCluster;
882 const double numAtomsPerCell = nbl->ncjInUse/static_cast<double>(grid.numCells())*numAtomsJCluster;
883 fprintf(fp, "nbl na_cj %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
884 nbl->na_cj, rl, nbl->ncjInUse, nbl->ncjInUse/static_cast<double>(grid.numCells()),
886 numAtomsPerCell/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid.numCells()*numAtomsJCluster/(dims.gridSize[XX]*dims.gridSize[YY]*dims.gridSize[ZZ])));
888 fprintf(fp, "nbl average j cell list length %.1f\n",
889 0.25*nbl->ncjInUse/std::max(static_cast<double>(nbl->ci.size()), 1.0));
891 int cs[SHIFTS] = { 0 };
893 for (const nbnxn_ci_t &ciEntry : nbl->ci)
895 cs[ciEntry.shift & NBNXN_CI_SHIFT] +=
896 ciEntry.cj_ind_end - ciEntry.cj_ind_start;
898 int j = ciEntry.cj_ind_start;
899 while (j < ciEntry.cj_ind_end &&
900 nbl->cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
906 fprintf(fp, "nbl cell pairs, total: %zu excl: %d %.1f%%\n",
907 nbl->cj.size(), npexcl, 100*npexcl/std::max(static_cast<double>(nbl->cj.size()), 1.0));
908 for (int s = 0; s < SHIFTS; s++)
912 fprintf(fp, "nbl shift %2d ncj %3d\n", s, cs[s]);
917 /* Print statistics of a pair lists, used for debug output */
918 static void print_nblist_statistics(FILE *fp, const NbnxnPairlistGpu *nbl,
919 const nbnxn_search *nbs, real rl)
921 const Grid &grid = nbs->grid[0];
922 const Grid::Dimensions &dims = grid.dimensions();
924 fprintf(fp, "nbl nsci %zu ncj4 %zu nsi %d excl4 %zu\n",
925 nbl->sci.size(), nbl->cj4.size(), nbl->nci_tot, nbl->excl.size());
926 const int numAtomsCluster = grid.geometry().numAtomsICluster;
927 const double numAtomsPerCell = nbl->nci_tot/static_cast<double>(grid.numClusters())*numAtomsCluster;
928 fprintf(fp, "nbl na_c %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
929 nbl->na_ci, rl, nbl->nci_tot, nbl->nci_tot/static_cast<double>(grid.numClusters()),
931 numAtomsPerCell/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid.numClusters()*numAtomsCluster/(dims.gridSize[XX]*dims.gridSize[YY]*dims.gridSize[ZZ])));
936 int c[c_gpuNumClusterPerCell + 1] = { 0 };
937 for (const nbnxn_sci_t &sci : nbl->sci)
940 for (int j4 = sci.cj4_ind_start; j4 < sci.cj4_ind_end; j4++)
942 for (int j = 0; j < c_nbnxnGpuJgroupSize; j++)
945 for (int si = 0; si < c_gpuNumClusterPerCell; si++)
947 if (nbl->cj4[j4].imei[0].imask & (1U << (j*c_gpuNumClusterPerCell + si)))
958 nsp_max = std::max(nsp_max, nsp);
960 if (!nbl->sci.empty())
962 sum_nsp /= nbl->sci.size();
963 sum_nsp2 /= nbl->sci.size();
965 fprintf(fp, "nbl #cluster-pairs: av %.1f stddev %.1f max %d\n",
966 sum_nsp, std::sqrt(sum_nsp2 - sum_nsp*sum_nsp), nsp_max);
968 if (!nbl->cj4.empty())
970 for (int b = 0; b <= c_gpuNumClusterPerCell; b++)
972 fprintf(fp, "nbl j-list #i-subcell %d %7d %4.1f\n",
973 b, c[b], 100.0*c[b]/size_t {nbl->cj4.size()*c_nbnxnGpuJgroupSize});
978 /* Returns a pointer to the exclusion mask for j-cluster-group \p cj4 and warp \p warp
979 * Generates a new exclusion entry when the j-cluster-group uses
980 * the default all-interaction mask at call time, so the returned mask
981 * can be modified when needed.
983 static nbnxn_excl_t *get_exclusion_mask(NbnxnPairlistGpu *nbl,
987 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
989 /* No exclusions set, make a new list entry */
990 const size_t oldSize = nbl->excl.size();
991 GMX_ASSERT(oldSize >= 1, "We should always have entry [0]");
992 /* Add entry with default values: no exclusions */
993 nbl->excl.resize(oldSize + 1);
994 nbl->cj4[cj4].imei[warp].excl_ind = oldSize;
997 return &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
1000 static void set_self_and_newton_excls_supersub(NbnxnPairlistGpu *nbl,
1001 int cj4_ind, int sj_offset,
1002 int i_cluster_in_cell)
1004 nbnxn_excl_t *excl[c_nbnxnGpuClusterpairSplit];
1006 /* Here we only set the set self and double pair exclusions */
1008 /* Reserve extra elements, so the resize() in get_exclusion_mask()
1009 * will not invalidate excl entries in the loop below
1011 nbl->excl.reserve(nbl->excl.size() + c_nbnxnGpuClusterpairSplit);
1012 for (int w = 0; w < c_nbnxnGpuClusterpairSplit; w++)
1014 excl[w] = get_exclusion_mask(nbl, cj4_ind, w);
1017 /* Only minor < major bits set */
1018 for (int ej = 0; ej < nbl->na_ci; ej++)
1021 for (int ei = ej; ei < nbl->na_ci; ei++)
1023 excl[w]->pair[(ej & (c_nbnxnGpuJgroupSize-1))*nbl->na_ci + ei] &=
1024 ~(1U << (sj_offset*c_gpuNumClusterPerCell + i_cluster_in_cell));
1029 /* Returns a diagonal or off-diagonal interaction mask for plain C lists */
1030 static unsigned int get_imask(gmx_bool rdiag, int ci, int cj)
1032 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
1035 /* Returns a diagonal or off-diagonal interaction mask for cj-size=2 */
1036 gmx_unused static unsigned int get_imask_simd_j2(gmx_bool rdiag, int ci, int cj)
1038 return (rdiag && ci*2 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_0 :
1039 (rdiag && ci*2+1 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_1 :
1040 NBNXN_INTERACTION_MASK_ALL));
1043 /* Returns a diagonal or off-diagonal interaction mask for cj-size=4 */
1044 gmx_unused static unsigned int get_imask_simd_j4(gmx_bool rdiag, int ci, int cj)
1046 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
1049 /* Returns a diagonal or off-diagonal interaction mask for cj-size=8 */
1050 gmx_unused static unsigned int get_imask_simd_j8(gmx_bool rdiag, int ci, int cj)
1052 return (rdiag && ci == cj*2 ? NBNXN_INTERACTION_MASK_DIAG_J8_0 :
1053 (rdiag && ci == cj*2+1 ? NBNXN_INTERACTION_MASK_DIAG_J8_1 :
1054 NBNXN_INTERACTION_MASK_ALL));
1058 #if GMX_SIMD_REAL_WIDTH == 2
1059 #define get_imask_simd_4xn get_imask_simd_j2
1061 #if GMX_SIMD_REAL_WIDTH == 4
1062 #define get_imask_simd_4xn get_imask_simd_j4
1064 #if GMX_SIMD_REAL_WIDTH == 8
1065 #define get_imask_simd_4xn get_imask_simd_j8
1066 #define get_imask_simd_2xnn get_imask_simd_j4
1068 #if GMX_SIMD_REAL_WIDTH == 16
1069 #define get_imask_simd_2xnn get_imask_simd_j8
1073 /* Plain C code for checking and adding cluster-pairs to the list.
1075 * \param[in] gridj The j-grid
1076 * \param[in,out] nbl The pair-list to store the cluster pairs in
1077 * \param[in] icluster The index of the i-cluster
1078 * \param[in] jclusterFirst The first cluster in the j-range
1079 * \param[in] jclusterLast The last cluster in the j-range
1080 * \param[in] excludeSubDiagonal Exclude atom pairs with i-index > j-index
1081 * \param[in] x_j Coordinates for the j-atom, in xyz format
1082 * \param[in] rlist2 The squared list cut-off
1083 * \param[in] rbb2 The squared cut-off for putting cluster-pairs in the list based on bounding box distance only
1084 * \param[in,out] numDistanceChecks The number of distance checks performed
1087 makeClusterListSimple(const Grid &jGrid,
1088 NbnxnPairlistCpu * nbl,
1092 bool excludeSubDiagonal,
1093 const real * gmx_restrict x_j,
1096 int * gmx_restrict numDistanceChecks)
1098 const nbnxn_bb_t * gmx_restrict bb_ci = nbl->work->iClusterData.bb.data();
1099 const real * gmx_restrict x_ci = nbl->work->iClusterData.x.data();
1104 while (!InRange && jclusterFirst <= jclusterLast)
1106 real d2 = subc_bb_dist2(0, bb_ci, jclusterFirst, jGrid.jBoundingBoxes());
1107 *numDistanceChecks += 2;
1109 /* Check if the distance is within the distance where
1110 * we use only the bounding box distance rbb,
1111 * or within the cut-off and there is at least one atom pair
1112 * within the cut-off.
1118 else if (d2 < rlist2)
1120 int cjf_gl = jGrid.cellOffset() + jclusterFirst;
1121 for (int i = 0; i < c_nbnxnCpuIClusterSize && !InRange; i++)
1123 for (int j = 0; j < c_nbnxnCpuIClusterSize; j++)
1125 InRange = InRange ||
1126 (gmx::square(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjf_gl*c_nbnxnCpuIClusterSize+j)*STRIDE_XYZ+XX]) +
1127 gmx::square(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjf_gl*c_nbnxnCpuIClusterSize+j)*STRIDE_XYZ+YY]) +
1128 gmx::square(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjf_gl*c_nbnxnCpuIClusterSize+j)*STRIDE_XYZ+ZZ]) < rlist2);
1131 *numDistanceChecks += c_nbnxnCpuIClusterSize*c_nbnxnCpuIClusterSize;
1144 while (!InRange && jclusterLast > jclusterFirst)
1146 real d2 = subc_bb_dist2(0, bb_ci, jclusterLast, jGrid.jBoundingBoxes());
1147 *numDistanceChecks += 2;
1149 /* Check if the distance is within the distance where
1150 * we use only the bounding box distance rbb,
1151 * or within the cut-off and there is at least one atom pair
1152 * within the cut-off.
1158 else if (d2 < rlist2)
1160 int cjl_gl = jGrid.cellOffset() + jclusterLast;
1161 for (int i = 0; i < c_nbnxnCpuIClusterSize && !InRange; i++)
1163 for (int j = 0; j < c_nbnxnCpuIClusterSize; j++)
1165 InRange = InRange ||
1166 (gmx::square(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjl_gl*c_nbnxnCpuIClusterSize+j)*STRIDE_XYZ+XX]) +
1167 gmx::square(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjl_gl*c_nbnxnCpuIClusterSize+j)*STRIDE_XYZ+YY]) +
1168 gmx::square(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjl_gl*c_nbnxnCpuIClusterSize+j)*STRIDE_XYZ+ZZ]) < rlist2);
1171 *numDistanceChecks += c_nbnxnCpuIClusterSize*c_nbnxnCpuIClusterSize;
1179 if (jclusterFirst <= jclusterLast)
1181 for (int jcluster = jclusterFirst; jcluster <= jclusterLast; jcluster++)
1183 /* Store cj and the interaction mask */
1185 cjEntry.cj = jGrid.cellOffset() + jcluster;
1186 cjEntry.excl = get_imask(excludeSubDiagonal, icluster, jcluster);
1187 nbl->cj.push_back(cjEntry);
1189 /* Increase the closing index in the i list */
1190 nbl->ci.back().cj_ind_end = nbl->cj.size();
1194 #ifdef GMX_NBNXN_SIMD_4XN
1195 #include "gromacs/nbnxm/pairlist_simd_4xm.h"
1197 #ifdef GMX_NBNXN_SIMD_2XNN
1198 #include "gromacs/nbnxm/pairlist_simd_2xmm.h"
1201 /* Plain C or SIMD4 code for making a pair list of super-cell sci vs scj.
1202 * Checks bounding box distances and possibly atom pair distances.
1204 static void make_cluster_list_supersub(const Grid &iGrid,
1206 NbnxnPairlistGpu *nbl,
1209 const bool excludeSubDiagonal,
1214 int *numDistanceChecks)
1216 NbnxnPairlistGpuWork &work = *nbl->work;
1219 const float *pbb_ci = work.iSuperClusterData.bbPacked.data();
1221 const nbnxn_bb_t *bb_ci = work.iSuperClusterData.bb.data();
1224 assert(c_nbnxnGpuClusterSize == iGrid.geometry().numAtomsICluster);
1225 assert(c_nbnxnGpuClusterSize == jGrid.geometry().numAtomsICluster);
1227 /* We generate the pairlist mainly based on bounding-box distances
1228 * and do atom pair distance based pruning on the GPU.
1229 * Only if a j-group contains a single cluster-pair, we try to prune
1230 * that pair based on atom distances on the CPU to avoid empty j-groups.
1232 #define PRUNE_LIST_CPU_ONE 1
1233 #define PRUNE_LIST_CPU_ALL 0
1235 #if PRUNE_LIST_CPU_ONE
1239 float *d2l = work.distanceBuffer.data();
1241 for (int subc = 0; subc < jGrid.numClustersPerCell()[scj]; subc++)
1243 const int cj4_ind = work.cj_ind/c_nbnxnGpuJgroupSize;
1244 const int cj_offset = work.cj_ind - cj4_ind*c_nbnxnGpuJgroupSize;
1245 const int cj = scj*c_gpuNumClusterPerCell + subc;
1247 const int cj_gl = jGrid.cellOffset()*c_gpuNumClusterPerCell + cj;
1250 if (excludeSubDiagonal && sci == scj)
1256 ci1 = iGrid.numClustersPerCell()[sci];
1260 /* Determine all ci1 bb distances in one call with SIMD4 */
1261 subc_bb_dist2_simd4_xxxx(jGrid.packedBoundingBoxes().data() + (cj >> STRIDE_PBB_2LOG)*NNBSBB_XXXX + (cj & (STRIDE_PBB-1)),
1263 *numDistanceChecks += c_nbnxnGpuClusterSize*2;
1267 unsigned int imask = 0;
1268 /* We use a fixed upper-bound instead of ci1 to help optimization */
1269 for (int ci = 0; ci < c_gpuNumClusterPerCell; ci++)
1277 /* Determine the bb distance between ci and cj */
1278 d2l[ci] = subc_bb_dist2(ci, bb_ci, cj, jGrid.jBoundingBoxes());
1279 *numDistanceChecks += 2;
1283 #if PRUNE_LIST_CPU_ALL
1284 /* Check if the distance is within the distance where
1285 * we use only the bounding box distance rbb,
1286 * or within the cut-off and there is at least one atom pair
1287 * within the cut-off. This check is very costly.
1289 *numDistanceChecks += c_nbnxnGpuClusterSize*c_nbnxnGpuClusterSize;
1292 clusterpair_in_range(work, ci, cj_gl, stride, x, rlist2)))
1294 /* Check if the distance between the two bounding boxes
1295 * in within the pair-list cut-off.
1300 /* Flag this i-subcell to be taken into account */
1301 imask |= (1U << (cj_offset*c_gpuNumClusterPerCell + ci));
1303 #if PRUNE_LIST_CPU_ONE
1311 #if PRUNE_LIST_CPU_ONE
1312 /* If we only found 1 pair, check if any atoms are actually
1313 * within the cut-off, so we could get rid of it.
1315 if (npair == 1 && d2l[ci_last] >= rbb2 &&
1316 !clusterpair_in_range(work, ci_last, cj_gl, stride, x, rlist2))
1318 imask &= ~(1U << (cj_offset*c_gpuNumClusterPerCell + ci_last));
1325 /* We have at least one cluster pair: add a j-entry */
1326 if (static_cast<size_t>(cj4_ind) == nbl->cj4.size())
1328 nbl->cj4.resize(nbl->cj4.size() + 1);
1330 nbnxn_cj4_t *cj4 = &nbl->cj4[cj4_ind];
1332 cj4->cj[cj_offset] = cj_gl;
1334 /* Set the exclusions for the ci==sj entry.
1335 * Here we don't bother to check if this entry is actually flagged,
1336 * as it will nearly always be in the list.
1338 if (excludeSubDiagonal && sci == scj)
1340 set_self_and_newton_excls_supersub(nbl, cj4_ind, cj_offset, subc);
1343 /* Copy the cluster interaction mask to the list */
1344 for (int w = 0; w < c_nbnxnGpuClusterpairSplit; w++)
1346 cj4->imei[w].imask |= imask;
1349 nbl->work->cj_ind++;
1351 /* Keep the count */
1352 nbl->nci_tot += npair;
1354 /* Increase the closing index in i super-cell list */
1355 nbl->sci.back().cj4_ind_end =
1356 (nbl->work->cj_ind + c_nbnxnGpuJgroupSize - 1)/c_nbnxnGpuJgroupSize;
1361 /* Returns how many contiguous j-clusters we have starting in the i-list */
1362 template <typename CjListType>
1363 static int numContiguousJClusters(const int cjIndexStart,
1364 const int cjIndexEnd,
1365 gmx::ArrayRef<const CjListType> cjList)
1367 const int firstJCluster = nblCj(cjList, cjIndexStart);
1369 int numContiguous = 0;
1371 while (cjIndexStart + numContiguous < cjIndexEnd &&
1372 nblCj(cjList, cjIndexStart + numContiguous) == firstJCluster + numContiguous)
1377 return numContiguous;
1381 * \brief Helper struct for efficient searching for excluded atoms in a j-list
1385 /*! \brief Constructs a j-list range from \p cjList with the given index range */
1386 template <typename CjListType>
1387 JListRanges(int cjIndexStart,
1389 gmx::ArrayRef<const CjListType> cjList);
1391 int cjIndexStart; //!< The start index in the j-list
1392 int cjIndexEnd; //!< The end index in the j-list
1393 int cjFirst; //!< The j-cluster with index cjIndexStart
1394 int cjLast; //!< The j-cluster with index cjIndexEnd-1
1395 int numDirect; //!< Up to cjIndexStart+numDirect the j-clusters are cjFirst + the index offset
1399 template <typename CjListType>
1400 JListRanges::JListRanges(int cjIndexStart,
1402 gmx::ArrayRef<const CjListType> cjList) :
1403 cjIndexStart(cjIndexStart),
1404 cjIndexEnd(cjIndexEnd)
1406 GMX_ASSERT(cjIndexEnd > cjIndexStart, "JListRanges should only be called with non-empty lists");
1408 cjFirst = nblCj(cjList, cjIndexStart);
1409 cjLast = nblCj(cjList, cjIndexEnd - 1);
1411 /* Determine how many contiguous j-cells we have starting
1412 * from the first i-cell. This number can be used to directly
1413 * calculate j-cell indices for excluded atoms.
1415 numDirect = numContiguousJClusters(cjIndexStart, cjIndexEnd, cjList);
1419 /* Return the index of \p jCluster in the given range or -1 when not present
1421 * Note: This code is executed very often and therefore performance is
1422 * important. It should be inlined and fully optimized.
1424 template <typename CjListType>
1426 findJClusterInJList(int jCluster,
1427 const JListRanges &ranges,
1428 gmx::ArrayRef<const CjListType> cjList)
1432 if (jCluster < ranges.cjFirst + ranges.numDirect)
1434 /* We can calculate the index directly using the offset */
1435 index = ranges.cjIndexStart + jCluster - ranges.cjFirst;
1439 /* Search for jCluster using bisection */
1441 int rangeStart = ranges.cjIndexStart + ranges.numDirect;
1442 int rangeEnd = ranges.cjIndexEnd;
1444 while (index == -1 && rangeStart < rangeEnd)
1446 rangeMiddle = (rangeStart + rangeEnd) >> 1;
1448 const int clusterMiddle = nblCj(cjList, rangeMiddle);
1450 if (jCluster == clusterMiddle)
1452 index = rangeMiddle;
1454 else if (jCluster < clusterMiddle)
1456 rangeEnd = rangeMiddle;
1460 rangeStart = rangeMiddle + 1;
1468 // TODO: Get rid of the two functions below by renaming sci to ci (or something better)
1470 /* Return the i-entry in the list we are currently operating on */
1471 static nbnxn_ci_t *getOpenIEntry(NbnxnPairlistCpu *nbl)
1473 return &nbl->ci.back();
1476 /* Return the i-entry in the list we are currently operating on */
1477 static nbnxn_sci_t *getOpenIEntry(NbnxnPairlistGpu *nbl)
1479 return &nbl->sci.back();
1482 /* Set all atom-pair exclusions for a simple type list i-entry
1484 * Set all atom-pair exclusions from the topology stored in exclusions
1485 * as masks in the pair-list for simple list entry iEntry.
1488 setExclusionsForIEntry(const nbnxn_search *nbs,
1489 NbnxnPairlistCpu *nbl,
1490 gmx_bool diagRemoved,
1492 const nbnxn_ci_t &iEntry,
1493 const t_blocka &exclusions)
1495 if (iEntry.cj_ind_end == iEntry.cj_ind_start)
1497 /* Empty list: no exclusions */
1501 const JListRanges ranges(iEntry.cj_ind_start, iEntry.cj_ind_end, gmx::makeConstArrayRef(nbl->cj));
1503 const int iCluster = iEntry.ci;
1505 gmx::ArrayRef<const int> cell = nbs->cell;
1507 /* Loop over the atoms in the i-cluster */
1508 for (int i = 0; i < nbl->na_ci; i++)
1510 const int iIndex = iCluster*nbl->na_ci + i;
1511 const int iAtom = nbs->a[iIndex];
1514 /* Loop over the topology-based exclusions for this i-atom */
1515 for (int exclIndex = exclusions.index[iAtom]; exclIndex < exclusions.index[iAtom + 1]; exclIndex++)
1517 const int jAtom = exclusions.a[exclIndex];
1521 /* The self exclusion are already set, save some time */
1525 /* Get the index of the j-atom in the nbnxn atom data */
1526 const int jIndex = cell[jAtom];
1528 /* Without shifts we only calculate interactions j>i
1529 * for one-way pair-lists.
1531 if (diagRemoved && jIndex <= iIndex)
1536 const int jCluster = (jIndex >> na_cj_2log);
1538 /* Could the cluster se be in our list? */
1539 if (jCluster >= ranges.cjFirst && jCluster <= ranges.cjLast)
1542 findJClusterInJList(jCluster, ranges,
1543 gmx::makeConstArrayRef(nbl->cj));
1547 /* We found an exclusion, clear the corresponding
1550 const int innerJ = jIndex - (jCluster << na_cj_2log);
1552 nbl->cj[index].excl &= ~(1U << ((i << na_cj_2log) + innerJ));
1560 /* Add a new i-entry to the FEP list and copy the i-properties */
1561 static inline void fep_list_new_nri_copy(t_nblist *nlist)
1563 /* Add a new i-entry */
1566 assert(nlist->nri < nlist->maxnri);
1568 /* Duplicate the last i-entry, except for jindex, which continues */
1569 nlist->iinr[nlist->nri] = nlist->iinr[nlist->nri-1];
1570 nlist->shift[nlist->nri] = nlist->shift[nlist->nri-1];
1571 nlist->gid[nlist->nri] = nlist->gid[nlist->nri-1];
1572 nlist->jindex[nlist->nri] = nlist->nrj;
1575 /* For load balancing of the free-energy lists over threads, we set
1576 * the maximum nrj size of an i-entry to 40. This leads to good
1577 * load balancing in the worst case scenario of a single perturbed
1578 * particle on 16 threads, while not introducing significant overhead.
1579 * Note that half of the perturbed pairs will anyhow end up in very small lists,
1580 * since non perturbed i-particles will see few perturbed j-particles).
1582 const int max_nrj_fep = 40;
1584 /* Exclude the perturbed pairs from the Verlet list. This is only done to avoid
1585 * singularities for overlapping particles (0/0), since the charges and
1586 * LJ parameters have been zeroed in the nbnxn data structure.
1587 * Simultaneously make a group pair list for the perturbed pairs.
1589 static void make_fep_list(const nbnxn_search *nbs,
1590 const nbnxn_atomdata_t *nbat,
1591 NbnxnPairlistCpu *nbl,
1592 gmx_bool bDiagRemoved,
1594 real gmx_unused shx,
1595 real gmx_unused shy,
1596 real gmx_unused shz,
1597 real gmx_unused rlist_fep2,
1602 int ci, cj_ind_start, cj_ind_end, cja, cjr;
1604 int gid_i = 0, gid_j, gid;
1605 int egp_shift, egp_mask;
1607 int ind_i, ind_j, ai, aj;
1609 gmx_bool bFEP_i, bFEP_i_all;
1611 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
1619 cj_ind_start = nbl_ci->cj_ind_start;
1620 cj_ind_end = nbl_ci->cj_ind_end;
1622 /* In worst case we have alternating energy groups
1623 * and create #atom-pair lists, which means we need the size
1624 * of a cluster pair (na_ci*na_cj) times the number of cj's.
1626 nri_max = nbl->na_ci*nbl->na_cj*(cj_ind_end - cj_ind_start);
1627 if (nlist->nri + nri_max > nlist->maxnri)
1629 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
1630 reallocate_nblist(nlist);
1633 const int numAtomsJCluster = jGrid.geometry().numAtomsJCluster;
1635 const nbnxn_atomdata_t::Params &nbatParams = nbat->params();
1637 const int ngid = nbatParams.nenergrp;
1639 /* TODO: Consider adding a check in grompp and changing this to an assert */
1640 const int numBitsInEnergyGroupIdsForAtomsInJCluster = sizeof(gid_cj)*8;
1641 if (ngid*numAtomsJCluster > numBitsInEnergyGroupIdsForAtomsInJCluster)
1643 gmx_fatal(FARGS, "The Verlet scheme with %dx%d kernels and free-energy only supports up to %zu energy groups",
1644 iGrid.geometry().numAtomsICluster, numAtomsJCluster,
1645 (sizeof(gid_cj)*8)/numAtomsJCluster);
1648 egp_shift = nbatParams.neg_2log;
1649 egp_mask = (1 << egp_shift) - 1;
1651 /* Loop over the atoms in the i sub-cell */
1653 for (int i = 0; i < nbl->na_ci; i++)
1655 ind_i = ci*nbl->na_ci + i;
1660 nlist->jindex[nri+1] = nlist->jindex[nri];
1661 nlist->iinr[nri] = ai;
1662 /* The actual energy group pair index is set later */
1663 nlist->gid[nri] = 0;
1664 nlist->shift[nri] = nbl_ci->shift & NBNXN_CI_SHIFT;
1666 bFEP_i = iGrid.atomIsPerturbed(ci - iGrid.cellOffset(), i);
1668 bFEP_i_all = bFEP_i_all && bFEP_i;
1670 if (nlist->nrj + (cj_ind_end - cj_ind_start)*nbl->na_cj > nlist->maxnrj)
1672 nlist->maxnrj = over_alloc_small(nlist->nrj + (cj_ind_end - cj_ind_start)*nbl->na_cj);
1673 srenew(nlist->jjnr, nlist->maxnrj);
1674 srenew(nlist->excl_fep, nlist->maxnrj);
1679 gid_i = (nbatParams.energrp[ci] >> (egp_shift*i)) & egp_mask;
1682 for (int cj_ind = cj_ind_start; cj_ind < cj_ind_end; cj_ind++)
1684 unsigned int fep_cj;
1686 cja = nbl->cj[cj_ind].cj;
1688 if (numAtomsJCluster == jGrid.geometry().numAtomsICluster)
1690 cjr = cja - jGrid.cellOffset();
1691 fep_cj = jGrid.fepBits(cjr);
1694 gid_cj = nbatParams.energrp[cja];
1697 else if (2*numAtomsJCluster == jGrid.geometry().numAtomsICluster)
1699 cjr = cja - jGrid.cellOffset()*2;
1700 /* Extract half of the ci fep/energrp mask */
1701 fep_cj = (jGrid.fepBits(cjr >> 1) >> ((cjr & 1)*numAtomsJCluster)) & ((1 << numAtomsJCluster) - 1);
1704 gid_cj = nbatParams.energrp[cja >> 1] >> ((cja & 1)*numAtomsJCluster*egp_shift) & ((1 << (numAtomsJCluster*egp_shift)) - 1);
1709 cjr = cja - (jGrid.cellOffset() >> 1);
1710 /* Combine two ci fep masks/energrp */
1711 fep_cj = jGrid.fepBits(cjr*2) + (jGrid.fepBits(cjr*2 + 1) << jGrid.geometry().numAtomsICluster);
1714 gid_cj = nbatParams.energrp[cja*2] + (nbatParams.energrp[cja*2+1] << (jGrid.geometry().numAtomsICluster*egp_shift));
1718 if (bFEP_i || fep_cj != 0)
1720 for (int j = 0; j < nbl->na_cj; j++)
1722 /* Is this interaction perturbed and not excluded? */
1723 ind_j = cja*nbl->na_cj + j;
1726 (bFEP_i || (fep_cj & (1 << j))) &&
1727 (!bDiagRemoved || ind_j >= ind_i))
1731 gid_j = (gid_cj >> (j*egp_shift)) & egp_mask;
1732 gid = GID(gid_i, gid_j, ngid);
1734 if (nlist->nrj > nlist->jindex[nri] &&
1735 nlist->gid[nri] != gid)
1737 /* Energy group pair changed: new list */
1738 fep_list_new_nri_copy(nlist);
1741 nlist->gid[nri] = gid;
1744 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
1746 fep_list_new_nri_copy(nlist);
1750 /* Add it to the FEP list */
1751 nlist->jjnr[nlist->nrj] = aj;
1752 nlist->excl_fep[nlist->nrj] = (nbl->cj[cj_ind].excl >> (i*nbl->na_cj + j)) & 1;
1755 /* Exclude it from the normal list.
1756 * Note that the charge has been set to zero,
1757 * but we need to avoid 0/0, as perturbed atoms
1758 * can be on top of each other.
1760 nbl->cj[cj_ind].excl &= ~(1U << (i*nbl->na_cj + j));
1766 if (nlist->nrj > nlist->jindex[nri])
1768 /* Actually add this new, non-empty, list */
1770 nlist->jindex[nlist->nri] = nlist->nrj;
1777 /* All interactions are perturbed, we can skip this entry */
1778 nbl_ci->cj_ind_end = cj_ind_start;
1779 nbl->ncjInUse -= cj_ind_end - cj_ind_start;
1783 /* Return the index of atom a within a cluster */
1784 static inline int cj_mod_cj4(int cj)
1786 return cj & (c_nbnxnGpuJgroupSize - 1);
1789 /* Convert a j-cluster to a cj4 group */
1790 static inline int cj_to_cj4(int cj)
1792 return cj/c_nbnxnGpuJgroupSize;
1795 /* Return the index of an j-atom within a warp */
1796 static inline int a_mod_wj(int a)
1798 return a & (c_nbnxnGpuClusterSize/c_nbnxnGpuClusterpairSplit - 1);
1801 /* As make_fep_list above, but for super/sub lists. */
1802 static void make_fep_list(const nbnxn_search *nbs,
1803 const nbnxn_atomdata_t *nbat,
1804 NbnxnPairlistGpu *nbl,
1805 gmx_bool bDiagRemoved,
1806 const nbnxn_sci_t *nbl_sci,
1817 int ind_i, ind_j, ai, aj;
1821 const nbnxn_cj4_t *cj4;
1823 const int numJClusterGroups = nbl_sci->numJClusterGroups();
1824 if (numJClusterGroups == 0)
1830 const int sci = nbl_sci->sci;
1832 const int cj4_ind_start = nbl_sci->cj4_ind_start;
1833 const int cj4_ind_end = nbl_sci->cj4_ind_end;
1835 /* Here we process one super-cell, max #atoms na_sc, versus a list
1836 * cj4 entries, each with max c_nbnxnGpuJgroupSize cj's, each
1837 * of size na_cj atoms.
1838 * On the GPU we don't support energy groups (yet).
1839 * So for each of the na_sc i-atoms, we need max one FEP list
1840 * for each max_nrj_fep j-atoms.
1842 nri_max = nbl->na_sc*nbl->na_cj*(1 + (numJClusterGroups*c_nbnxnGpuJgroupSize)/max_nrj_fep);
1843 if (nlist->nri + nri_max > nlist->maxnri)
1845 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
1846 reallocate_nblist(nlist);
1849 /* Loop over the atoms in the i super-cluster */
1850 for (int c = 0; c < c_gpuNumClusterPerCell; c++)
1852 c_abs = sci*c_gpuNumClusterPerCell + c;
1854 for (int i = 0; i < nbl->na_ci; i++)
1856 ind_i = c_abs*nbl->na_ci + i;
1861 nlist->jindex[nri+1] = nlist->jindex[nri];
1862 nlist->iinr[nri] = ai;
1863 /* With GPUs, energy groups are not supported */
1864 nlist->gid[nri] = 0;
1865 nlist->shift[nri] = nbl_sci->shift & NBNXN_CI_SHIFT;
1867 bFEP_i = iGrid.atomIsPerturbed(c_abs - iGrid.cellOffset()*c_gpuNumClusterPerCell, i);
1869 xi = nbat->x()[ind_i*nbat->xstride+XX] + shx;
1870 yi = nbat->x()[ind_i*nbat->xstride+YY] + shy;
1871 zi = nbat->x()[ind_i*nbat->xstride+ZZ] + shz;
1873 const int nrjMax = nlist->nrj + numJClusterGroups*c_nbnxnGpuJgroupSize*nbl->na_cj;
1874 if (nrjMax > nlist->maxnrj)
1876 nlist->maxnrj = over_alloc_small(nrjMax);
1877 srenew(nlist->jjnr, nlist->maxnrj);
1878 srenew(nlist->excl_fep, nlist->maxnrj);
1881 for (int cj4_ind = cj4_ind_start; cj4_ind < cj4_ind_end; cj4_ind++)
1883 cj4 = &nbl->cj4[cj4_ind];
1885 for (int gcj = 0; gcj < c_nbnxnGpuJgroupSize; gcj++)
1887 if ((cj4->imei[0].imask & (1U << (gcj*c_gpuNumClusterPerCell + c))) == 0)
1889 /* Skip this ci for this cj */
1894 cj4->cj[gcj] - jGrid.cellOffset()*c_gpuNumClusterPerCell;
1896 if (bFEP_i || jGrid.clusterIsPerturbed(cjr))
1898 for (int j = 0; j < nbl->na_cj; j++)
1900 /* Is this interaction perturbed and not excluded? */
1901 ind_j = (jGrid.cellOffset()*c_gpuNumClusterPerCell + cjr)*nbl->na_cj + j;
1904 (bFEP_i || jGrid.atomIsPerturbed(cjr, j)) &&
1905 (!bDiagRemoved || ind_j >= ind_i))
1908 unsigned int excl_bit;
1911 const int jHalf = j/(c_nbnxnGpuClusterSize/c_nbnxnGpuClusterpairSplit);
1912 nbnxn_excl_t *excl =
1913 get_exclusion_mask(nbl, cj4_ind, jHalf);
1915 excl_pair = a_mod_wj(j)*nbl->na_ci + i;
1916 excl_bit = (1U << (gcj*c_gpuNumClusterPerCell + c));
1918 dx = nbat->x()[ind_j*nbat->xstride+XX] - xi;
1919 dy = nbat->x()[ind_j*nbat->xstride+YY] - yi;
1920 dz = nbat->x()[ind_j*nbat->xstride+ZZ] - zi;
1922 /* The unpruned GPU list has more than 2/3
1923 * of the atom pairs beyond rlist. Using
1924 * this list will cause a lot of overhead
1925 * in the CPU FEP kernels, especially
1926 * relative to the fast GPU kernels.
1927 * So we prune the FEP list here.
1929 if (dx*dx + dy*dy + dz*dz < rlist_fep2)
1931 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
1933 fep_list_new_nri_copy(nlist);
1937 /* Add it to the FEP list */
1938 nlist->jjnr[nlist->nrj] = aj;
1939 nlist->excl_fep[nlist->nrj] = (excl->pair[excl_pair] & excl_bit) ? 1 : 0;
1943 /* Exclude it from the normal list.
1944 * Note that the charge and LJ parameters have
1945 * been set to zero, but we need to avoid 0/0,
1946 * as perturbed atoms can be on top of each other.
1948 excl->pair[excl_pair] &= ~excl_bit;
1952 /* Note that we could mask out this pair in imask
1953 * if all i- and/or all j-particles are perturbed.
1954 * But since the perturbed pairs on the CPU will
1955 * take an order of magnitude more time, the GPU
1956 * will finish before the CPU and there is no gain.
1962 if (nlist->nrj > nlist->jindex[nri])
1964 /* Actually add this new, non-empty, list */
1966 nlist->jindex[nlist->nri] = nlist->nrj;
1973 /* Set all atom-pair exclusions for a GPU type list i-entry
1975 * Sets all atom-pair exclusions from the topology stored in exclusions
1976 * as masks in the pair-list for i-super-cluster list entry iEntry.
1979 setExclusionsForIEntry(const nbnxn_search *nbs,
1980 NbnxnPairlistGpu *nbl,
1981 gmx_bool diagRemoved,
1982 int gmx_unused na_cj_2log,
1983 const nbnxn_sci_t &iEntry,
1984 const t_blocka &exclusions)
1986 if (iEntry.numJClusterGroups() == 0)
1992 /* Set the search ranges using start and end j-cluster indices.
1993 * Note that here we can not use cj4_ind_end, since the last cj4
1994 * can be only partially filled, so we use cj_ind.
1996 const JListRanges ranges(iEntry.cj4_ind_start*c_nbnxnGpuJgroupSize,
1998 gmx::makeConstArrayRef(nbl->cj4));
2000 GMX_ASSERT(nbl->na_ci == c_nbnxnGpuClusterSize, "na_ci should match the GPU cluster size");
2001 constexpr int c_clusterSize = c_nbnxnGpuClusterSize;
2002 constexpr int c_superClusterSize = c_nbnxnGpuNumClusterPerSupercluster*c_nbnxnGpuClusterSize;
2004 const int iSuperCluster = iEntry.sci;
2006 gmx::ArrayRef<const int> cell = nbs->cell;
2008 /* Loop over the atoms in the i super-cluster */
2009 for (int i = 0; i < c_superClusterSize; i++)
2011 const int iIndex = iSuperCluster*c_superClusterSize + i;
2012 const int iAtom = nbs->a[iIndex];
2015 const int iCluster = i/c_clusterSize;
2017 /* Loop over the topology-based exclusions for this i-atom */
2018 for (int exclIndex = exclusions.index[iAtom]; exclIndex < exclusions.index[iAtom + 1]; exclIndex++)
2020 const int jAtom = exclusions.a[exclIndex];
2024 /* The self exclusions are already set, save some time */
2028 /* Get the index of the j-atom in the nbnxn atom data */
2029 const int jIndex = cell[jAtom];
2031 /* Without shifts we only calculate interactions j>i
2032 * for one-way pair-lists.
2034 /* NOTE: We would like to use iIndex on the right hand side,
2035 * but that makes this routine 25% slower with gcc6/7.
2036 * Even using c_superClusterSize makes it slower.
2037 * Either of these changes triggers peeling of the exclIndex
2038 * loop, which apparently leads to far less efficient code.
2040 if (diagRemoved && jIndex <= iSuperCluster*nbl->na_sc + i)
2045 const int jCluster = jIndex/c_clusterSize;
2047 /* Check whether the cluster is in our list? */
2048 if (jCluster >= ranges.cjFirst && jCluster <= ranges.cjLast)
2051 findJClusterInJList(jCluster, ranges,
2052 gmx::makeConstArrayRef(nbl->cj4));
2056 /* We found an exclusion, clear the corresponding
2059 const unsigned int pairMask = (1U << (cj_mod_cj4(index)*c_gpuNumClusterPerCell + iCluster));
2060 /* Check if the i-cluster interacts with the j-cluster */
2061 if (nbl_imask0(nbl, index) & pairMask)
2063 const int innerI = (i & (c_clusterSize - 1));
2064 const int innerJ = (jIndex & (c_clusterSize - 1));
2066 /* Determine which j-half (CUDA warp) we are in */
2067 const int jHalf = innerJ/(c_clusterSize/c_nbnxnGpuClusterpairSplit);
2069 nbnxn_excl_t *interactionMask =
2070 get_exclusion_mask(nbl, cj_to_cj4(index), jHalf);
2072 interactionMask->pair[a_mod_wj(innerJ)*c_clusterSize + innerI] &= ~pairMask;
2081 /* Make a new ci entry at the back of nbl->ci */
2082 static void addNewIEntry(NbnxnPairlistCpu *nbl, int ci, int shift, int flags)
2086 ciEntry.shift = shift;
2087 /* Store the interaction flags along with the shift */
2088 ciEntry.shift |= flags;
2089 ciEntry.cj_ind_start = nbl->cj.size();
2090 ciEntry.cj_ind_end = nbl->cj.size();
2091 nbl->ci.push_back(ciEntry);
2094 /* Make a new sci entry at index nbl->nsci */
2095 static void addNewIEntry(NbnxnPairlistGpu *nbl, int sci, int shift, int gmx_unused flags)
2097 nbnxn_sci_t sciEntry;
2099 sciEntry.shift = shift;
2100 sciEntry.cj4_ind_start = nbl->cj4.size();
2101 sciEntry.cj4_ind_end = nbl->cj4.size();
2103 nbl->sci.push_back(sciEntry);
2106 /* Sort the simple j-list cj on exclusions.
2107 * Entries with exclusions will all be sorted to the beginning of the list.
2109 static void sort_cj_excl(nbnxn_cj_t *cj, int ncj,
2110 NbnxnPairlistCpuWork *work)
2112 work->cj.resize(ncj);
2114 /* Make a list of the j-cells involving exclusions */
2116 for (int j = 0; j < ncj; j++)
2118 if (cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
2120 work->cj[jnew++] = cj[j];
2123 /* Check if there are exclusions at all or not just the first entry */
2124 if (!((jnew == 0) ||
2125 (jnew == 1 && cj[0].excl != NBNXN_INTERACTION_MASK_ALL)))
2127 for (int j = 0; j < ncj; j++)
2129 if (cj[j].excl == NBNXN_INTERACTION_MASK_ALL)
2131 work->cj[jnew++] = cj[j];
2134 for (int j = 0; j < ncj; j++)
2136 cj[j] = work->cj[j];
2141 /* Close this simple list i entry */
2142 static void closeIEntry(NbnxnPairlistCpu *nbl,
2143 int gmx_unused sp_max_av,
2144 gmx_bool gmx_unused progBal,
2145 float gmx_unused nsp_tot_est,
2146 int gmx_unused thread,
2147 int gmx_unused nthread)
2149 nbnxn_ci_t &ciEntry = nbl->ci.back();
2151 /* All content of the new ci entry have already been filled correctly,
2152 * we only need to sort and increase counts or remove the entry when empty.
2154 const int jlen = ciEntry.cj_ind_end - ciEntry.cj_ind_start;
2157 sort_cj_excl(nbl->cj.data() + ciEntry.cj_ind_start, jlen, nbl->work);
2159 /* The counts below are used for non-bonded pair/flop counts
2160 * and should therefore match the available kernel setups.
2162 if (!(ciEntry.shift & NBNXN_CI_DO_COUL(0)))
2164 nbl->work->ncj_noq += jlen;
2166 else if ((ciEntry.shift & NBNXN_CI_HALF_LJ(0)) ||
2167 !(ciEntry.shift & NBNXN_CI_DO_LJ(0)))
2169 nbl->work->ncj_hlj += jlen;
2174 /* Entry is empty: remove it */
2179 /* Split sci entry for load balancing on the GPU.
2180 * Splitting ensures we have enough lists to fully utilize the whole GPU.
2181 * With progBal we generate progressively smaller lists, which improves
2182 * load balancing. As we only know the current count on our own thread,
2183 * we will need to estimate the current total amount of i-entries.
2184 * As the lists get concatenated later, this estimate depends
2185 * both on nthread and our own thread index.
2187 static void split_sci_entry(NbnxnPairlistGpu *nbl,
2189 gmx_bool progBal, float nsp_tot_est,
2190 int thread, int nthread)
2198 /* Estimate the total numbers of ci's of the nblist combined
2199 * over all threads using the target number of ci's.
2201 nsp_est = (nsp_tot_est*thread)/nthread + nbl->nci_tot;
2203 /* The first ci blocks should be larger, to avoid overhead.
2204 * The last ci blocks should be smaller, to improve load balancing.
2205 * The factor 3/2 makes the first block 3/2 times the target average
2206 * and ensures that the total number of blocks end up equal to
2207 * that of equally sized blocks of size nsp_target_av.
2209 nsp_max = static_cast<int>(nsp_target_av*(nsp_tot_est*1.5/(nsp_est + nsp_tot_est)));
2213 nsp_max = nsp_target_av;
2216 const int cj4_start = nbl->sci.back().cj4_ind_start;
2217 const int cj4_end = nbl->sci.back().cj4_ind_end;
2218 const int j4len = cj4_end - cj4_start;
2220 if (j4len > 1 && j4len*c_gpuNumClusterPerCell*c_nbnxnGpuJgroupSize > nsp_max)
2222 /* Modify the last ci entry and process the cj4's again */
2228 for (int cj4 = cj4_start; cj4 < cj4_end; cj4++)
2230 int nsp_cj4_p = nsp_cj4;
2231 /* Count the number of cluster pairs in this cj4 group */
2233 for (int p = 0; p < c_gpuNumClusterPerCell*c_nbnxnGpuJgroupSize; p++)
2235 nsp_cj4 += (nbl->cj4[cj4].imei[0].imask >> p) & 1;
2238 /* If adding the current cj4 with nsp_cj4 pairs get us further
2239 * away from our target nsp_max, split the list before this cj4.
2241 if (nsp > 0 && nsp_max - nsp < nsp + nsp_cj4 - nsp_max)
2243 /* Split the list at cj4 */
2244 nbl->sci.back().cj4_ind_end = cj4;
2245 /* Create a new sci entry */
2247 sciNew.sci = nbl->sci.back().sci;
2248 sciNew.shift = nbl->sci.back().shift;
2249 sciNew.cj4_ind_start = cj4;
2250 nbl->sci.push_back(sciNew);
2253 nsp_cj4_e = nsp_cj4_p;
2259 /* Put the remaining cj4's in the last sci entry */
2260 nbl->sci.back().cj4_ind_end = cj4_end;
2262 /* Possibly balance out the last two sci's
2263 * by moving the last cj4 of the second last sci.
2265 if (nsp_sci - nsp_cj4_e >= nsp + nsp_cj4_e)
2267 GMX_ASSERT(nbl->sci.size() >= 2, "We expect at least two elements");
2268 nbl->sci[nbl->sci.size() - 2].cj4_ind_end--;
2269 nbl->sci[nbl->sci.size() - 1].cj4_ind_start--;
2274 /* Clost this super/sub list i entry */
2275 static void closeIEntry(NbnxnPairlistGpu *nbl,
2277 gmx_bool progBal, float nsp_tot_est,
2278 int thread, int nthread)
2280 nbnxn_sci_t &sciEntry = *getOpenIEntry(nbl);
2282 /* All content of the new ci entry have already been filled correctly,
2283 * we only need to, potentially, split or remove the entry when empty.
2285 int j4len = sciEntry.numJClusterGroups();
2288 /* We can only have complete blocks of 4 j-entries in a list,
2289 * so round the count up before closing.
2291 int ncj4 = (nbl->work->cj_ind + c_nbnxnGpuJgroupSize - 1)/c_nbnxnGpuJgroupSize;
2292 nbl->work->cj_ind = ncj4*c_nbnxnGpuJgroupSize;
2296 /* Measure the size of the new entry and potentially split it */
2297 split_sci_entry(nbl, nsp_max_av, progBal, nsp_tot_est,
2303 /* Entry is empty: remove it */
2304 nbl->sci.pop_back();
2308 /* Syncs the working array before adding another grid pair to the GPU list */
2309 static void sync_work(NbnxnPairlistCpu gmx_unused *nbl)
2313 /* Syncs the working array before adding another grid pair to the GPU list */
2314 static void sync_work(NbnxnPairlistGpu *nbl)
2316 nbl->work->cj_ind = nbl->cj4.size()*c_nbnxnGpuJgroupSize;
2319 /* Clears an NbnxnPairlistCpu data structure */
2320 static void clear_pairlist(NbnxnPairlistCpu *nbl)
2326 nbl->ciOuter.clear();
2327 nbl->cjOuter.clear();
2329 nbl->work->ncj_noq = 0;
2330 nbl->work->ncj_hlj = 0;
2333 /* Clears an NbnxnPairlistGpu data structure */
2334 static void clear_pairlist(NbnxnPairlistGpu *nbl)
2338 nbl->excl.resize(1);
2342 /* Clears a group scheme pair list */
2343 static void clear_pairlist_fep(t_nblist *nl)
2347 if (nl->jindex == nullptr)
2349 snew(nl->jindex, 1);
2354 /* Sets a simple list i-cell bounding box, including PBC shift */
2355 static inline void set_icell_bb_simple(gmx::ArrayRef<const nbnxn_bb_t> bb,
2357 real shx, real shy, real shz,
2360 bb_ci->lower.x = bb[ci].lower.x + shx;
2361 bb_ci->lower.y = bb[ci].lower.y + shy;
2362 bb_ci->lower.z = bb[ci].lower.z + shz;
2363 bb_ci->upper.x = bb[ci].upper.x + shx;
2364 bb_ci->upper.y = bb[ci].upper.y + shy;
2365 bb_ci->upper.z = bb[ci].upper.z + shz;
2368 /* Sets a simple list i-cell bounding box, including PBC shift */
2369 static inline void set_icell_bb(const Grid &iGrid,
2371 real shx, real shy, real shz,
2372 NbnxnPairlistCpuWork *work)
2374 set_icell_bb_simple(iGrid.iBoundingBoxes(), ci, shx, shy, shz,
2375 &work->iClusterData.bb[0]);
2379 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
2380 static void set_icell_bbxxxx_supersub(gmx::ArrayRef<const float> bb,
2382 real shx, real shy, real shz,
2385 int ia = ci*(c_gpuNumClusterPerCell >> STRIDE_PBB_2LOG)*NNBSBB_XXXX;
2386 for (int m = 0; m < (c_gpuNumClusterPerCell >> STRIDE_PBB_2LOG)*NNBSBB_XXXX; m += NNBSBB_XXXX)
2388 for (int i = 0; i < STRIDE_PBB; i++)
2390 bb_ci[m+0*STRIDE_PBB+i] = bb[ia+m+0*STRIDE_PBB+i] + shx;
2391 bb_ci[m+1*STRIDE_PBB+i] = bb[ia+m+1*STRIDE_PBB+i] + shy;
2392 bb_ci[m+2*STRIDE_PBB+i] = bb[ia+m+2*STRIDE_PBB+i] + shz;
2393 bb_ci[m+3*STRIDE_PBB+i] = bb[ia+m+3*STRIDE_PBB+i] + shx;
2394 bb_ci[m+4*STRIDE_PBB+i] = bb[ia+m+4*STRIDE_PBB+i] + shy;
2395 bb_ci[m+5*STRIDE_PBB+i] = bb[ia+m+5*STRIDE_PBB+i] + shz;
2401 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
2402 gmx_unused static void set_icell_bb_supersub(gmx::ArrayRef<const nbnxn_bb_t> bb,
2404 real shx, real shy, real shz,
2407 for (int i = 0; i < c_gpuNumClusterPerCell; i++)
2409 set_icell_bb_simple(bb, ci*c_gpuNumClusterPerCell+i,
2415 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
2416 gmx_unused static void set_icell_bb(const Grid &iGrid,
2418 real shx, real shy, real shz,
2419 NbnxnPairlistGpuWork *work)
2422 set_icell_bbxxxx_supersub(iGrid.packedBoundingBoxes(), ci, shx, shy, shz,
2423 work->iSuperClusterData.bbPacked.data());
2425 set_icell_bb_supersub(iGrid.iBoundingBoxes(), ci, shx, shy, shz,
2426 work->iSuperClusterData.bb.data());
2430 /* Copies PBC shifted i-cell atom coordinates x,y,z to working array */
2431 static void icell_set_x_simple(int ci,
2432 real shx, real shy, real shz,
2433 int stride, const real *x,
2434 NbnxnPairlistCpuWork::IClusterData *iClusterData)
2436 const int ia = ci*c_nbnxnCpuIClusterSize;
2438 for (int i = 0; i < c_nbnxnCpuIClusterSize; i++)
2440 iClusterData->x[i*STRIDE_XYZ+XX] = x[(ia+i)*stride+XX] + shx;
2441 iClusterData->x[i*STRIDE_XYZ+YY] = x[(ia+i)*stride+YY] + shy;
2442 iClusterData->x[i*STRIDE_XYZ+ZZ] = x[(ia+i)*stride+ZZ] + shz;
2446 static void icell_set_x(int ci,
2447 real shx, real shy, real shz,
2448 int stride, const real *x,
2449 const Nbnxm::KernelType kernelType,
2450 NbnxnPairlistCpuWork *work)
2455 #ifdef GMX_NBNXN_SIMD_4XN
2456 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
2457 icell_set_x_simd_4xn(ci, shx, shy, shz, stride, x, work);
2460 #ifdef GMX_NBNXN_SIMD_2XNN
2461 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
2462 icell_set_x_simd_2xnn(ci, shx, shy, shz, stride, x, work);
2466 case Nbnxm::KernelType::Cpu4x4_PlainC:
2467 icell_set_x_simple(ci, shx, shy, shz, stride, x, &work->iClusterData);
2470 GMX_ASSERT(false, "Unhandled case");
2475 /* Copies PBC shifted super-cell atom coordinates x,y,z to working array */
2476 static void icell_set_x(int ci,
2477 real shx, real shy, real shz,
2478 int stride, const real *x,
2479 Nbnxm::KernelType gmx_unused kernelType,
2480 NbnxnPairlistGpuWork *work)
2482 #if !GMX_SIMD4_HAVE_REAL
2484 real * x_ci = work->iSuperClusterData.x.data();
2486 int ia = ci*c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize;
2487 for (int i = 0; i < c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize; i++)
2489 x_ci[i*DIM + XX] = x[(ia+i)*stride + XX] + shx;
2490 x_ci[i*DIM + YY] = x[(ia+i)*stride + YY] + shy;
2491 x_ci[i*DIM + ZZ] = x[(ia+i)*stride + ZZ] + shz;
2494 #else /* !GMX_SIMD4_HAVE_REAL */
2496 real * x_ci = work->iSuperClusterData.xSimd.data();
2498 for (int si = 0; si < c_gpuNumClusterPerCell; si++)
2500 for (int i = 0; i < c_nbnxnGpuClusterSize; i += GMX_SIMD4_WIDTH)
2502 int io = si*c_nbnxnGpuClusterSize + i;
2503 int ia = ci*c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize + io;
2504 for (int j = 0; j < GMX_SIMD4_WIDTH; j++)
2506 x_ci[io*DIM + j + XX*GMX_SIMD4_WIDTH] = x[(ia + j)*stride + XX] + shx;
2507 x_ci[io*DIM + j + YY*GMX_SIMD4_WIDTH] = x[(ia + j)*stride + YY] + shy;
2508 x_ci[io*DIM + j + ZZ*GMX_SIMD4_WIDTH] = x[(ia + j)*stride + ZZ] + shz;
2513 #endif /* !GMX_SIMD4_HAVE_REAL */
2516 static real minimum_subgrid_size_xy(const Grid &grid)
2518 const Grid::Dimensions &dims = grid.dimensions();
2520 if (grid.geometry().isSimple)
2522 return std::min(dims.cellSize[XX], dims.cellSize[YY]);
2526 return std::min(dims.cellSize[XX]/c_gpuNumClusterPerCellX,
2527 dims.cellSize[YY]/c_gpuNumClusterPerCellY);
2531 static real effective_buffer_1x1_vs_MxN(const Grid &iGrid,
2534 const real eff_1x1_buffer_fac_overest = 0.1;
2536 /* Determine an atom-pair list cut-off buffer size for atom pairs,
2537 * to be added to rlist (including buffer) used for MxN.
2538 * This is for converting an MxN list to a 1x1 list. This means we can't
2539 * use the normal buffer estimate, as we have an MxN list in which
2540 * some atom pairs beyond rlist are missing. We want to capture
2541 * the beneficial effect of buffering by extra pairs just outside rlist,
2542 * while removing the useless pairs that are further away from rlist.
2543 * (Also the buffer could have been set manually not using the estimate.)
2544 * This buffer size is an overestimate.
2545 * We add 10% of the smallest grid sub-cell dimensions.
2546 * Note that the z-size differs per cell and we don't use this,
2547 * so we overestimate.
2548 * With PME, the 10% value gives a buffer that is somewhat larger
2549 * than the effective buffer with a tolerance of 0.005 kJ/mol/ps.
2550 * Smaller tolerances or using RF lead to a smaller effective buffer,
2551 * so 10% gives a safe overestimate.
2553 return eff_1x1_buffer_fac_overest*(minimum_subgrid_size_xy(iGrid) +
2554 minimum_subgrid_size_xy(jGrid));
2557 /* Estimates the interaction volume^2 for non-local interactions */
2558 static real nonlocal_vol2(const struct gmx_domdec_zones_t *zones, const rvec ls, real r)
2566 /* Here we simply add up the volumes of 1, 2 or 3 1D decomposition
2567 * not home interaction volume^2. As these volumes are not additive,
2568 * this is an overestimate, but it would only be significant in the limit
2569 * of small cells, where we anyhow need to split the lists into
2570 * as small parts as possible.
2573 for (int z = 0; z < zones->n; z++)
2575 if (zones->shift[z][XX] + zones->shift[z][YY] + zones->shift[z][ZZ] == 1)
2580 for (int d = 0; d < DIM; d++)
2582 if (zones->shift[z][d] == 0)
2586 za *= zones->size[z].x1[d] - zones->size[z].x0[d];
2590 /* 4 octants of a sphere */
2591 vold_est = 0.25*M_PI*r*r*r*r;
2592 /* 4 quarter pie slices on the edges */
2593 vold_est += 4*cl*M_PI/6.0*r*r*r;
2594 /* One rectangular volume on a face */
2595 vold_est += ca*0.5*r*r;
2597 vol2_est_tot += vold_est*za;
2601 return vol2_est_tot;
2604 /* Estimates the average size of a full j-list for super/sub setup */
2605 static void get_nsubpair_target(const nbnxn_search *nbs,
2606 const InteractionLocality iloc,
2608 const int min_ci_balanced,
2609 int *nsubpair_target,
2610 float *nsubpair_tot_est)
2612 /* The target value of 36 seems to be the optimum for Kepler.
2613 * Maxwell is less sensitive to the exact value.
2615 const int nsubpair_target_min = 36;
2616 real r_eff_sup, vol_est, nsp_est, nsp_est_nl;
2618 const Grid &grid = nbs->grid[0];
2620 /* We don't need to balance list sizes if:
2621 * - We didn't request balancing.
2622 * - The number of grid cells >= the number of lists requested,
2623 * since we will always generate at least #cells lists.
2624 * - We don't have any cells, since then there won't be any lists.
2626 if (min_ci_balanced <= 0 || grid.numCells() >= min_ci_balanced || grid.numCells() == 0)
2628 /* nsubpair_target==0 signals no balancing */
2629 *nsubpair_target = 0;
2630 *nsubpair_tot_est = 0;
2636 const int numAtomsCluster = grid.geometry().numAtomsICluster;
2637 const Grid::Dimensions &dims = grid.dimensions();
2639 ls[XX] = dims.cellSize[XX]/c_gpuNumClusterPerCellX;
2640 ls[YY] = dims.cellSize[YY]/c_gpuNumClusterPerCellY;
2641 ls[ZZ] = numAtomsCluster/(dims.atomDensity*ls[XX]*ls[YY]);
2643 /* The formulas below are a heuristic estimate of the average nsj per si*/
2644 r_eff_sup = rlist + nbnxn_get_rlist_effective_inc(numAtomsCluster, ls);
2646 if (!nbs->DomDec || nbs->zones->n == 1)
2653 gmx::square(dims.atomDensity/numAtomsCluster)*
2654 nonlocal_vol2(nbs->zones, ls, r_eff_sup);
2657 if (iloc == InteractionLocality::Local)
2659 /* Sub-cell interacts with itself */
2660 vol_est = ls[XX]*ls[YY]*ls[ZZ];
2661 /* 6/2 rectangular volume on the faces */
2662 vol_est += (ls[XX]*ls[YY] + ls[XX]*ls[ZZ] + ls[YY]*ls[ZZ])*r_eff_sup;
2663 /* 12/2 quarter pie slices on the edges */
2664 vol_est += 2*(ls[XX] + ls[YY] + ls[ZZ])*0.25*M_PI*gmx::square(r_eff_sup);
2665 /* 4 octants of a sphere */
2666 vol_est += 0.5*4.0/3.0*M_PI*gmx::power3(r_eff_sup);
2668 /* Estimate the number of cluster pairs as the local number of
2669 * clusters times the volume they interact with times the density.
2671 nsp_est = grid.numClusters()*vol_est*dims.atomDensity/numAtomsCluster;
2673 /* Subtract the non-local pair count */
2674 nsp_est -= nsp_est_nl;
2676 /* For small cut-offs nsp_est will be an underesimate.
2677 * With DD nsp_est_nl is an overestimate so nsp_est can get negative.
2678 * So to avoid too small or negative nsp_est we set a minimum of
2679 * all cells interacting with all 3^3 direct neighbors (3^3-1)/2+1=14.
2680 * This might be a slight overestimate for small non-periodic groups of
2681 * atoms as will occur for a local domain with DD, but for small
2682 * groups of atoms we'll anyhow be limited by nsubpair_target_min,
2683 * so this overestimation will not matter.
2685 nsp_est = std::max(nsp_est, grid.numClusters()*14._real);
2689 fprintf(debug, "nsp_est local %5.1f non-local %5.1f\n",
2690 nsp_est, nsp_est_nl);
2695 nsp_est = nsp_est_nl;
2698 /* Thus the (average) maximum j-list size should be as follows.
2699 * Since there is overhead, we shouldn't make the lists too small
2700 * (and we can't chop up j-groups) so we use a minimum target size of 36.
2702 *nsubpair_target = std::max(nsubpair_target_min,
2703 roundToInt(nsp_est/min_ci_balanced));
2704 *nsubpair_tot_est = static_cast<int>(nsp_est);
2708 fprintf(debug, "nbl nsp estimate %.1f, nsubpair_target %d\n",
2709 nsp_est, *nsubpair_target);
2713 /* Debug list print function */
2714 static void print_nblist_ci_cj(FILE *fp, const NbnxnPairlistCpu *nbl)
2716 for (const nbnxn_ci_t &ciEntry : nbl->ci)
2718 fprintf(fp, "ci %4d shift %2d ncj %3d\n",
2719 ciEntry.ci, ciEntry.shift,
2720 ciEntry.cj_ind_end - ciEntry.cj_ind_start);
2722 for (int j = ciEntry.cj_ind_start; j < ciEntry.cj_ind_end; j++)
2724 fprintf(fp, " cj %5d imask %x\n",
2731 /* Debug list print function */
2732 static void print_nblist_sci_cj(FILE *fp, const NbnxnPairlistGpu *nbl)
2734 for (const nbnxn_sci_t &sci : nbl->sci)
2736 fprintf(fp, "ci %4d shift %2d ncj4 %2d\n",
2738 sci.numJClusterGroups());
2741 for (int j4 = sci.cj4_ind_start; j4 < sci.cj4_ind_end; j4++)
2743 for (int j = 0; j < c_nbnxnGpuJgroupSize; j++)
2745 fprintf(fp, " sj %5d imask %x\n",
2747 nbl->cj4[j4].imei[0].imask);
2748 for (int si = 0; si < c_gpuNumClusterPerCell; si++)
2750 if (nbl->cj4[j4].imei[0].imask & (1U << (j*c_gpuNumClusterPerCell + si)))
2757 fprintf(fp, "ci %4d shift %2d ncj4 %2d ncp %3d\n",
2759 sci.numJClusterGroups(),
2764 /* Combine pair lists *nbl generated on multiple threads nblc */
2765 static void combine_nblists(int nnbl, NbnxnPairlistGpu **nbl,
2766 NbnxnPairlistGpu *nblc)
2768 int nsci = nblc->sci.size();
2769 int ncj4 = nblc->cj4.size();
2770 int nexcl = nblc->excl.size();
2771 for (int i = 0; i < nnbl; i++)
2773 nsci += nbl[i]->sci.size();
2774 ncj4 += nbl[i]->cj4.size();
2775 nexcl += nbl[i]->excl.size();
2778 /* Resize with the final, combined size, so we can fill in parallel */
2779 /* NOTE: For better performance we should use default initialization */
2780 nblc->sci.resize(nsci);
2781 nblc->cj4.resize(ncj4);
2782 nblc->excl.resize(nexcl);
2784 /* Each thread should copy its own data to the combined arrays,
2785 * as otherwise data will go back and forth between different caches.
2787 #if GMX_OPENMP && !(defined __clang_analyzer__)
2788 int nthreads = gmx_omp_nthreads_get(emntPairsearch);
2791 #pragma omp parallel for num_threads(nthreads) schedule(static)
2792 for (int n = 0; n < nnbl; n++)
2796 /* Determine the offset in the combined data for our thread.
2797 * Note that the original sizes in nblc are lost.
2799 int sci_offset = nsci;
2800 int cj4_offset = ncj4;
2801 int excl_offset = nexcl;
2803 for (int i = n; i < nnbl; i++)
2805 sci_offset -= nbl[i]->sci.size();
2806 cj4_offset -= nbl[i]->cj4.size();
2807 excl_offset -= nbl[i]->excl.size();
2810 const NbnxnPairlistGpu &nbli = *nbl[n];
2812 for (size_t i = 0; i < nbli.sci.size(); i++)
2814 nblc->sci[sci_offset + i] = nbli.sci[i];
2815 nblc->sci[sci_offset + i].cj4_ind_start += cj4_offset;
2816 nblc->sci[sci_offset + i].cj4_ind_end += cj4_offset;
2819 for (size_t j4 = 0; j4 < nbli.cj4.size(); j4++)
2821 nblc->cj4[cj4_offset + j4] = nbli.cj4[j4];
2822 nblc->cj4[cj4_offset + j4].imei[0].excl_ind += excl_offset;
2823 nblc->cj4[cj4_offset + j4].imei[1].excl_ind += excl_offset;
2826 for (size_t j4 = 0; j4 < nbli.excl.size(); j4++)
2828 nblc->excl[excl_offset + j4] = nbli.excl[j4];
2831 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
2834 for (int n = 0; n < nnbl; n++)
2836 nblc->nci_tot += nbl[n]->nci_tot;
2840 static void balance_fep_lists(const nbnxn_search *nbs,
2841 nbnxn_pairlist_set_t *nbl_lists)
2844 int nri_tot, nrj_tot, nrj_target;
2848 nnbl = nbl_lists->nnbl;
2852 /* Nothing to balance */
2856 /* Count the total i-lists and pairs */
2859 for (int th = 0; th < nnbl; th++)
2861 nri_tot += nbl_lists->nbl_fep[th]->nri;
2862 nrj_tot += nbl_lists->nbl_fep[th]->nrj;
2865 nrj_target = (nrj_tot + nnbl - 1)/nnbl;
2867 assert(gmx_omp_nthreads_get(emntNonbonded) == nnbl);
2869 #pragma omp parallel for schedule(static) num_threads(nnbl)
2870 for (int th = 0; th < nnbl; th++)
2874 t_nblist *nbl = nbs->work[th].nbl_fep.get();
2876 /* Note that here we allocate for the total size, instead of
2877 * a per-thread esimate (which is hard to obtain).
2879 if (nri_tot > nbl->maxnri)
2881 nbl->maxnri = over_alloc_large(nri_tot);
2882 reallocate_nblist(nbl);
2884 if (nri_tot > nbl->maxnri || nrj_tot > nbl->maxnrj)
2886 nbl->maxnrj = over_alloc_small(nrj_tot);
2887 srenew(nbl->jjnr, nbl->maxnrj);
2888 srenew(nbl->excl_fep, nbl->maxnrj);
2891 clear_pairlist_fep(nbl);
2893 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
2896 /* Loop over the source lists and assign and copy i-entries */
2898 nbld = nbs->work[th_dest].nbl_fep.get();
2899 for (int th = 0; th < nnbl; th++)
2903 nbls = nbl_lists->nbl_fep[th];
2905 for (int i = 0; i < nbls->nri; i++)
2909 /* The number of pairs in this i-entry */
2910 nrj = nbls->jindex[i+1] - nbls->jindex[i];
2912 /* Decide if list th_dest is too large and we should procede
2913 * to the next destination list.
2915 if (th_dest+1 < nnbl && nbld->nrj > 0 &&
2916 nbld->nrj + nrj - nrj_target > nrj_target - nbld->nrj)
2919 nbld = nbs->work[th_dest].nbl_fep.get();
2922 nbld->iinr[nbld->nri] = nbls->iinr[i];
2923 nbld->gid[nbld->nri] = nbls->gid[i];
2924 nbld->shift[nbld->nri] = nbls->shift[i];
2926 for (int j = nbls->jindex[i]; j < nbls->jindex[i+1]; j++)
2928 nbld->jjnr[nbld->nrj] = nbls->jjnr[j];
2929 nbld->excl_fep[nbld->nrj] = nbls->excl_fep[j];
2933 nbld->jindex[nbld->nri] = nbld->nrj;
2937 /* Swap the list pointers */
2938 for (int th = 0; th < nnbl; th++)
2940 t_nblist *nbl_tmp = nbs->work[th].nbl_fep.release();
2941 nbs->work[th].nbl_fep.reset(nbl_lists->nbl_fep[th]);
2942 nbl_lists->nbl_fep[th] = nbl_tmp;
2946 fprintf(debug, "nbl_fep[%d] nri %4d nrj %4d\n",
2948 nbl_lists->nbl_fep[th]->nri,
2949 nbl_lists->nbl_fep[th]->nrj);
2954 /* Returns the next ci to be processes by our thread */
2955 static gmx_bool next_ci(const Grid &grid,
2956 int nth, int ci_block,
2957 int *ci_x, int *ci_y,
2963 if (*ci_b == ci_block)
2965 /* Jump to the next block assigned to this task */
2966 *ci += (nth - 1)*ci_block;
2970 if (*ci >= grid.numCells())
2975 while (*ci >= grid.firstCellInColumn(*ci_x*grid.dimensions().numCells[YY] + *ci_y + 1))
2978 if (*ci_y == grid.dimensions().numCells[YY])
2988 /* Returns the distance^2 for which we put cell pairs in the list
2989 * without checking atom pair distances. This is usually < rlist^2.
2991 static float boundingbox_only_distance2(const Grid::Dimensions &iGridDims,
2992 const Grid::Dimensions &jGridDims,
2996 /* If the distance between two sub-cell bounding boxes is less
2997 * than this distance, do not check the distance between
2998 * all particle pairs in the sub-cell, since then it is likely
2999 * that the box pair has atom pairs within the cut-off.
3000 * We use the nblist cut-off minus 0.5 times the average x/y diagonal
3001 * spacing of the sub-cells. Around 40% of the checked pairs are pruned.
3002 * Using more than 0.5 gains at most 0.5%.
3003 * If forces are calculated more than twice, the performance gain
3004 * in the force calculation outweighs the cost of checking.
3005 * Note that with subcell lists, the atom-pair distance check
3006 * is only performed when only 1 out of 8 sub-cells in within range,
3007 * this is because the GPU is much faster than the cpu.
3012 bbx = 0.5*(iGridDims.cellSize[XX] + jGridDims.cellSize[XX]);
3013 bby = 0.5*(iGridDims.cellSize[YY] + jGridDims.cellSize[YY]);
3016 bbx /= c_gpuNumClusterPerCellX;
3017 bby /= c_gpuNumClusterPerCellY;
3020 rbb2 = std::max(0.0, rlist - 0.5*std::sqrt(bbx*bbx + bby*bby));
3026 return (float)((1+GMX_FLOAT_EPS)*rbb2);
3030 static int get_ci_block_size(const Grid &iGrid,
3031 gmx_bool bDomDec, int nth)
3033 const int ci_block_enum = 5;
3034 const int ci_block_denom = 11;
3035 const int ci_block_min_atoms = 16;
3038 /* Here we decide how to distribute the blocks over the threads.
3039 * We use prime numbers to try to avoid that the grid size becomes
3040 * a multiple of the number of threads, which would lead to some
3041 * threads getting "inner" pairs and others getting boundary pairs,
3042 * which in turns will lead to load imbalance between threads.
3043 * Set the block size as 5/11/ntask times the average number of cells
3044 * in a y,z slab. This should ensure a quite uniform distribution
3045 * of the grid parts of the different thread along all three grid
3046 * zone boundaries with 3D domain decomposition. At the same time
3047 * the blocks will not become too small.
3049 ci_block = (iGrid.numCells()*ci_block_enum)/(ci_block_denom*iGrid.dimensions().numCells[XX]*nth);
3051 const int numAtomsPerCell = iGrid.geometry().numAtomsPerCell;
3053 /* Ensure the blocks are not too small: avoids cache invalidation */
3054 if (ci_block*numAtomsPerCell < ci_block_min_atoms)
3056 ci_block = (ci_block_min_atoms + numAtomsPerCell - 1)/numAtomsPerCell;
3059 /* Without domain decomposition
3060 * or with less than 3 blocks per task, divide in nth blocks.
3062 if (!bDomDec || nth*3*ci_block > iGrid.numCells())
3064 ci_block = (iGrid.numCells() + nth - 1)/nth;
3067 if (ci_block > 1 && (nth - 1)*ci_block >= iGrid.numCells())
3069 /* Some threads have no work. Although reducing the block size
3070 * does not decrease the block count on the first few threads,
3071 * with GPUs better mixing of "upper" cells that have more empty
3072 * clusters results in a somewhat lower max load over all threads.
3073 * Without GPUs the regime of so few atoms per thread is less
3074 * performance relevant, but with 8-wide SIMD the same reasoning
3075 * applies, since the pair list uses 4 i-atom "sub-clusters".
3083 /* Returns the number of bits to right-shift a cluster index to obtain
3084 * the corresponding force buffer flag index.
3086 static int getBufferFlagShift(int numAtomsPerCluster)
3088 int bufferFlagShift = 0;
3089 while ((numAtomsPerCluster << bufferFlagShift) < NBNXN_BUFFERFLAG_SIZE)
3094 return bufferFlagShift;
3097 static bool pairlistIsSimple(const NbnxnPairlistCpu gmx_unused &pairlist)
3102 static bool pairlistIsSimple(const NbnxnPairlistGpu gmx_unused &pairlist)
3107 static void makeClusterListWrapper(NbnxnPairlistCpu *nbl,
3108 const Grid gmx_unused &iGrid,
3111 const int firstCell,
3113 const bool excludeSubDiagonal,
3114 const nbnxn_atomdata_t *nbat,
3117 const Nbnxm::KernelType kernelType,
3118 int *numDistanceChecks)
3122 case Nbnxm::KernelType::Cpu4x4_PlainC:
3123 makeClusterListSimple(jGrid,
3124 nbl, ci, firstCell, lastCell,
3130 #ifdef GMX_NBNXN_SIMD_4XN
3131 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
3132 makeClusterListSimd4xn(jGrid,
3133 nbl, ci, firstCell, lastCell,
3140 #ifdef GMX_NBNXN_SIMD_2XNN
3141 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
3142 makeClusterListSimd2xnn(jGrid,
3143 nbl, ci, firstCell, lastCell,
3151 GMX_ASSERT(false, "Unhandled kernel type");
3155 static void makeClusterListWrapper(NbnxnPairlistGpu *nbl,
3156 const Grid &gmx_unused iGrid,
3159 const int firstCell,
3161 const bool excludeSubDiagonal,
3162 const nbnxn_atomdata_t *nbat,
3165 Nbnxm::KernelType gmx_unused kernelType,
3166 int *numDistanceChecks)
3168 for (int cj = firstCell; cj <= lastCell; cj++)
3170 make_cluster_list_supersub(iGrid, jGrid,
3173 nbat->xstride, nbat->x().data(),
3179 static int getNumSimpleJClustersInList(const NbnxnPairlistCpu &nbl)
3181 return nbl.cj.size();
3184 static int getNumSimpleJClustersInList(const gmx_unused NbnxnPairlistGpu &nbl)
3189 static void incrementNumSimpleJClustersInList(NbnxnPairlistCpu *nbl,
3192 nbl->ncjInUse += nbl->cj.size() - ncj_old_j;
3195 static void incrementNumSimpleJClustersInList(NbnxnPairlistGpu gmx_unused *nbl,
3196 int gmx_unused ncj_old_j)
3200 static void checkListSizeConsistency(const NbnxnPairlistCpu &nbl,
3201 const bool haveFreeEnergy)
3203 GMX_RELEASE_ASSERT(static_cast<size_t>(nbl.ncjInUse) == nbl.cj.size() || haveFreeEnergy,
3204 "Without free-energy all cj pair-list entries should be in use. "
3205 "Note that subsequent code does not make use of the equality, "
3206 "this check is only here to catch bugs");
3209 static void checkListSizeConsistency(const NbnxnPairlistGpu gmx_unused &nbl,
3210 bool gmx_unused haveFreeEnergy)
3212 /* We currently can not check consistency here */
3215 /* Set the buffer flags for newly added entries in the list */
3216 static void setBufferFlags(const NbnxnPairlistCpu &nbl,
3217 const int ncj_old_j,
3218 const int gridj_flag_shift,
3219 gmx_bitmask_t *gridj_flag,
3222 if (gmx::ssize(nbl.cj) > ncj_old_j)
3224 int cbFirst = nbl.cj[ncj_old_j].cj >> gridj_flag_shift;
3225 int cbLast = nbl.cj.back().cj >> gridj_flag_shift;
3226 for (int cb = cbFirst; cb <= cbLast; cb++)
3228 bitmask_init_bit(&gridj_flag[cb], th);
3233 static void setBufferFlags(const NbnxnPairlistGpu gmx_unused &nbl,
3234 int gmx_unused ncj_old_j,
3235 int gmx_unused gridj_flag_shift,
3236 gmx_bitmask_t gmx_unused *gridj_flag,
3239 GMX_ASSERT(false, "This function should never be called");
3242 /* Generates the part of pair-list nbl assigned to our thread */
3243 template <typename T>
3244 static void nbnxn_make_pairlist_part(const nbnxn_search *nbs,
3247 nbnxn_search_work_t *work,
3248 const nbnxn_atomdata_t *nbat,
3249 const t_blocka &exclusions,
3251 const Nbnxm::KernelType kernelType,
3253 gmx_bool bFBufferFlag,
3256 float nsubpair_tot_est,
3263 real rlist2, rl_fep2 = 0;
3265 int ci_b, ci, ci_x, ci_y, ci_xy;
3267 real bx0, bx1, by0, by1, bz0, bz1;
3269 real d2cx, d2z, d2z_cx, d2z_cy, d2zx, d2zxy, d2xy;
3270 int cxf, cxl, cyf, cyf_x, cyl;
3271 int numDistanceChecks;
3272 int gridi_flag_shift = 0, gridj_flag_shift = 0;
3273 gmx_bitmask_t *gridj_flag = nullptr;
3274 int ncj_old_i, ncj_old_j;
3276 nbs_cycle_start(&work->cc[enbsCCsearch]);
3278 if (jGrid.geometry().isSimple != pairlistIsSimple(*nbl) ||
3279 iGrid.geometry().isSimple != pairlistIsSimple(*nbl))
3281 gmx_incons("Grid incompatible with pair-list");
3285 GMX_ASSERT(nbl->na_ci == jGrid.geometry().numAtomsICluster,
3286 "The cluster sizes in the list and grid should match");
3287 nbl->na_cj = Nbnxm::JClusterSizePerKernelType[kernelType];
3288 na_cj_2log = get_2log(nbl->na_cj);
3294 /* Determine conversion of clusters to flag blocks */
3295 gridi_flag_shift = getBufferFlagShift(nbl->na_ci);
3296 gridj_flag_shift = getBufferFlagShift(nbl->na_cj);
3298 gridj_flag = work->buffer_flags.flag;
3301 copy_mat(nbs->box, box);
3303 rlist2 = nbl->rlist*nbl->rlist;
3305 if (nbs->bFEP && !pairlistIsSimple(*nbl))
3307 /* Determine an atom-pair list cut-off distance for FEP atom pairs.
3308 * We should not simply use rlist, since then we would not have
3309 * the small, effective buffering of the NxN lists.
3310 * The buffer is on overestimate, but the resulting cost for pairs
3311 * beyond rlist is neglible compared to the FEP pairs within rlist.
3313 rl_fep2 = nbl->rlist + effective_buffer_1x1_vs_MxN(iGrid, jGrid);
3317 fprintf(debug, "nbl_fep atom-pair rlist %f\n", rl_fep2);
3319 rl_fep2 = rl_fep2*rl_fep2;
3322 const Grid::Dimensions &iGridDims = iGrid.dimensions();
3323 const Grid::Dimensions &jGridDims = jGrid.dimensions();
3325 rbb2 = boundingbox_only_distance2(iGridDims, jGridDims, nbl->rlist, pairlistIsSimple(*nbl));
3329 fprintf(debug, "nbl bounding box only distance %f\n", std::sqrt(rbb2));
3332 const bool isIntraGridList = (&iGrid == &jGrid);
3334 /* Set the shift range */
3335 for (int d = 0; d < DIM; d++)
3337 /* Check if we need periodicity shifts.
3338 * Without PBC or with domain decomposition we don't need them.
3340 if (d >= ePBC2npbcdim(nbs->ePBC) || nbs->dd_dim[d])
3346 const real listRangeCellToCell =
3347 listRangeForGridCellToGridCell(rlist, iGrid.dimensions(), jGrid.dimensions());
3349 box[XX][XX] - fabs(box[YY][XX]) - fabs(box[ZZ][XX]) < listRangeCellToCell)
3359 const bool bSimple = pairlistIsSimple(*nbl);
3360 gmx::ArrayRef<const nbnxn_bb_t> bb_i;
3362 gmx::ArrayRef<const float> pbb_i;
3365 bb_i = iGrid.iBoundingBoxes();
3369 pbb_i = iGrid.packedBoundingBoxes();
3372 /* We use the normal bounding box format for both grid types */
3373 bb_i = iGrid.iBoundingBoxes();
3375 gmx::ArrayRef<const BoundingBox1D> bbcz_i = iGrid.zBoundingBoxes();
3376 gmx::ArrayRef<const int> flags_i = iGrid.clusterFlags();
3377 gmx::ArrayRef<const BoundingBox1D> bbcz_j = jGrid.zBoundingBoxes();
3378 int cell0_i = iGrid.cellOffset();
3382 fprintf(debug, "nbl nc_i %d col.av. %.1f ci_block %d\n",
3383 iGrid.numCells(), iGrid.numCells()/static_cast<double>(iGrid.numColumns()), ci_block);
3386 numDistanceChecks = 0;
3388 const real listRangeBBToJCell2 = gmx::square(listRangeForBoundingBoxToGridCell(rlist, jGrid.dimensions()));
3390 /* Initially ci_b and ci to 1 before where we want them to start,
3391 * as they will both be incremented in next_ci.
3394 ci = th*ci_block - 1;
3397 while (next_ci(iGrid, nth, ci_block, &ci_x, &ci_y, &ci_b, &ci))
3399 if (bSimple && flags_i[ci] == 0)
3404 ncj_old_i = getNumSimpleJClustersInList(*nbl);
3407 if (!isIntraGridList && shp[XX] == 0)
3411 bx1 = bb_i[ci].upper.x;
3415 bx1 = iGridDims.lowerCorner[XX] + (ci_x+1)*iGridDims.cellSize[XX];
3417 if (bx1 < jGridDims.lowerCorner[XX])
3419 d2cx = gmx::square(jGridDims.lowerCorner[XX] - bx1);
3421 if (d2cx >= listRangeBBToJCell2)
3428 ci_xy = ci_x*iGridDims.numCells[YY] + ci_y;
3430 /* Loop over shift vectors in three dimensions */
3431 for (int tz = -shp[ZZ]; tz <= shp[ZZ]; tz++)
3433 const real shz = tz*box[ZZ][ZZ];
3435 bz0 = bbcz_i[ci].lower + shz;
3436 bz1 = bbcz_i[ci].upper + shz;
3444 d2z = gmx::square(bz1);
3448 d2z = gmx::square(bz0 - box[ZZ][ZZ]);
3451 d2z_cx = d2z + d2cx;
3453 if (d2z_cx >= rlist2)
3458 bz1_frac = bz1/iGrid.numCellsInColumn(ci_xy);
3463 /* The check with bz1_frac close to or larger than 1 comes later */
3465 for (int ty = -shp[YY]; ty <= shp[YY]; ty++)
3467 const real shy = ty*box[YY][YY] + tz*box[ZZ][YY];
3471 by0 = bb_i[ci].lower.y + shy;
3472 by1 = bb_i[ci].upper.y + shy;
3476 by0 = iGridDims.lowerCorner[YY] + (ci_y )*iGridDims.cellSize[YY] + shy;
3477 by1 = iGridDims.lowerCorner[YY] + (ci_y + 1)*iGridDims.cellSize[YY] + shy;
3480 get_cell_range<YY>(by0, by1,
3491 if (by1 < jGridDims.lowerCorner[YY])
3493 d2z_cy += gmx::square(jGridDims.lowerCorner[YY] - by1);
3495 else if (by0 > jGridDims.upperCorner[YY])
3497 d2z_cy += gmx::square(by0 - jGridDims.upperCorner[YY]);
3500 for (int tx = -shp[XX]; tx <= shp[XX]; tx++)
3502 const int shift = XYZ2IS(tx, ty, tz);
3504 const bool excludeSubDiagonal = (isIntraGridList && shift == CENTRAL);
3506 if (c_pbcShiftBackward && isIntraGridList && shift > CENTRAL)
3511 const real shx = tx*box[XX][XX] + ty*box[YY][XX] + tz*box[ZZ][XX];
3515 bx0 = bb_i[ci].lower.x + shx;
3516 bx1 = bb_i[ci].upper.x + shx;
3520 bx0 = iGridDims.lowerCorner[XX] + (ci_x )*iGridDims.cellSize[XX] + shx;
3521 bx1 = iGridDims.lowerCorner[XX] + (ci_x+1)*iGridDims.cellSize[XX] + shx;
3524 get_cell_range<XX>(bx0, bx1,
3534 addNewIEntry(nbl, cell0_i+ci, shift, flags_i[ci]);
3536 if ((!c_pbcShiftBackward || excludeSubDiagonal) &&
3539 /* Leave the pairs with i > j.
3540 * x is the major index, so skip half of it.
3545 set_icell_bb(iGrid, ci, shx, shy, shz,
3548 icell_set_x(cell0_i+ci, shx, shy, shz,
3549 nbat->xstride, nbat->x().data(),
3553 for (int cx = cxf; cx <= cxl; cx++)
3556 if (jGridDims.lowerCorner[XX] + cx*jGridDims.cellSize[XX] > bx1)
3558 d2zx += gmx::square(jGridDims.lowerCorner[XX] + cx*jGridDims.cellSize[XX] - bx1);
3560 else if (jGridDims.lowerCorner[XX] + (cx+1)*jGridDims.cellSize[XX] < bx0)
3562 d2zx += gmx::square(jGridDims.lowerCorner[XX] + (cx+1)*jGridDims.cellSize[XX] - bx0);
3565 if (isIntraGridList &&
3567 (!c_pbcShiftBackward || shift == CENTRAL) &&
3570 /* Leave the pairs with i > j.
3571 * Skip half of y when i and j have the same x.
3580 for (int cy = cyf_x; cy <= cyl; cy++)
3582 const int columnStart = jGrid.firstCellInColumn(cx*jGridDims.numCells[YY] + cy);
3583 const int columnEnd = jGrid.firstCellInColumn(cx*jGridDims.numCells[YY] + cy + 1);
3586 if (jGridDims.lowerCorner[YY] + cy*jGridDims.cellSize[YY] > by1)
3588 d2zxy += gmx::square(jGridDims.lowerCorner[YY] + cy*jGridDims.cellSize[YY] - by1);
3590 else if (jGridDims.lowerCorner[YY] + (cy + 1)*jGridDims.cellSize[YY] < by0)
3592 d2zxy += gmx::square(jGridDims.lowerCorner[YY] + (cy + 1)*jGridDims.cellSize[YY] - by0);
3594 if (columnStart < columnEnd && d2zxy < listRangeBBToJCell2)
3596 /* To improve efficiency in the common case
3597 * of a homogeneous particle distribution,
3598 * we estimate the index of the middle cell
3599 * in range (midCell). We search down and up
3600 * starting from this index.
3602 * Note that the bbcz_j array contains bounds
3603 * for i-clusters, thus for clusters of 4 atoms.
3604 * For the common case where the j-cluster size
3605 * is 8, we could step with a stride of 2,
3606 * but we do not do this because it would
3607 * complicate this code even more.
3609 int midCell = columnStart + static_cast<int>(bz1_frac*(columnEnd - columnStart));
3610 if (midCell >= columnEnd)
3612 midCell = columnEnd - 1;
3617 /* Find the lowest cell that can possibly
3619 * Check if we hit the bottom of the grid,
3620 * if the j-cell is below the i-cell and if so,
3621 * if it is within range.
3623 int downTestCell = midCell;
3624 while (downTestCell >= columnStart &&
3625 (bbcz_j[downTestCell].upper >= bz0 ||
3626 d2xy + gmx::square(bbcz_j[downTestCell].upper - bz0) < rlist2))
3630 int firstCell = downTestCell + 1;
3632 /* Find the highest cell that can possibly
3634 * Check if we hit the top of the grid,
3635 * if the j-cell is above the i-cell and if so,
3636 * if it is within range.
3638 int upTestCell = midCell + 1;
3639 while (upTestCell < columnEnd &&
3640 (bbcz_j[upTestCell].lower <= bz1 ||
3641 d2xy + gmx::square(bbcz_j[upTestCell].lower - bz1) < rlist2))
3645 int lastCell = upTestCell - 1;
3647 #define NBNXN_REFCODE 0
3650 /* Simple reference code, for debugging,
3651 * overrides the more complex code above.
3653 firstCell = columnEnd;
3655 for (int k = columnStart; k < columnEnd; k++)
3657 if (d2xy + gmx::square(bbcz_j[k*NNBSBB_D + 1] - bz0) < rlist2 &&
3662 if (d2xy + gmx::square(bbcz_j[k*NNBSBB_D] - bz1) < rlist2 &&
3671 if (isIntraGridList)
3673 /* We want each atom/cell pair only once,
3674 * only use cj >= ci.
3676 if (!c_pbcShiftBackward || shift == CENTRAL)
3678 firstCell = std::max(firstCell, ci);
3682 if (firstCell <= lastCell)
3684 GMX_ASSERT(firstCell >= columnStart && lastCell < columnEnd, "The range should reside within the current grid column");
3686 /* For f buffer flags with simple lists */
3687 ncj_old_j = getNumSimpleJClustersInList(*nbl);
3689 makeClusterListWrapper(nbl,
3691 jGrid, firstCell, lastCell,
3696 &numDistanceChecks);
3700 setBufferFlags(*nbl, ncj_old_j, gridj_flag_shift,
3704 incrementNumSimpleJClustersInList(nbl, ncj_old_j);
3710 /* Set the exclusions for this ci list */
3711 setExclusionsForIEntry(nbs,
3715 *getOpenIEntry(nbl),
3720 make_fep_list(nbs, nbat, nbl,
3725 iGrid, jGrid, nbl_fep);
3728 /* Close this ci list */
3731 progBal, nsubpair_tot_est,
3737 if (bFBufferFlag && getNumSimpleJClustersInList(*nbl) > ncj_old_i)
3739 bitmask_init_bit(&(work->buffer_flags.flag[(iGrid.cellOffset() + ci) >> gridi_flag_shift]), th);
3743 work->ndistc = numDistanceChecks;
3745 nbs_cycle_stop(&work->cc[enbsCCsearch]);
3747 checkListSizeConsistency(*nbl, nbs->bFEP);
3751 fprintf(debug, "number of distance checks %d\n", numDistanceChecks);
3753 print_nblist_statistics(debug, nbl, nbs, rlist);
3757 fprintf(debug, "nbl FEP list pairs: %d\n", nbl_fep->nrj);
3762 static void reduce_buffer_flags(const nbnxn_search *nbs,
3764 const nbnxn_buffer_flags_t *dest)
3766 for (int s = 0; s < nsrc; s++)
3768 gmx_bitmask_t * flag = nbs->work[s].buffer_flags.flag;
3770 for (int b = 0; b < dest->nflag; b++)
3772 bitmask_union(&(dest->flag[b]), flag[b]);
3777 static void print_reduction_cost(const nbnxn_buffer_flags_t *flags, int nout)
3779 int nelem, nkeep, ncopy, nred, out;
3780 gmx_bitmask_t mask_0;
3786 bitmask_init_bit(&mask_0, 0);
3787 for (int b = 0; b < flags->nflag; b++)
3789 if (bitmask_is_equal(flags->flag[b], mask_0))
3791 /* Only flag 0 is set, no copy of reduction required */
3795 else if (!bitmask_is_zero(flags->flag[b]))
3798 for (out = 0; out < nout; out++)
3800 if (bitmask_is_set(flags->flag[b], out))
3817 fprintf(debug, "nbnxn reduction: #flag %d #list %d elem %4.2f, keep %4.2f copy %4.2f red %4.2f\n",
3819 nelem/static_cast<double>(flags->nflag),
3820 nkeep/static_cast<double>(flags->nflag),
3821 ncopy/static_cast<double>(flags->nflag),
3822 nred/static_cast<double>(flags->nflag));
3825 /* Copies the list entries from src to dest when cjStart <= *cjGlobal < cjEnd.
3826 * *cjGlobal is updated with the cj count in src.
3827 * When setFlags==true, flag bit t is set in flag for all i and j clusters.
3829 template<bool setFlags>
3830 static void copySelectedListRange(const nbnxn_ci_t * gmx_restrict srcCi,
3831 const NbnxnPairlistCpu * gmx_restrict src,
3832 NbnxnPairlistCpu * gmx_restrict dest,
3833 gmx_bitmask_t *flag,
3834 int iFlagShift, int jFlagShift, int t)
3836 const int ncj = srcCi->cj_ind_end - srcCi->cj_ind_start;
3838 dest->ci.push_back(*srcCi);
3839 dest->ci.back().cj_ind_start = dest->cj.size();
3840 dest->ci.back().cj_ind_end = dest->cj.size() + ncj;
3844 bitmask_init_bit(&flag[srcCi->ci >> iFlagShift], t);
3847 for (int j = srcCi->cj_ind_start; j < srcCi->cj_ind_end; j++)
3849 dest->cj.push_back(src->cj[j]);
3853 /* NOTE: This is relatively expensive, since this
3854 * operation is done for all elements in the list,
3855 * whereas at list generation this is done only
3856 * once for each flag entry.
3858 bitmask_init_bit(&flag[src->cj[j].cj >> jFlagShift], t);
3863 /* This routine re-balances the pairlists such that all are nearly equally
3864 * sized. Only whole i-entries are moved between lists. These are moved
3865 * between the ends of the lists, such that the buffer reduction cost should
3866 * not change significantly.
3867 * Note that all original reduction flags are currently kept. This can lead
3868 * to reduction of parts of the force buffer that could be avoided. But since
3869 * the original lists are quite balanced, this will only give minor overhead.
3871 static void rebalanceSimpleLists(int numLists,
3872 NbnxnPairlistCpu * const * const srcSet,
3873 NbnxnPairlistCpu **destSet,
3874 gmx::ArrayRef<nbnxn_search_work_t> searchWork)
3877 for (int s = 0; s < numLists; s++)
3879 ncjTotal += srcSet[s]->ncjInUse;
3881 int ncjTarget = (ncjTotal + numLists - 1)/numLists;
3883 #pragma omp parallel num_threads(numLists)
3885 int t = gmx_omp_get_thread_num();
3887 int cjStart = ncjTarget* t;
3888 int cjEnd = ncjTarget*(t + 1);
3890 /* The destination pair-list for task/thread t */
3891 NbnxnPairlistCpu *dest = destSet[t];
3893 clear_pairlist(dest);
3894 dest->na_cj = srcSet[0]->na_cj;
3896 /* Note that the flags in the work struct (still) contain flags
3897 * for all entries that are present in srcSet->nbl[t].
3899 gmx_bitmask_t *flag = searchWork[t].buffer_flags.flag;
3901 int iFlagShift = getBufferFlagShift(dest->na_ci);
3902 int jFlagShift = getBufferFlagShift(dest->na_cj);
3905 for (int s = 0; s < numLists && cjGlobal < cjEnd; s++)
3907 const NbnxnPairlistCpu *src = srcSet[s];
3909 if (cjGlobal + src->ncjInUse > cjStart)
3911 for (gmx::index i = 0; i < gmx::ssize(src->ci) && cjGlobal < cjEnd; i++)
3913 const nbnxn_ci_t *srcCi = &src->ci[i];
3914 int ncj = srcCi->cj_ind_end - srcCi->cj_ind_start;
3915 if (cjGlobal >= cjStart)
3917 /* If the source list is not our own, we need to set
3918 * extra flags (the template bool parameter).
3922 copySelectedListRange
3925 flag, iFlagShift, jFlagShift, t);
3929 copySelectedListRange
3932 dest, flag, iFlagShift, jFlagShift, t);
3940 cjGlobal += src->ncjInUse;
3944 dest->ncjInUse = dest->cj.size();
3948 int ncjTotalNew = 0;
3949 for (int s = 0; s < numLists; s++)
3951 ncjTotalNew += destSet[s]->ncjInUse;
3953 GMX_RELEASE_ASSERT(ncjTotalNew == ncjTotal, "The total size of the lists before and after rebalancing should match");
3957 /* Returns if the pairlists are so imbalanced that it is worth rebalancing. */
3958 static bool checkRebalanceSimpleLists(const nbnxn_pairlist_set_t *listSet)
3960 int numLists = listSet->nnbl;
3963 for (int s = 0; s < numLists; s++)
3965 ncjMax = std::max(ncjMax, listSet->nbl[s]->ncjInUse);
3966 ncjTotal += listSet->nbl[s]->ncjInUse;
3970 fprintf(debug, "Pair-list ncjMax %d ncjTotal %d\n", ncjMax, ncjTotal);
3972 /* The rebalancing adds 3% extra time to the search. Heuristically we
3973 * determined that under common conditions the non-bonded kernel balance
3974 * improvement will outweigh this when the imbalance is more than 3%.
3975 * But this will, obviously, depend on search vs kernel time and nstlist.
3977 const real rebalanceTolerance = 1.03;
3979 return numLists*ncjMax > ncjTotal*rebalanceTolerance;
3982 /* Perform a count (linear) sort to sort the smaller lists to the end.
3983 * This avoids load imbalance on the GPU, as large lists will be
3984 * scheduled and executed first and the smaller lists later.
3985 * Load balancing between multi-processors only happens at the end
3986 * and there smaller lists lead to more effective load balancing.
3987 * The sorting is done on the cj4 count, not on the actual pair counts.
3988 * Not only does this make the sort faster, but it also results in
3989 * better load balancing than using a list sorted on exact load.
3990 * This function swaps the pointer in the pair list to avoid a copy operation.
3992 static void sort_sci(NbnxnPairlistGpu *nbl)
3994 if (nbl->cj4.size() <= nbl->sci.size())
3996 /* nsci = 0 or all sci have size 1, sorting won't change the order */
4000 NbnxnPairlistGpuWork &work = *nbl->work;
4002 /* We will distinguish differences up to double the average */
4003 const int m = (2*nbl->cj4.size())/nbl->sci.size();
4005 /* Resize work.sci_sort so we can sort into it */
4006 work.sci_sort.resize(nbl->sci.size());
4008 std::vector<int> &sort = work.sortBuffer;
4009 /* Set up m + 1 entries in sort, initialized at 0 */
4011 sort.resize(m + 1, 0);
4012 /* Count the entries of each size */
4013 for (const nbnxn_sci_t &sci : nbl->sci)
4015 int i = std::min(m, sci.numJClusterGroups());
4018 /* Calculate the offset for each count */
4021 for (int i = m - 1; i >= 0; i--)
4024 sort[i] = sort[i + 1] + s0;
4028 /* Sort entries directly into place */
4029 gmx::ArrayRef<nbnxn_sci_t> sci_sort = work.sci_sort;
4030 for (const nbnxn_sci_t &sci : nbl->sci)
4032 int i = std::min(m, sci.numJClusterGroups());
4033 sci_sort[sort[i]++] = sci;
4036 /* Swap the sci pointers so we use the new, sorted list */
4037 std::swap(nbl->sci, work.sci_sort);
4041 nonbonded_verlet_t::PairlistSets::construct(const InteractionLocality iLocality,
4043 nbnxn_atomdata_t *nbat,
4044 const t_blocka *excl,
4045 const Nbnxm::KernelType kernelType,
4049 nbnxn_pairlist_set_t *nbl_list = &pairlistSet(iLocality);
4051 const real rlist = nbl_list->params.rlistOuter;
4053 int nsubpair_target;
4054 float nsubpair_tot_est;
4057 gmx_bool CombineNBLists;
4059 int np_tot, np_noq, np_hlj, nap;
4061 nnbl = nbl_list->nnbl;
4062 CombineNBLists = nbl_list->bCombined;
4066 fprintf(debug, "ns making %d nblists\n", nnbl);
4069 nbat->bUseBufferFlags = (nbat->out.size() > 1);
4070 /* We should re-init the flags before making the first list */
4071 if (nbat->bUseBufferFlags && iLocality == InteractionLocality::Local)
4073 init_buffer_flags(&nbat->buffer_flags, nbat->numAtoms());
4077 if (iLocality == InteractionLocality::Local)
4079 /* Only zone (grid) 0 vs 0 */
4084 nzi = nbs->zones->nizone;
4087 if (!nbl_list->bSimple && minimumIlistCountForGpuBalancing_ > 0)
4089 get_nsubpair_target(nbs, iLocality, rlist, minimumIlistCountForGpuBalancing_,
4090 &nsubpair_target, &nsubpair_tot_est);
4094 nsubpair_target = 0;
4095 nsubpair_tot_est = 0;
4098 /* Clear all pair-lists */
4099 for (int th = 0; th < nnbl; th++)
4101 if (nbl_list->bSimple)
4103 clear_pairlist(nbl_list->nbl[th]);
4107 clear_pairlist(nbl_list->nblGpu[th]);
4112 clear_pairlist_fep(nbl_list->nbl_fep[th]);
4116 for (int zi = 0; zi < nzi; zi++)
4118 const Grid &iGrid = nbs->grid[zi];
4122 if (iLocality == InteractionLocality::Local)
4129 zj0 = nbs->zones->izone[zi].j0;
4130 zj1 = nbs->zones->izone[zi].j1;
4136 for (int zj = zj0; zj < zj1; zj++)
4138 const Grid &jGrid = nbs->grid[zj];
4142 fprintf(debug, "ns search grid %d vs %d\n", zi, zj);
4145 nbs_cycle_start(&nbs->cc[enbsCCsearch]);
4147 ci_block = get_ci_block_size(iGrid, nbs->DomDec, nnbl);
4149 /* With GPU: generate progressively smaller lists for
4150 * load balancing for local only or non-local with 2 zones.
4152 progBal = (iLocality == InteractionLocality::Local || nbs->zones->n <= 2);
4154 #pragma omp parallel for num_threads(nnbl) schedule(static)
4155 for (int th = 0; th < nnbl; th++)
4159 /* Re-init the thread-local work flag data before making
4160 * the first list (not an elegant conditional).
4162 if (nbat->bUseBufferFlags && ((zi == 0 && zj == 0)))
4164 init_buffer_flags(&nbs->work[th].buffer_flags, nbat->numAtoms());
4167 if (CombineNBLists && th > 0)
4169 GMX_ASSERT(!nbl_list->bSimple, "Can only combine GPU lists");
4171 clear_pairlist(nbl_list->nblGpu[th]);
4174 /* Divide the i super cell equally over the nblists */
4175 if (nbl_list->bSimple)
4177 nbnxn_make_pairlist_part(nbs, iGrid, jGrid,
4178 &nbs->work[th], nbat, *excl,
4182 nbat->bUseBufferFlags,
4184 progBal, nsubpair_tot_est,
4187 nbl_list->nbl_fep[th]);
4191 nbnxn_make_pairlist_part(nbs, iGrid, jGrid,
4192 &nbs->work[th], nbat, *excl,
4196 nbat->bUseBufferFlags,
4198 progBal, nsubpair_tot_est,
4200 nbl_list->nblGpu[th],
4201 nbl_list->nbl_fep[th]);
4204 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
4206 nbs_cycle_stop(&nbs->cc[enbsCCsearch]);
4211 for (int th = 0; th < nnbl; th++)
4213 inc_nrnb(nrnb, eNR_NBNXN_DIST2, nbs->work[th].ndistc);
4215 if (nbl_list->bSimple)
4217 NbnxnPairlistCpu *nbl = nbl_list->nbl[th];
4218 np_tot += nbl->cj.size();
4219 np_noq += nbl->work->ncj_noq;
4220 np_hlj += nbl->work->ncj_hlj;
4224 NbnxnPairlistGpu *nbl = nbl_list->nblGpu[th];
4225 /* This count ignores potential subsequent pair pruning */
4226 np_tot += nbl->nci_tot;
4229 if (nbl_list->bSimple)
4231 nap = nbl_list->nbl[0]->na_ci*nbl_list->nbl[0]->na_cj;
4235 nap = gmx::square(nbl_list->nblGpu[0]->na_ci);
4237 nbl_list->natpair_ljq = (np_tot - np_noq)*nap - np_hlj*nap/2;
4238 nbl_list->natpair_lj = np_noq*nap;
4239 nbl_list->natpair_q = np_hlj*nap/2;
4241 if (CombineNBLists && nnbl > 1)
4243 GMX_ASSERT(!nbl_list->bSimple, "Can only combine GPU lists");
4244 NbnxnPairlistGpu **nbl = nbl_list->nblGpu;
4246 nbs_cycle_start(&nbs->cc[enbsCCcombine]);
4248 combine_nblists(nnbl-1, nbl+1, nbl[0]);
4250 nbs_cycle_stop(&nbs->cc[enbsCCcombine]);
4255 if (nbl_list->bSimple)
4257 if (nnbl > 1 && checkRebalanceSimpleLists(nbl_list))
4259 rebalanceSimpleLists(nbl_list->nnbl, nbl_list->nbl, nbl_list->nbl_work, nbs->work);
4261 /* Swap the pointer of the sets of pair lists */
4262 NbnxnPairlistCpu **tmp = nbl_list->nbl;
4263 nbl_list->nbl = nbl_list->nbl_work;
4264 nbl_list->nbl_work = tmp;
4269 /* Sort the entries on size, large ones first */
4270 if (CombineNBLists || nnbl == 1)
4272 sort_sci(nbl_list->nblGpu[0]);
4276 #pragma omp parallel for num_threads(nnbl) schedule(static)
4277 for (int th = 0; th < nnbl; th++)
4281 sort_sci(nbl_list->nblGpu[th]);
4283 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
4288 if (nbat->bUseBufferFlags)
4290 reduce_buffer_flags(nbs, nbl_list->nnbl, &nbat->buffer_flags);
4295 /* Balance the free-energy lists over all the threads */
4296 balance_fep_lists(nbs, nbl_list);
4299 if (nbl_list->bSimple)
4301 /* This is a fresh list, so not pruned, stored using ci.
4302 * ciOuter is invalid at this point.
4304 GMX_ASSERT(nbl_list->nbl[0]->ciOuter.empty(), "ciOuter is invalid so it should be empty");
4307 if (iLocality == Nbnxm::InteractionLocality::Local)
4309 outerListCreationStep_ = step;
4313 GMX_RELEASE_ASSERT(outerListCreationStep_ == step,
4314 "Outer list should be created at the same step as the inner list");
4317 /* Special performance logging stuff (env.var. GMX_NBNXN_CYCLE) */
4318 if (iLocality == InteractionLocality::Local)
4320 nbs->search_count++;
4322 if (nbs->print_cycles &&
4323 (!nbs->DomDec || iLocality == InteractionLocality::NonLocal) &&
4324 nbs->search_count % 100 == 0)
4326 nbs_cycle_print(stderr, nbs);
4329 /* If we have more than one list, they either got rebalancing (CPU)
4330 * or combined (GPU), so we should dump the final result to debug.
4332 if (debug && nbl_list->nnbl > 1)
4334 if (nbl_list->bSimple)
4336 for (int t = 0; t < nbl_list->nnbl; t++)
4338 print_nblist_statistics(debug, nbl_list->nbl[t], nbs, rlist);
4343 print_nblist_statistics(debug, nbl_list->nblGpu[0], nbs, rlist);
4351 if (nbl_list->bSimple)
4353 for (int t = 0; t < nbl_list->nnbl; t++)
4355 print_nblist_ci_cj(debug, nbl_list->nbl[t]);
4360 print_nblist_sci_cj(debug, nbl_list->nblGpu[0]);
4364 if (nbat->bUseBufferFlags)
4366 print_reduction_cost(&nbat->buffer_flags, nbl_list->nnbl);
4370 if (params_.useDynamicPruning && nbl_list->bSimple)
4372 nbnxnPrepareListForDynamicPruning(nbl_list);
4377 nonbonded_verlet_t::constructPairlist(const Nbnxm::InteractionLocality iLocality,
4378 const t_blocka *excl,
4382 pairlistSets_->construct(iLocality, nbs.get(), nbat.get(), excl,
4383 kernelSetup_.kernelType,
4388 /* Launch the transfer of the pairlist to the GPU.
4390 * NOTE: The launch overhead is currently not timed separately
4392 Nbnxm::gpu_init_pairlist(gpu_nbv,
4393 pairlistSets().pairlistSet(iLocality).nblGpu[0],
4398 void nbnxnPrepareListForDynamicPruning(nbnxn_pairlist_set_t *listSet)
4400 GMX_RELEASE_ASSERT(listSet->bSimple, "Should only be called for simple lists");
4402 /* TODO: Restructure the lists so we have actual outer and inner
4403 * list objects so we can set a single pointer instead of
4404 * swapping several pointers.
4407 for (int i = 0; i < listSet->nnbl; i++)
4409 NbnxnPairlistCpu &list = *listSet->nbl[i];
4411 /* The search produced a list in ci/cj.
4412 * Swap the list pointers so we get the outer list is ciOuter,cjOuter
4413 * and we can prune that to get an inner list in ci/cj.
4415 GMX_RELEASE_ASSERT(list.ciOuter.empty() && list.cjOuter.empty(),
4416 "The outer lists should be empty before preparation");
4418 std::swap(list.ci, list.ciOuter);
4419 std::swap(list.cj, list.cjOuter);