2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016 by the GROMACS development team.
5 * Copyright (c) 2017,2018,2019,2020,2021, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
49 #include "gromacs/domdec/domdec_struct.h"
50 #include "gromacs/gmxlib/nrnb.h"
51 #include "gromacs/math/functions.h"
52 #include "gromacs/math/utilities.h"
53 #include "gromacs/math/vec.h"
54 #include "gromacs/mdlib/gmx_omp_nthreads.h"
55 #include "gromacs/mdtypes/group.h"
56 #include "gromacs/mdtypes/md_enums.h"
57 #include "gromacs/mdtypes/nblist.h"
58 #include "gromacs/nbnxm/atomdata.h"
59 #include "gromacs/nbnxm/gpu_data_mgmt.h"
60 #include "gromacs/pbcutil/ishift.h"
61 #include "gromacs/pbcutil/pbc.h"
62 #include "gromacs/simd/simd.h"
63 #include "gromacs/simd/vector_operations.h"
64 #include "gromacs/utility/exceptions.h"
65 #include "gromacs/utility/fatalerror.h"
66 #include "gromacs/utility/gmxomp.h"
67 #include "gromacs/utility/listoflists.h"
68 #include "gromacs/utility/smalloc.h"
70 #include "boundingboxes.h"
71 #include "clusterdistancekerneltype.h"
73 #include "nbnxm_geometry.h"
74 #include "nbnxm_simd.h"
75 #include "pairlistset.h"
76 #include "pairlistsets.h"
77 #include "pairlistwork.h"
78 #include "pairsearch.h"
80 using namespace gmx; // TODO: Remove when this file is moved into gmx namespace
82 using BoundingBox = Nbnxm::BoundingBox; // TODO: Remove when refactoring this file
83 using BoundingBox1D = Nbnxm::BoundingBox1D; // TODO: Remove when refactoring this file
85 using Grid = Nbnxm::Grid; // TODO: Remove when refactoring this file
87 // Convenience alias for partial Nbnxn namespace usage
88 using InteractionLocality = gmx::InteractionLocality;
90 /* We shift the i-particles backward for PBC.
91 * This leads to more conditionals than shifting forward.
92 * We do this to get more balanced pair lists.
94 constexpr bool c_pbcShiftBackward = true;
96 /* Layout for the nonbonded NxN pair lists */
97 enum class NbnxnLayout
99 NoSimd4x4, // i-cluster size 4, j-cluster size 4
100 Simd4xN, // i-cluster size 4, j-cluster size SIMD width
101 Simd2xNN, // i-cluster size 4, j-cluster size half SIMD width
102 Gpu8x8x8 // i-cluster size 8, j-cluster size 8 + super-clustering
105 #if defined(GMX_NBNXN_SIMD_4XN) || defined(GMX_NBNXN_SIMD_2XNN)
106 /* Returns the j-cluster size */
107 template<NbnxnLayout layout>
108 static constexpr int jClusterSize()
110 static_assert(layout == NbnxnLayout::NoSimd4x4 || layout == NbnxnLayout::Simd4xN
111 || layout == NbnxnLayout::Simd2xNN,
112 "Currently jClusterSize only supports CPU layouts");
114 return layout == NbnxnLayout::Simd4xN
115 ? GMX_SIMD_REAL_WIDTH
116 : (layout == NbnxnLayout::Simd2xNN ? GMX_SIMD_REAL_WIDTH / 2 : c_nbnxnCpuIClusterSize);
119 /*! \brief Returns the j-cluster index given the i-cluster index.
121 * \tparam jClusterSize The number of atoms in a j-cluster
122 * \tparam jSubClusterIndex The j-sub-cluster index (0/1), used when size(j-cluster) <
123 * size(i-cluster) \param[in] ci The i-cluster index
125 template<int jClusterSize, int jSubClusterIndex>
126 static inline int cjFromCi(int ci)
128 static_assert(jClusterSize == c_nbnxnCpuIClusterSize / 2 || jClusterSize == c_nbnxnCpuIClusterSize
129 || jClusterSize == c_nbnxnCpuIClusterSize * 2,
130 "Only j-cluster sizes 2, 4 and 8 are currently implemented");
132 static_assert(jSubClusterIndex == 0 || jSubClusterIndex == 1,
133 "Only sub-cluster indices 0 and 1 are supported");
135 if (jClusterSize == c_nbnxnCpuIClusterSize / 2)
137 if (jSubClusterIndex == 0)
143 return ((ci + 1) << 1) - 1;
146 else if (jClusterSize == c_nbnxnCpuIClusterSize)
156 /*! \brief Returns the j-cluster index given the i-cluster index.
158 * \tparam layout The pair-list layout
159 * \tparam jSubClusterIndex The j-sub-cluster index (0/1), used when size(j-cluster) <
160 * size(i-cluster) \param[in] ci The i-cluster index
162 template<NbnxnLayout layout, int jSubClusterIndex>
163 static inline int cjFromCi(int ci)
165 constexpr int clusterSize = jClusterSize<layout>();
167 return cjFromCi<clusterSize, jSubClusterIndex>(ci);
170 /* Returns the nbnxn coordinate data index given the i-cluster index */
171 template<NbnxnLayout layout>
172 static inline int xIndexFromCi(int ci)
174 constexpr int clusterSize = jClusterSize<layout>();
176 static_assert(clusterSize == c_nbnxnCpuIClusterSize / 2 || clusterSize == c_nbnxnCpuIClusterSize
177 || clusterSize == c_nbnxnCpuIClusterSize * 2,
178 "Only j-cluster sizes 2, 4 and 8 are currently implemented");
180 if (clusterSize <= c_nbnxnCpuIClusterSize)
182 /* Coordinates are stored packed in groups of 4 */
183 return ci * STRIDE_P4;
187 /* Coordinates packed in 8, i-cluster size is half the packing width */
188 return (ci >> 1) * STRIDE_P8 + (ci & 1) * (c_packX8 >> 1);
192 /* Returns the nbnxn coordinate data index given the j-cluster index */
193 template<NbnxnLayout layout>
194 static inline int xIndexFromCj(int cj)
196 constexpr int clusterSize = jClusterSize<layout>();
198 static_assert(clusterSize == c_nbnxnCpuIClusterSize / 2 || clusterSize == c_nbnxnCpuIClusterSize
199 || clusterSize == c_nbnxnCpuIClusterSize * 2,
200 "Only j-cluster sizes 2, 4 and 8 are currently implemented");
202 if (clusterSize == c_nbnxnCpuIClusterSize / 2)
204 /* Coordinates are stored packed in groups of 4 */
205 return (cj >> 1) * STRIDE_P4 + (cj & 1) * (c_packX4 >> 1);
207 else if (clusterSize == c_nbnxnCpuIClusterSize)
209 /* Coordinates are stored packed in groups of 4 */
210 return cj * STRIDE_P4;
214 /* Coordinates are stored packed in groups of 8 */
215 return cj * STRIDE_P8;
218 #endif // defined(GMX_NBNXN_SIMD_4XN) || defined(GMX_NBNXN_SIMD_2XNN)
220 static constexpr int sizeNeededForBufferFlags(const int numAtoms)
222 return (numAtoms + NBNXN_BUFFERFLAG_SIZE - 1) / NBNXN_BUFFERFLAG_SIZE;
225 // Resets current flags to 0 and adds more flags if needed.
226 static void resizeAndZeroBufferFlags(std::vector<gmx_bitmask_t>* flags, const int numAtoms)
229 flags->resize(sizeNeededForBufferFlags(numAtoms), gmx_bitmask_t{ 0 });
233 /* Returns the pair-list cutoff between a bounding box and a grid cell given an atom-to-atom pair-list cutoff
235 * Given a cutoff distance between atoms, this functions returns the cutoff
236 * distance2 between a bounding box of a group of atoms and a grid cell.
237 * Since atoms can be geometrically outside of the cell they have been
238 * assigned to (when atom groups instead of individual atoms are assigned
239 * to cells), this distance returned can be larger than the input.
241 static real listRangeForBoundingBoxToGridCell(real rlist, const Grid::Dimensions& gridDims)
243 return rlist + gridDims.maxAtomGroupRadius;
245 /* Returns the pair-list cutoff between a grid cells given an atom-to-atom pair-list cutoff
247 * Given a cutoff distance between atoms, this functions returns the cutoff
248 * distance2 between two grid cells.
249 * Since atoms can be geometrically outside of the cell they have been
250 * assigned to (when atom groups instead of individual atoms are assigned
251 * to cells), this distance returned can be larger than the input.
253 static real listRangeForGridCellToGridCell(real rlist,
254 const Grid::Dimensions& iGridDims,
255 const Grid::Dimensions& jGridDims)
257 return rlist + iGridDims.maxAtomGroupRadius + jGridDims.maxAtomGroupRadius;
260 /* Determines the cell range along one dimension that
261 * the bounding box b0 - b1 sees.
265 get_cell_range(real b0, real b1, const Grid::Dimensions& jGridDims, real d2, real rlist, int* cf, int* cl)
267 real listRangeBBToCell2 = gmx::square(listRangeForBoundingBoxToGridCell(rlist, jGridDims));
268 real distanceInCells = (b0 - jGridDims.lowerCorner[dim]) * jGridDims.invCellSize[dim];
269 *cf = std::max(static_cast<int>(distanceInCells), 0);
272 && d2 + gmx::square((b0 - jGridDims.lowerCorner[dim]) - (*cf - 1 + 1) * jGridDims.cellSize[dim])
273 < listRangeBBToCell2)
278 *cl = std::min(static_cast<int>((b1 - jGridDims.lowerCorner[dim]) * jGridDims.invCellSize[dim]),
279 jGridDims.numCells[dim] - 1);
280 while (*cl < jGridDims.numCells[dim] - 1
281 && d2 + gmx::square((*cl + 1) * jGridDims.cellSize[dim] - (b1 - jGridDims.lowerCorner[dim]))
282 < listRangeBBToCell2)
288 /* Reference code calculating the distance^2 between two bounding boxes */
290 static float box_dist2(float bx0, float bx1, float by0,
291 float by1, float bz0, float bz1,
292 const BoundingBox *bb)
295 float dl, dh, dm, dm0;
299 dl = bx0 - bb->upper.x;
300 dh = bb->lower.x - bx1;
301 dm = std::max(dl, dh);
302 dm0 = std::max(dm, 0.0f);
305 dl = by0 - bb->upper.y;
306 dh = bb->lower.y - by1;
307 dm = std::max(dl, dh);
308 dm0 = std::max(dm, 0.0f);
311 dl = bz0 - bb->upper.z;
312 dh = bb->lower.z - bz1;
313 dm = std::max(dl, dh);
314 dm0 = std::max(dm, 0.0f);
321 #if !NBNXN_SEARCH_BB_SIMD4
323 /*! \brief Plain C code calculating the distance^2 between two bounding boxes in xyz0 format
325 * \param[in] bb_i First bounding box
326 * \param[in] bb_j Second bounding box
328 static float clusterBoundingBoxDistance2(const BoundingBox& bb_i, const BoundingBox& bb_j)
330 float dl = bb_i.lower.x - bb_j.upper.x;
331 float dh = bb_j.lower.x - bb_i.upper.x;
332 float dm = std::max(dl, dh);
333 float dm0 = std::max(dm, 0.0F);
334 float d2 = dm0 * dm0;
336 dl = bb_i.lower.y - bb_j.upper.y;
337 dh = bb_j.lower.y - bb_i.upper.y;
338 dm = std::max(dl, dh);
339 dm0 = std::max(dm, 0.0F);
342 dl = bb_i.lower.z - bb_j.upper.z;
343 dh = bb_j.lower.z - bb_i.upper.z;
344 dm = std::max(dl, dh);
345 dm0 = std::max(dm, 0.0F);
351 #else /* NBNXN_SEARCH_BB_SIMD4 */
353 /*! \brief 4-wide SIMD code calculating the distance^2 between two bounding boxes in xyz0 format
355 * \param[in] bb_i First bounding box, should be aligned for 4-wide SIMD
356 * \param[in] bb_j Second bounding box, should be aligned for 4-wide SIMD
358 static float clusterBoundingBoxDistance2(const BoundingBox& bb_i, const BoundingBox& bb_j)
360 // TODO: During SIMDv2 transition only some archs use namespace (remove when done)
363 const Simd4Float bb_i_S0 = load4(bb_i.lower.ptr());
364 const Simd4Float bb_i_S1 = load4(bb_i.upper.ptr());
365 const Simd4Float bb_j_S0 = load4(bb_j.lower.ptr());
366 const Simd4Float bb_j_S1 = load4(bb_j.upper.ptr());
368 const Simd4Float dl_S = bb_i_S0 - bb_j_S1;
369 const Simd4Float dh_S = bb_j_S0 - bb_i_S1;
371 const Simd4Float dm_S = max(dl_S, dh_S);
372 const Simd4Float dm0_S = max(dm_S, simd4SetZeroF());
374 return dotProduct(dm0_S, dm0_S);
377 /* Calculate bb bounding distances of bb_i[si,...,si+3] and store them in d2 */
378 template<int boundingBoxStart>
379 static inline void gmx_simdcall clusterBoundingBoxDistance2_xxxx_simd4_inner(const float* bb_i,
381 const Simd4Float xj_l,
382 const Simd4Float yj_l,
383 const Simd4Float zj_l,
384 const Simd4Float xj_h,
385 const Simd4Float yj_h,
386 const Simd4Float zj_h)
388 constexpr int stride = c_packedBoundingBoxesDimSize;
390 const int shi = boundingBoxStart * Nbnxm::c_numBoundingBoxBounds1D * DIM;
392 const Simd4Float zero = setZero();
394 const Simd4Float xi_l = load4(bb_i + shi + 0 * stride);
395 const Simd4Float yi_l = load4(bb_i + shi + 1 * stride);
396 const Simd4Float zi_l = load4(bb_i + shi + 2 * stride);
397 const Simd4Float xi_h = load4(bb_i + shi + 3 * stride);
398 const Simd4Float yi_h = load4(bb_i + shi + 4 * stride);
399 const Simd4Float zi_h = load4(bb_i + shi + 5 * stride);
401 const Simd4Float dx_0 = xi_l - xj_h;
402 const Simd4Float dy_0 = yi_l - yj_h;
403 const Simd4Float dz_0 = zi_l - zj_h;
405 const Simd4Float dx_1 = xj_l - xi_h;
406 const Simd4Float dy_1 = yj_l - yi_h;
407 const Simd4Float dz_1 = zj_l - zi_h;
409 const Simd4Float mx = max(dx_0, dx_1);
410 const Simd4Float my = max(dy_0, dy_1);
411 const Simd4Float mz = max(dz_0, dz_1);
413 const Simd4Float m0x = max(mx, zero);
414 const Simd4Float m0y = max(my, zero);
415 const Simd4Float m0z = max(mz, zero);
417 const Simd4Float d2x = m0x * m0x;
418 const Simd4Float d2y = m0y * m0y;
419 const Simd4Float d2z = m0z * m0z;
421 const Simd4Float d2s = d2x + d2y;
422 const Simd4Float d2t = d2s + d2z;
424 store4(d2 + boundingBoxStart, d2t);
427 /* 4-wide SIMD code for nsi bb distances for bb format xxxxyyyyzzzz */
428 static void clusterBoundingBoxDistance2_xxxx_simd4(const float* bb_j, const int nsi, const float* bb_i, float* d2)
430 constexpr int stride = c_packedBoundingBoxesDimSize;
432 // TODO: During SIMDv2 transition only some archs use namespace (remove when done)
435 const Simd4Float xj_l = Simd4Float(bb_j[0 * stride]);
436 const Simd4Float yj_l = Simd4Float(bb_j[1 * stride]);
437 const Simd4Float zj_l = Simd4Float(bb_j[2 * stride]);
438 const Simd4Float xj_h = Simd4Float(bb_j[3 * stride]);
439 const Simd4Float yj_h = Simd4Float(bb_j[4 * stride]);
440 const Simd4Float zj_h = Simd4Float(bb_j[5 * stride]);
442 /* Here we "loop" over si (0,stride) from 0 to nsi with step stride.
443 * But as we know the number of iterations is 1 or 2, we unroll manually.
445 clusterBoundingBoxDistance2_xxxx_simd4_inner<0>(bb_i, d2, xj_l, yj_l, zj_l, xj_h, yj_h, zj_h);
448 clusterBoundingBoxDistance2_xxxx_simd4_inner<stride>(bb_i, d2, xj_l, yj_l, zj_l, xj_h, yj_h, zj_h);
452 #endif /* NBNXN_SEARCH_BB_SIMD4 */
455 /* Returns if any atom pair from two clusters is within distance sqrt(rlist2) */
456 static inline gmx_bool
457 clusterpair_in_range(const NbnxnPairlistGpuWork& work, int si, int csj, int stride, const real* x_j, real rlist2)
459 #if !GMX_SIMD4_HAVE_REAL
462 * All coordinates are stored as xyzxyz...
465 const real* x_i = work.iSuperClusterData.x.data();
467 for (int i = 0; i < c_nbnxnGpuClusterSize; i++)
469 int i0 = (si * c_nbnxnGpuClusterSize + i) * DIM;
470 for (int j = 0; j < c_nbnxnGpuClusterSize; j++)
472 int j0 = (csj * c_nbnxnGpuClusterSize + j) * stride;
474 real d2 = gmx::square(x_i[i0] - x_j[j0]) + gmx::square(x_i[i0 + 1] - x_j[j0 + 1])
475 + gmx::square(x_i[i0 + 2] - x_j[j0 + 2]);
486 #else /* !GMX_SIMD4_HAVE_REAL */
488 /* 4-wide SIMD version.
489 * The coordinates x_i are stored as xxxxyyyy..., x_j is stored xyzxyz...
490 * Using 8-wide AVX(2) is not faster on Intel Sandy Bridge and Haswell.
492 static_assert(c_nbnxnGpuClusterSize == 8 || c_nbnxnGpuClusterSize == 4,
493 "A cluster is hard-coded to 4/8 atoms.");
495 Simd4Real rc2_S{ rlist2 };
497 const real* x_i = work.iSuperClusterData.xSimd.data();
499 int dim_stride = c_nbnxnGpuClusterSize * DIM;
500 Simd4Real ix_S0 = load4(x_i + si * dim_stride + 0 * GMX_SIMD4_WIDTH);
501 Simd4Real iy_S0 = load4(x_i + si * dim_stride + 1 * GMX_SIMD4_WIDTH);
502 Simd4Real iz_S0 = load4(x_i + si * dim_stride + 2 * GMX_SIMD4_WIDTH);
504 Simd4Real ix_S1, iy_S1, iz_S1;
505 if (c_nbnxnGpuClusterSize == 8)
507 ix_S1 = load4(x_i + si * dim_stride + 3 * GMX_SIMD4_WIDTH);
508 iy_S1 = load4(x_i + si * dim_stride + 4 * GMX_SIMD4_WIDTH);
509 iz_S1 = load4(x_i + si * dim_stride + 5 * GMX_SIMD4_WIDTH);
511 /* We loop from the outer to the inner particles to maximize
512 * the chance that we find a pair in range quickly and return.
514 int j0 = csj * c_nbnxnGpuClusterSize;
515 int j1 = j0 + c_nbnxnGpuClusterSize - 1;
518 Simd4Real jx0_S, jy0_S, jz0_S;
519 Simd4Real jx1_S, jy1_S, jz1_S;
521 Simd4Real dx_S0, dy_S0, dz_S0;
522 Simd4Real dx_S1, dy_S1, dz_S1;
523 Simd4Real dx_S2, dy_S2, dz_S2;
524 Simd4Real dx_S3, dy_S3, dz_S3;
535 Simd4Bool wco_any_S01, wco_any_S23, wco_any_S;
537 jx0_S = Simd4Real(x_j[j0 * stride + 0]);
538 jy0_S = Simd4Real(x_j[j0 * stride + 1]);
539 jz0_S = Simd4Real(x_j[j0 * stride + 2]);
541 jx1_S = Simd4Real(x_j[j1 * stride + 0]);
542 jy1_S = Simd4Real(x_j[j1 * stride + 1]);
543 jz1_S = Simd4Real(x_j[j1 * stride + 2]);
545 /* Calculate distance */
546 dx_S0 = ix_S0 - jx0_S;
547 dy_S0 = iy_S0 - jy0_S;
548 dz_S0 = iz_S0 - jz0_S;
549 dx_S2 = ix_S0 - jx1_S;
550 dy_S2 = iy_S0 - jy1_S;
551 dz_S2 = iz_S0 - jz1_S;
552 if (c_nbnxnGpuClusterSize == 8)
554 dx_S1 = ix_S1 - jx0_S;
555 dy_S1 = iy_S1 - jy0_S;
556 dz_S1 = iz_S1 - jz0_S;
557 dx_S3 = ix_S1 - jx1_S;
558 dy_S3 = iy_S1 - jy1_S;
559 dz_S3 = iz_S1 - jz1_S;
562 /* rsq = dx*dx+dy*dy+dz*dz */
563 rsq_S0 = norm2(dx_S0, dy_S0, dz_S0);
564 rsq_S2 = norm2(dx_S2, dy_S2, dz_S2);
565 if (c_nbnxnGpuClusterSize == 8)
567 rsq_S1 = norm2(dx_S1, dy_S1, dz_S1);
568 rsq_S3 = norm2(dx_S3, dy_S3, dz_S3);
571 wco_S0 = (rsq_S0 < rc2_S);
572 wco_S2 = (rsq_S2 < rc2_S);
573 if (c_nbnxnGpuClusterSize == 8)
575 wco_S1 = (rsq_S1 < rc2_S);
576 wco_S3 = (rsq_S3 < rc2_S);
578 if (c_nbnxnGpuClusterSize == 8)
580 wco_any_S01 = wco_S0 || wco_S1;
581 wco_any_S23 = wco_S2 || wco_S3;
582 wco_any_S = wco_any_S01 || wco_any_S23;
586 wco_any_S = wco_S0 || wco_S2;
589 if (anyTrue(wco_any_S))
600 #endif /* !GMX_SIMD4_HAVE_REAL */
603 /* Returns the j-cluster index for index cjIndex in a cj list */
604 static inline int nblCj(gmx::ArrayRef<const nbnxn_cj_t> cjList, int cjIndex)
606 return cjList[cjIndex].cj;
609 /* Returns the j-cluster index for index cjIndex in a cj4 list */
610 static inline int nblCj(gmx::ArrayRef<const nbnxn_cj4_t> cj4List, int cjIndex)
612 return cj4List[cjIndex / c_nbnxnGpuJgroupSize].cj[cjIndex & (c_nbnxnGpuJgroupSize - 1)];
615 /* Returns the i-interaction mask of the j sub-cell for index cj_ind */
616 static unsigned int nbl_imask0(const NbnxnPairlistGpu* nbl, int cj_ind)
618 return nbl->cj4[cj_ind / c_nbnxnGpuJgroupSize].imei[0].imask;
621 NbnxnPairlistCpu::NbnxnPairlistCpu() :
622 na_ci(c_nbnxnCpuIClusterSize),
627 work(std::make_unique<NbnxnPairlistCpuWork>())
631 NbnxnPairlistGpu::NbnxnPairlistGpu(gmx::PinningPolicy pinningPolicy) :
632 na_ci(c_nbnxnGpuClusterSize),
633 na_cj(c_nbnxnGpuClusterSize),
634 na_sc(c_gpuNumClusterPerCell * c_nbnxnGpuClusterSize),
636 sci({}, { pinningPolicy }),
637 cj4({}, { pinningPolicy }),
638 excl({}, { pinningPolicy }),
640 work(std::make_unique<NbnxnPairlistGpuWork>())
642 static_assert(c_nbnxnGpuNumClusterPerSupercluster == c_gpuNumClusterPerCell,
643 "The search code assumes that the a super-cluster matches a search grid cell");
645 static_assert(sizeof(cj4[0].imei[0].imask) * 8 >= c_nbnxnGpuJgroupSize * c_gpuNumClusterPerCell,
646 "The i super-cluster cluster interaction mask does not contain a sufficient "
649 static_assert(sizeof(excl[0]) * 8 >= c_nbnxnGpuJgroupSize * c_gpuNumClusterPerCell,
650 "The GPU exclusion mask does not contain a sufficient number of bits");
652 // We always want a first entry without any exclusions
656 // TODO: Move to pairlistset.cpp
657 PairlistSet::PairlistSet(const InteractionLocality locality, const PairlistParams& pairlistParams) :
659 params_(pairlistParams),
660 combineLists_(sc_isGpuPairListType[pairlistParams.pairlistType]), // Currently GPU lists are always combined
661 isCpuType_(!sc_isGpuPairListType[pairlistParams.pairlistType])
664 const int numLists = gmx_omp_nthreads_get(emntNonbonded);
666 if (!combineLists_ && numLists > NBNXN_BUFFERFLAG_MAX_THREADS)
669 "%d OpenMP threads were requested. Since the non-bonded force buffer reduction "
670 "is prohibitively slow with more than %d threads, we do not allow this. Use %d "
671 "or less OpenMP threads.",
673 NBNXN_BUFFERFLAG_MAX_THREADS,
674 NBNXN_BUFFERFLAG_MAX_THREADS);
679 cpuLists_.resize(numLists);
682 cpuListsWork_.resize(numLists);
687 /* Only list 0 is used on the GPU, use normal allocation for i>0 */
688 gpuLists_.emplace_back(gmx::PinningPolicy::PinnedIfSupported);
689 /* Lists 0 to numLists are use for constructing lists in parallel
690 * on the CPU using numLists threads (and then merged into list 0).
692 for (int i = 1; i < numLists; i++)
694 gpuLists_.emplace_back(gmx::PinningPolicy::CannotBePinned);
699 fepLists_.resize(numLists);
701 /* Execute in order to avoid memory interleaving between threads */
702 #pragma omp parallel for num_threads(numLists) schedule(static)
703 for (int i = 0; i < numLists; i++)
707 /* We used to allocate all normal lists locally on each thread
708 * as well. The question is if allocating the object on the
709 * master thread (but all contained list memory thread local)
710 * impacts performance.
712 fepLists_[i] = std::make_unique<t_nblist>();
714 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
719 /* Print statistics of a pair list, used for debug output */
720 static void print_nblist_statistics(FILE* fp,
721 const NbnxnPairlistCpu& nbl,
722 const Nbnxm::GridSet& gridSet,
725 const Grid& grid = gridSet.grids()[0];
726 const Grid::Dimensions& dims = grid.dimensions();
728 fprintf(fp, "nbl nci %zu ncj %d\n", nbl.ci.size(), nbl.ncjInUse);
729 const int numAtomsJCluster = grid.geometry().numAtomsJCluster;
730 const double numAtomsPerCell = nbl.ncjInUse / static_cast<double>(grid.numCells()) * numAtomsJCluster;
732 "nbl na_cj %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
736 nbl.ncjInUse / static_cast<double>(grid.numCells()),
739 / (0.5 * 4.0 / 3.0 * M_PI * rl * rl * rl * grid.numCells() * numAtomsJCluster
740 / (dims.gridSize[XX] * dims.gridSize[YY] * dims.gridSize[ZZ])));
743 "nbl average j cell list length %.1f\n",
744 0.25 * nbl.ncjInUse / std::max(static_cast<double>(nbl.ci.size()), 1.0));
746 int cs[SHIFTS] = { 0 };
748 for (const nbnxn_ci_t& ciEntry : nbl.ci)
750 cs[ciEntry.shift & NBNXN_CI_SHIFT] += ciEntry.cj_ind_end - ciEntry.cj_ind_start;
752 int j = ciEntry.cj_ind_start;
753 while (j < ciEntry.cj_ind_end && nbl.cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
760 "nbl cell pairs, total: %zu excl: %d %.1f%%\n",
763 100 * npexcl / std::max(static_cast<double>(nbl.cj.size()), 1.0));
764 for (int s = 0; s < SHIFTS; s++)
768 fprintf(fp, "nbl shift %2d ncj %3d\n", s, cs[s]);
773 /* Print statistics of a pair lists, used for debug output */
774 static void print_nblist_statistics(FILE* fp,
775 const NbnxnPairlistGpu& nbl,
776 const Nbnxm::GridSet& gridSet,
779 const Grid& grid = gridSet.grids()[0];
780 const Grid::Dimensions& dims = grid.dimensions();
783 "nbl nsci %zu ncj4 %zu nsi %d excl4 %zu\n",
788 const int numAtomsCluster = grid.geometry().numAtomsICluster;
789 const double numAtomsPerCell = nbl.nci_tot / static_cast<double>(grid.numClusters()) * numAtomsCluster;
791 "nbl na_c %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
795 nbl.nci_tot / static_cast<double>(grid.numClusters()),
798 / (0.5 * 4.0 / 3.0 * M_PI * rl * rl * rl * grid.numClusters() * numAtomsCluster
799 / (dims.gridSize[XX] * dims.gridSize[YY] * dims.gridSize[ZZ])));
804 int c[c_gpuNumClusterPerCell + 1] = { 0 };
805 for (const nbnxn_sci_t& sci : nbl.sci)
808 for (int j4 = sci.cj4_ind_start; j4 < sci.cj4_ind_end; j4++)
810 for (int j = 0; j < c_nbnxnGpuJgroupSize; j++)
813 for (int si = 0; si < c_gpuNumClusterPerCell; si++)
815 if (nbl.cj4[j4].imei[0].imask & (1U << (j * c_gpuNumClusterPerCell + si)))
825 sum_nsp2 += nsp * nsp;
826 nsp_max = std::max(nsp_max, nsp);
828 if (!nbl.sci.empty())
830 sum_nsp /= nbl.sci.size();
831 sum_nsp2 /= nbl.sci.size();
834 "nbl #cluster-pairs: av %.1f stddev %.1f max %d\n",
836 std::sqrt(sum_nsp2 - sum_nsp * sum_nsp),
839 if (!nbl.cj4.empty())
841 for (int b = 0; b <= c_gpuNumClusterPerCell; b++)
844 "nbl j-list #i-subcell %d %7d %4.1f\n",
847 100.0 * c[b] / size_t{ nbl.cj4.size() * c_nbnxnGpuJgroupSize });
852 /* Returns a reference to the exclusion mask for j-cluster-group \p cj4 and warp \p warp
853 * Generates a new exclusion entry when the j-cluster-group uses
854 * the default all-interaction mask at call time, so the returned mask
855 * can be modified when needed.
857 static nbnxn_excl_t& get_exclusion_mask(NbnxnPairlistGpu* nbl, int cj4, int warp)
859 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
861 /* No exclusions set, make a new list entry */
862 const size_t oldSize = nbl->excl.size();
863 GMX_ASSERT(oldSize >= 1, "We should always have entry [0]");
864 /* Add entry with default values: no exclusions */
865 nbl->excl.resize(oldSize + 1);
866 nbl->cj4[cj4].imei[warp].excl_ind = oldSize;
869 return nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
872 /* Sets self exclusions and excludes half of the double pairs in the self cluster-pair \p nbl->cj4[cj4Index].cj[jOffsetInGroup]
874 * \param[in,out] nbl The cluster pair list
875 * \param[in] cj4Index The j-cluster group index into \p nbl->cj4
876 * \param[in] jOffsetInGroup The j-entry offset in \p nbl->cj4[cj4Index]
877 * \param[in] iClusterInCell The i-cluster index in the cell
879 static void setSelfAndNewtonExclusionsGpu(NbnxnPairlistGpu* nbl,
881 const int jOffsetInGroup,
882 const int iClusterInCell)
884 constexpr int numJatomsPerPart = c_nbnxnGpuClusterSize / c_nbnxnGpuClusterpairSplit;
886 /* The exclusions are stored separately for each part of the split */
887 for (int part = 0; part < c_nbnxnGpuClusterpairSplit; part++)
889 const int jOffset = part * numJatomsPerPart;
890 /* Make a new exclusion mask entry for each part, if we don't already have one yet */
891 nbnxn_excl_t& excl = get_exclusion_mask(nbl, cj4Index, part);
893 /* Set all bits with j-index <= i-index */
894 for (int jIndexInPart = 0; jIndexInPart < numJatomsPerPart; jIndexInPart++)
896 for (int i = jOffset + jIndexInPart; i < c_nbnxnGpuClusterSize; i++)
898 excl.pair[jIndexInPart * c_nbnxnGpuClusterSize + i] &=
899 ~(1U << (jOffsetInGroup * c_gpuNumClusterPerCell + iClusterInCell));
905 /* Returns a diagonal or off-diagonal interaction mask for plain C lists */
906 static unsigned int get_imask(gmx_bool rdiag, int ci, int cj)
908 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
911 /* Returns a diagonal or off-diagonal interaction mask for cj-size=2 */
912 gmx_unused static unsigned int get_imask_simd_j2(gmx_bool rdiag, int ci, int cj)
914 return (rdiag && ci * 2 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_0
915 : (rdiag && ci * 2 + 1 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_1
916 : NBNXN_INTERACTION_MASK_ALL));
919 /* Returns a diagonal or off-diagonal interaction mask for cj-size=4 */
920 gmx_unused static unsigned int get_imask_simd_j4(gmx_bool rdiag, int ci, int cj)
922 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
925 /* Returns a diagonal or off-diagonal interaction mask for cj-size=8 */
926 gmx_unused static unsigned int get_imask_simd_j8(gmx_bool rdiag, int ci, int cj)
928 return (rdiag && ci == cj * 2 ? NBNXN_INTERACTION_MASK_DIAG_J8_0
929 : (rdiag && ci == cj * 2 + 1 ? NBNXN_INTERACTION_MASK_DIAG_J8_1
930 : NBNXN_INTERACTION_MASK_ALL));
934 # if GMX_SIMD_REAL_WIDTH == 2
935 # define get_imask_simd_4xn get_imask_simd_j2
937 # if GMX_SIMD_REAL_WIDTH == 4
938 # define get_imask_simd_4xn get_imask_simd_j4
940 # if GMX_SIMD_REAL_WIDTH == 8
941 # define get_imask_simd_4xn get_imask_simd_j8
942 # define get_imask_simd_2xnn get_imask_simd_j4
944 # if GMX_SIMD_REAL_WIDTH == 16
945 # define get_imask_simd_2xnn get_imask_simd_j8
949 /* Plain C code for checking and adding cluster-pairs to the list.
951 * \param[in] gridj The j-grid
952 * \param[in,out] nbl The pair-list to store the cluster pairs in
953 * \param[in] icluster The index of the i-cluster
954 * \param[in] jclusterFirst The first cluster in the j-range
955 * \param[in] jclusterLast The last cluster in the j-range
956 * \param[in] excludeSubDiagonal Exclude atom pairs with i-index > j-index
957 * \param[in] x_j Coordinates for the j-atom, in xyz format
958 * \param[in] rlist2 The squared list cut-off
959 * \param[in] rbb2 The squared cut-off for putting cluster-pairs in the list based on bounding box distance only
960 * \param[in,out] numDistanceChecks The number of distance checks performed
962 static void makeClusterListSimple(const Grid& jGrid,
963 NbnxnPairlistCpu* nbl,
967 bool excludeSubDiagonal,
968 const real* gmx_restrict x_j,
971 int* gmx_restrict numDistanceChecks)
973 const BoundingBox* gmx_restrict bb_ci = nbl->work->iClusterData.bb.data();
974 const real* gmx_restrict x_ci = nbl->work->iClusterData.x.data();
976 bool InRange = false;
977 while (!InRange && jclusterFirst <= jclusterLast)
979 real d2 = clusterBoundingBoxDistance2(bb_ci[0], jGrid.jBoundingBoxes()[jclusterFirst]);
980 *numDistanceChecks += 2;
982 /* Check if the distance is within the distance where
983 * we use only the bounding box distance rbb,
984 * or within the cut-off and there is at least one atom pair
985 * within the cut-off.
991 else if (d2 < rlist2)
993 int cjf_gl = jGrid.cellOffset() + jclusterFirst;
994 for (int i = 0; i < c_nbnxnCpuIClusterSize && !InRange; i++)
996 for (int j = 0; j < c_nbnxnCpuIClusterSize; j++)
1000 || (gmx::square(x_ci[i * STRIDE_XYZ + XX]
1001 - x_j[(cjf_gl * c_nbnxnCpuIClusterSize + j) * STRIDE_XYZ + XX])
1002 + gmx::square(x_ci[i * STRIDE_XYZ + YY]
1003 - x_j[(cjf_gl * c_nbnxnCpuIClusterSize + j) * STRIDE_XYZ + YY])
1004 + gmx::square(x_ci[i * STRIDE_XYZ + ZZ]
1005 - x_j[(cjf_gl * c_nbnxnCpuIClusterSize + j) * STRIDE_XYZ + ZZ])
1009 *numDistanceChecks += c_nbnxnCpuIClusterSize * c_nbnxnCpuIClusterSize;
1022 while (!InRange && jclusterLast > jclusterFirst)
1024 real d2 = clusterBoundingBoxDistance2(bb_ci[0], jGrid.jBoundingBoxes()[jclusterLast]);
1025 *numDistanceChecks += 2;
1027 /* Check if the distance is within the distance where
1028 * we use only the bounding box distance rbb,
1029 * or within the cut-off and there is at least one atom pair
1030 * within the cut-off.
1036 else if (d2 < rlist2)
1038 int cjl_gl = jGrid.cellOffset() + jclusterLast;
1039 for (int i = 0; i < c_nbnxnCpuIClusterSize && !InRange; i++)
1041 for (int j = 0; j < c_nbnxnCpuIClusterSize; j++)
1045 || (gmx::square(x_ci[i * STRIDE_XYZ + XX]
1046 - x_j[(cjl_gl * c_nbnxnCpuIClusterSize + j) * STRIDE_XYZ + XX])
1047 + gmx::square(x_ci[i * STRIDE_XYZ + YY]
1048 - x_j[(cjl_gl * c_nbnxnCpuIClusterSize + j) * STRIDE_XYZ + YY])
1049 + gmx::square(x_ci[i * STRIDE_XYZ + ZZ]
1050 - x_j[(cjl_gl * c_nbnxnCpuIClusterSize + j) * STRIDE_XYZ + ZZ])
1054 *numDistanceChecks += c_nbnxnCpuIClusterSize * c_nbnxnCpuIClusterSize;
1062 if (jclusterFirst <= jclusterLast)
1064 for (int jcluster = jclusterFirst; jcluster <= jclusterLast; jcluster++)
1066 /* Store cj and the interaction mask */
1068 cjEntry.cj = jGrid.cellOffset() + jcluster;
1069 cjEntry.excl = get_imask(excludeSubDiagonal, icluster, jcluster);
1070 nbl->cj.push_back(cjEntry);
1072 /* Increase the closing index in the i list */
1073 nbl->ci.back().cj_ind_end = nbl->cj.size();
1077 #ifdef GMX_NBNXN_SIMD_4XN
1078 # include "pairlist_simd_4xm.h"
1080 #ifdef GMX_NBNXN_SIMD_2XNN
1081 # include "pairlist_simd_2xmm.h"
1084 /* Plain C or SIMD4 code for making a pair list of super-cell sci vs scj.
1085 * Checks bounding box distances and possibly atom pair distances.
1087 static void make_cluster_list_supersub(const Grid& iGrid,
1089 NbnxnPairlistGpu* nbl,
1092 const bool excludeSubDiagonal,
1097 int* numDistanceChecks)
1099 NbnxnPairlistGpuWork& work = *nbl->work;
1102 const float* pbb_ci = work.iSuperClusterData.bbPacked.data();
1104 const BoundingBox* bb_ci = work.iSuperClusterData.bb.data();
1107 assert(c_nbnxnGpuClusterSize == iGrid.geometry().numAtomsICluster);
1108 assert(c_nbnxnGpuClusterSize == jGrid.geometry().numAtomsICluster);
1110 /* We generate the pairlist mainly based on bounding-box distances
1111 * and do atom pair distance based pruning on the GPU.
1112 * Only if a j-group contains a single cluster-pair, we try to prune
1113 * that pair based on atom distances on the CPU to avoid empty j-groups.
1115 #define PRUNE_LIST_CPU_ONE 1
1116 #define PRUNE_LIST_CPU_ALL 0
1118 #if PRUNE_LIST_CPU_ONE
1122 float* d2l = work.distanceBuffer.data();
1124 for (int subc = 0; subc < jGrid.numClustersPerCell()[scj]; subc++)
1126 const int cj4_ind = work.cj_ind / c_nbnxnGpuJgroupSize;
1127 const int cj_offset = work.cj_ind - cj4_ind * c_nbnxnGpuJgroupSize;
1128 const int cj = scj * c_gpuNumClusterPerCell + subc;
1130 const int cj_gl = jGrid.cellOffset() * c_gpuNumClusterPerCell + cj;
1132 int ci1 = (excludeSubDiagonal && sci == scj) ? subc + 1 : iGrid.numClustersPerCell()[sci];
1136 /* Determine all ci1 bb distances in one call with SIMD4 */
1137 const int offset = packedBoundingBoxesIndex(cj) + (cj & (c_packedBoundingBoxesDimSize - 1));
1138 clusterBoundingBoxDistance2_xxxx_simd4(
1139 jGrid.packedBoundingBoxes().data() + offset, ci1, pbb_ci, d2l);
1140 *numDistanceChecks += c_nbnxnGpuClusterSize * 2;
1144 unsigned int imask = 0;
1145 /* We use a fixed upper-bound instead of ci1 to help optimization */
1146 for (int ci = 0; ci < c_gpuNumClusterPerCell; ci++)
1154 /* Determine the bb distance between ci and cj */
1155 d2l[ci] = clusterBoundingBoxDistance2(bb_ci[ci], jGrid.jBoundingBoxes()[cj]);
1156 *numDistanceChecks += 2;
1160 #if PRUNE_LIST_CPU_ALL
1161 /* Check if the distance is within the distance where
1162 * we use only the bounding box distance rbb,
1163 * or within the cut-off and there is at least one atom pair
1164 * within the cut-off. This check is very costly.
1166 *numDistanceChecks += c_nbnxnGpuClusterSize * c_nbnxnGpuClusterSize;
1167 if (d2 < rbb2 || (d2 < rlist2 && clusterpair_in_range(work, ci, cj_gl, stride, x, rlist2)))
1169 /* Check if the distance between the two bounding boxes
1170 * in within the pair-list cut-off.
1175 /* Flag this i-subcell to be taken into account */
1176 imask |= (1U << (cj_offset * c_gpuNumClusterPerCell + ci));
1178 #if PRUNE_LIST_CPU_ONE
1186 #if PRUNE_LIST_CPU_ONE
1187 /* If we only found 1 pair, check if any atoms are actually
1188 * within the cut-off, so we could get rid of it.
1190 if (npair == 1 && d2l[ci_last] >= rbb2
1191 && !clusterpair_in_range(work, ci_last, cj_gl, stride, x, rlist2))
1193 imask &= ~(1U << (cj_offset * c_gpuNumClusterPerCell + ci_last));
1200 /* We have at least one cluster pair: add a j-entry */
1201 if (static_cast<size_t>(cj4_ind) == nbl->cj4.size())
1203 nbl->cj4.resize(nbl->cj4.size() + 1);
1205 nbnxn_cj4_t* cj4 = &nbl->cj4[cj4_ind];
1207 cj4->cj[cj_offset] = cj_gl;
1209 /* Set the exclusions for the ci==sj entry.
1210 * Here we don't bother to check if this entry is actually flagged,
1211 * as it will nearly always be in the list.
1213 if (excludeSubDiagonal && sci == scj)
1215 setSelfAndNewtonExclusionsGpu(nbl, cj4_ind, cj_offset, subc);
1218 /* Copy the cluster interaction mask to the list */
1219 for (int w = 0; w < c_nbnxnGpuClusterpairSplit; w++)
1221 cj4->imei[w].imask |= imask;
1224 nbl->work->cj_ind++;
1226 /* Keep the count */
1227 nbl->nci_tot += npair;
1229 /* Increase the closing index in i super-cell list */
1230 nbl->sci.back().cj4_ind_end =
1231 (nbl->work->cj_ind + c_nbnxnGpuJgroupSize - 1) / c_nbnxnGpuJgroupSize;
1236 /* Returns how many contiguous j-clusters we have starting in the i-list */
1237 template<typename CjListType>
1238 static int numContiguousJClusters(const int cjIndexStart,
1239 const int cjIndexEnd,
1240 gmx::ArrayRef<const CjListType> cjList)
1242 const int firstJCluster = nblCj(cjList, cjIndexStart);
1244 int numContiguous = 0;
1246 while (cjIndexStart + numContiguous < cjIndexEnd
1247 && nblCj(cjList, cjIndexStart + numContiguous) == firstJCluster + numContiguous)
1252 return numContiguous;
1256 * \brief Helper struct for efficient searching for excluded atoms in a j-list
1260 /*! \brief Constructs a j-list range from \p cjList with the given index range */
1261 template<typename CjListType>
1262 JListRanges(int cjIndexStart, int cjIndexEnd, gmx::ArrayRef<const CjListType> cjList);
1264 int cjIndexStart; //!< The start index in the j-list
1265 int cjIndexEnd; //!< The end index in the j-list
1266 int cjFirst; //!< The j-cluster with index cjIndexStart
1267 int cjLast; //!< The j-cluster with index cjIndexEnd-1
1268 int numDirect; //!< Up to cjIndexStart+numDirect the j-clusters are cjFirst + the index offset
1272 template<typename CjListType>
1273 JListRanges::JListRanges(int cjIndexStart, int cjIndexEnd, gmx::ArrayRef<const CjListType> cjList) :
1274 cjIndexStart(cjIndexStart),
1275 cjIndexEnd(cjIndexEnd)
1277 GMX_ASSERT(cjIndexEnd > cjIndexStart, "JListRanges should only be called with non-empty lists");
1279 cjFirst = nblCj(cjList, cjIndexStart);
1280 cjLast = nblCj(cjList, cjIndexEnd - 1);
1282 /* Determine how many contiguous j-cells we have starting
1283 * from the first i-cell. This number can be used to directly
1284 * calculate j-cell indices for excluded atoms.
1286 numDirect = numContiguousJClusters(cjIndexStart, cjIndexEnd, cjList);
1290 /* Return the index of \p jCluster in the given range or -1 when not present
1292 * Note: This code is executed very often and therefore performance is
1293 * important. It should be inlined and fully optimized.
1295 template<typename CjListType>
1296 static inline int findJClusterInJList(int jCluster,
1297 const JListRanges& ranges,
1298 gmx::ArrayRef<const CjListType> cjList)
1300 if (jCluster < ranges.cjFirst + ranges.numDirect)
1302 /* We can calculate the index directly using the offset */
1303 return ranges.cjIndexStart + jCluster - ranges.cjFirst;
1307 /* Search for jCluster using bisection */
1309 int rangeStart = ranges.cjIndexStart + ranges.numDirect;
1310 int rangeEnd = ranges.cjIndexEnd;
1311 while (index == -1 && rangeStart < rangeEnd)
1313 int rangeMiddle = (rangeStart + rangeEnd) >> 1;
1315 const int clusterMiddle = nblCj(cjList, rangeMiddle);
1317 if (jCluster == clusterMiddle)
1319 index = rangeMiddle;
1321 else if (jCluster < clusterMiddle)
1323 rangeEnd = rangeMiddle;
1327 rangeStart = rangeMiddle + 1;
1334 // TODO: Get rid of the two functions below by renaming sci to ci (or something better)
1336 /* Return the i-entry in the list we are currently operating on */
1337 static nbnxn_ci_t* getOpenIEntry(NbnxnPairlistCpu* nbl)
1339 return &nbl->ci.back();
1342 /* Return the i-entry in the list we are currently operating on */
1343 static nbnxn_sci_t* getOpenIEntry(NbnxnPairlistGpu* nbl)
1345 return &nbl->sci.back();
1348 /* Set all atom-pair exclusions for a simple type list i-entry
1350 * Set all atom-pair exclusions from the topology stored in exclusions
1351 * as masks in the pair-list for simple list entry iEntry.
1353 static void setExclusionsForIEntry(const Nbnxm::GridSet& gridSet,
1354 NbnxnPairlistCpu* nbl,
1355 gmx_bool diagRemoved,
1357 const nbnxn_ci_t& iEntry,
1358 const ListOfLists<int>& exclusions)
1360 if (iEntry.cj_ind_end == iEntry.cj_ind_start)
1362 /* Empty list: no exclusions */
1366 const JListRanges ranges(iEntry.cj_ind_start, iEntry.cj_ind_end, gmx::makeConstArrayRef(nbl->cj));
1368 const int iCluster = iEntry.ci;
1370 gmx::ArrayRef<const int> cell = gridSet.cells();
1371 gmx::ArrayRef<const int> atomIndices = gridSet.atomIndices();
1373 /* Loop over the atoms in the i-cluster */
1374 for (int i = 0; i < nbl->na_ci; i++)
1376 const int iIndex = iCluster * nbl->na_ci + i;
1377 const int iAtom = atomIndices[iIndex];
1380 /* Loop over the topology-based exclusions for this i-atom */
1381 for (const int jAtom : exclusions[iAtom])
1385 /* The self exclusion are already set, save some time */
1389 /* Get the index of the j-atom in the nbnxn atom data */
1390 const int jIndex = cell[jAtom];
1392 /* Without shifts we only calculate interactions j>i
1393 * for one-way pair-lists.
1395 if (diagRemoved && jIndex <= iIndex)
1400 const int jCluster = (jIndex >> na_cj_2log);
1402 /* Could the cluster se be in our list? */
1403 if (jCluster >= ranges.cjFirst && jCluster <= ranges.cjLast)
1406 findJClusterInJList(jCluster, ranges, gmx::makeConstArrayRef(nbl->cj));
1410 /* We found an exclusion, clear the corresponding
1413 const int innerJ = jIndex - (jCluster << na_cj_2log);
1415 nbl->cj[index].excl &= ~(1U << ((i << na_cj_2log) + innerJ));
1423 /* Add a new i-entry to the FEP list and copy the i-properties */
1424 static inline void fep_list_new_nri_copy(t_nblist* nlist)
1426 /* Add a new i-entry */
1429 assert(nlist->nri < nlist->maxnri);
1431 /* Duplicate the last i-entry, except for jindex, which continues */
1432 nlist->iinr[nlist->nri] = nlist->iinr[nlist->nri - 1];
1433 nlist->shift[nlist->nri] = nlist->shift[nlist->nri - 1];
1434 nlist->gid[nlist->nri] = nlist->gid[nlist->nri - 1];
1435 nlist->jindex[nlist->nri] = nlist->nrj;
1438 /* Rellocate FEP list for size nl->maxnri, TODO: replace by C++ */
1439 static void reallocate_nblist(t_nblist* nl)
1441 nl->iinr.resize(nl->maxnri);
1442 nl->gid.resize(nl->maxnri);
1443 nl->shift.resize(nl->maxnri);
1444 nl->jindex.resize(nl->maxnri + 1);
1447 /* For load balancing of the free-energy lists over threads, we set
1448 * the maximum nrj size of an i-entry to 40. This leads to good
1449 * load balancing in the worst case scenario of a single perturbed
1450 * particle on 16 threads, while not introducing significant overhead.
1451 * Note that half of the perturbed pairs will anyhow end up in very small lists,
1452 * since non perturbed i-particles will see few perturbed j-particles).
1454 const int max_nrj_fep = 40;
1456 /* Exclude the perturbed pairs from the Verlet list. This is only done to avoid
1457 * singularities for overlapping particles (0/0), since the charges and
1458 * LJ parameters have been zeroed in the nbnxn data structure.
1459 * Simultaneously make a group pair list for the perturbed pairs.
1461 static void make_fep_list(gmx::ArrayRef<const int> atomIndices,
1462 const nbnxn_atomdata_t* nbat,
1463 NbnxnPairlistCpu* nbl,
1464 gmx_bool bDiagRemoved,
1466 real gmx_unused shx,
1467 real gmx_unused shy,
1468 real gmx_unused shz,
1469 real gmx_unused rlist_fep2,
1477 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
1483 const int ci = nbl_ci->ci;
1485 const int cj_ind_start = nbl_ci->cj_ind_start;
1486 const int cj_ind_end = nbl_ci->cj_ind_end;
1488 /* In worst case we have alternating energy groups
1489 * and create #atom-pair lists, which means we need the size
1490 * of a cluster pair (na_ci*na_cj) times the number of cj's.
1492 const int nri_max = nbl->na_ci * nbl->na_cj * (cj_ind_end - cj_ind_start);
1493 if (nlist->nri + nri_max > nlist->maxnri)
1495 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
1496 reallocate_nblist(nlist);
1499 const int numAtomsJCluster = jGrid.geometry().numAtomsJCluster;
1501 const nbnxn_atomdata_t::Params& nbatParams = nbat->params();
1503 const int ngid = nbatParams.nenergrp;
1505 /* TODO: Consider adding a check in grompp and changing this to an assert */
1506 const int numBitsInEnergyGroupIdsForAtomsInJCluster = sizeof(gid_cj) * 8;
1507 if (ngid * numAtomsJCluster > numBitsInEnergyGroupIdsForAtomsInJCluster)
1510 "The Verlet scheme with %dx%d kernels and free-energy only supports up to %zu "
1512 iGrid.geometry().numAtomsICluster,
1514 (sizeof(gid_cj) * 8) / numAtomsJCluster);
1517 const int egp_shift = nbatParams.neg_2log;
1518 const int egp_mask = (1 << egp_shift) - 1;
1520 /* Loop over the atoms in the i sub-cell */
1521 bool bFEP_i_all = true;
1522 for (int i = 0; i < nbl->na_ci; i++)
1524 const int ind_i = ci * nbl->na_ci + i;
1525 const int ai = atomIndices[ind_i];
1528 int nri = nlist->nri;
1529 nlist->jindex[nri + 1] = nlist->jindex[nri];
1530 nlist->iinr[nri] = ai;
1531 /* The actual energy group pair index is set later */
1532 nlist->gid[nri] = 0;
1533 nlist->shift[nri] = nbl_ci->shift & NBNXN_CI_SHIFT;
1535 bool bFEP_i = iGrid.atomIsPerturbed(ci - iGrid.cellOffset(), i);
1537 bFEP_i_all = bFEP_i_all && bFEP_i;
1539 if (nlist->nrj + (cj_ind_end - cj_ind_start) * nbl->na_cj > nlist->maxnrj)
1541 nlist->maxnrj = over_alloc_small(nlist->nrj + (cj_ind_end - cj_ind_start) * nbl->na_cj);
1542 nlist->jjnr.resize(nlist->maxnrj);
1543 nlist->excl_fep.resize(nlist->maxnrj);
1548 gid_i = (nbatParams.energrp[ci] >> (egp_shift * i)) & egp_mask;
1551 for (int cj_ind = cj_ind_start; cj_ind < cj_ind_end; cj_ind++)
1553 unsigned int fep_cj = 0U;
1556 const int cja = nbl->cj[cj_ind].cj;
1558 if (numAtomsJCluster == jGrid.geometry().numAtomsICluster)
1560 const int cjr = cja - jGrid.cellOffset();
1561 fep_cj = jGrid.fepBits(cjr);
1564 gid_cj = nbatParams.energrp[cja];
1567 else if (2 * numAtomsJCluster == jGrid.geometry().numAtomsICluster)
1569 const int cjr = cja - jGrid.cellOffset() * 2;
1570 /* Extract half of the ci fep/energrp mask */
1571 fep_cj = (jGrid.fepBits(cjr >> 1) >> ((cjr & 1) * numAtomsJCluster))
1572 & ((1 << numAtomsJCluster) - 1);
1575 gid_cj = nbatParams.energrp[cja >> 1] >> ((cja & 1) * numAtomsJCluster * egp_shift)
1576 & ((1 << (numAtomsJCluster * egp_shift)) - 1);
1581 const int cjr = cja - (jGrid.cellOffset() >> 1);
1582 /* Combine two ci fep masks/energrp */
1583 fep_cj = jGrid.fepBits(cjr * 2)
1584 + (jGrid.fepBits(cjr * 2 + 1) << jGrid.geometry().numAtomsICluster);
1587 gid_cj = nbatParams.energrp[cja * 2]
1588 + (nbatParams.energrp[cja * 2 + 1]
1589 << (jGrid.geometry().numAtomsICluster * egp_shift));
1593 if (bFEP_i || fep_cj != 0)
1595 for (int j = 0; j < nbl->na_cj; j++)
1597 /* Is this interaction perturbed and not excluded? */
1598 const int ind_j = cja * nbl->na_cj + j;
1599 const int aj = atomIndices[ind_j];
1600 if (aj >= 0 && (bFEP_i || (fep_cj & (1 << j))) && (!bDiagRemoved || ind_j >= ind_i))
1604 const int gid_j = (gid_cj >> (j * egp_shift)) & egp_mask;
1605 const int gid = GID(gid_i, gid_j, ngid);
1607 if (nlist->nrj > nlist->jindex[nri] && nlist->gid[nri] != gid)
1609 /* Energy group pair changed: new list */
1610 fep_list_new_nri_copy(nlist);
1613 nlist->gid[nri] = gid;
1616 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
1618 fep_list_new_nri_copy(nlist);
1622 /* Add it to the FEP list */
1623 nlist->jjnr[nlist->nrj] = aj;
1624 nlist->excl_fep[nlist->nrj] =
1625 (nbl->cj[cj_ind].excl >> (i * nbl->na_cj + j)) & 1;
1628 /* Exclude it from the normal list.
1629 * Note that the charge has been set to zero,
1630 * but we need to avoid 0/0, as perturbed atoms
1631 * can be on top of each other.
1633 nbl->cj[cj_ind].excl &= ~(1U << (i * nbl->na_cj + j));
1639 if (nlist->nrj > nlist->jindex[nri])
1641 /* Actually add this new, non-empty, list */
1643 nlist->jindex[nlist->nri] = nlist->nrj;
1650 /* All interactions are perturbed, we can skip this entry */
1651 nbl_ci->cj_ind_end = cj_ind_start;
1652 nbl->ncjInUse -= cj_ind_end - cj_ind_start;
1656 /* Return the index of atom a within a cluster */
1657 static inline int cj_mod_cj4(int cj)
1659 return cj & (c_nbnxnGpuJgroupSize - 1);
1662 /* Convert a j-cluster to a cj4 group */
1663 static inline int cj_to_cj4(int cj)
1665 return cj / c_nbnxnGpuJgroupSize;
1668 /* Return the index of an j-atom within a warp */
1669 static inline int a_mod_wj(int a)
1671 return a & (c_nbnxnGpuClusterSize / c_nbnxnGpuClusterpairSplit - 1);
1674 /* As make_fep_list above, but for super/sub lists. */
1675 static void make_fep_list(gmx::ArrayRef<const int> atomIndices,
1676 const nbnxn_atomdata_t* nbat,
1677 NbnxnPairlistGpu* nbl,
1678 gmx_bool bDiagRemoved,
1679 const nbnxn_sci_t* nbl_sci,
1688 const int numJClusterGroups = nbl_sci->numJClusterGroups();
1689 if (numJClusterGroups == 0)
1695 const int sci = nbl_sci->sci;
1697 const int cj4_ind_start = nbl_sci->cj4_ind_start;
1698 const int cj4_ind_end = nbl_sci->cj4_ind_end;
1700 /* Here we process one super-cell, max #atoms na_sc, versus a list
1701 * cj4 entries, each with max c_nbnxnGpuJgroupSize cj's, each
1702 * of size na_cj atoms.
1703 * On the GPU we don't support energy groups (yet).
1704 * So for each of the na_sc i-atoms, we need max one FEP list
1705 * for each max_nrj_fep j-atoms.
1708 nbl->na_sc * nbl->na_cj * (1 + (numJClusterGroups * c_nbnxnGpuJgroupSize) / max_nrj_fep);
1709 if (nlist->nri + nri_max > nlist->maxnri)
1711 nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
1712 reallocate_nblist(nlist);
1715 /* Loop over the atoms in the i super-cluster */
1716 for (int c = 0; c < c_gpuNumClusterPerCell; c++)
1718 const int c_abs = sci * c_gpuNumClusterPerCell + c;
1720 for (int i = 0; i < nbl->na_ci; i++)
1722 const int ind_i = c_abs * nbl->na_ci + i;
1723 const int ai = atomIndices[ind_i];
1726 int nri = nlist->nri;
1727 nlist->jindex[nri + 1] = nlist->jindex[nri];
1728 nlist->iinr[nri] = ai;
1729 /* With GPUs, energy groups are not supported */
1730 nlist->gid[nri] = 0;
1731 nlist->shift[nri] = nbl_sci->shift & NBNXN_CI_SHIFT;
1734 iGrid.atomIsPerturbed(c_abs - iGrid.cellOffset() * c_gpuNumClusterPerCell, i);
1736 real xi = nbat->x()[ind_i * nbat->xstride + XX] + shx;
1737 real yi = nbat->x()[ind_i * nbat->xstride + YY] + shy;
1738 real zi = nbat->x()[ind_i * nbat->xstride + ZZ] + shz;
1740 const int nrjMax = nlist->nrj + numJClusterGroups * c_nbnxnGpuJgroupSize * nbl->na_cj;
1741 if (nrjMax > nlist->maxnrj)
1743 nlist->maxnrj = over_alloc_small(nrjMax);
1744 nlist->jjnr.resize(nlist->maxnrj);
1745 nlist->excl_fep.resize(nlist->maxnrj);
1748 for (int cj4_ind = cj4_ind_start; cj4_ind < cj4_ind_end; cj4_ind++)
1750 const nbnxn_cj4_t* cj4 = &nbl->cj4[cj4_ind];
1752 for (int gcj = 0; gcj < c_nbnxnGpuJgroupSize; gcj++)
1754 if ((cj4->imei[0].imask & (1U << (gcj * c_gpuNumClusterPerCell + c))) == 0)
1756 /* Skip this ci for this cj */
1760 const int cjr = cj4->cj[gcj] - jGrid.cellOffset() * c_gpuNumClusterPerCell;
1762 if (bFEP_i || jGrid.clusterIsPerturbed(cjr))
1764 for (int j = 0; j < nbl->na_cj; j++)
1766 /* Is this interaction perturbed and not excluded? */
1768 (jGrid.cellOffset() * c_gpuNumClusterPerCell + cjr) * nbl->na_cj + j;
1769 const int aj = atomIndices[ind_j];
1770 if (aj >= 0 && (bFEP_i || jGrid.atomIsPerturbed(cjr, j))
1771 && (!bDiagRemoved || ind_j >= ind_i))
1774 j / (c_nbnxnGpuClusterSize / c_nbnxnGpuClusterpairSplit);
1775 nbnxn_excl_t& excl = get_exclusion_mask(nbl, cj4_ind, jHalf);
1777 int excl_pair = a_mod_wj(j) * nbl->na_ci + i;
1778 unsigned int excl_bit = (1U << (gcj * c_gpuNumClusterPerCell + c));
1780 real dx = nbat->x()[ind_j * nbat->xstride + XX] - xi;
1781 real dy = nbat->x()[ind_j * nbat->xstride + YY] - yi;
1782 real dz = nbat->x()[ind_j * nbat->xstride + ZZ] - zi;
1784 /* The unpruned GPU list has more than 2/3
1785 * of the atom pairs beyond rlist. Using
1786 * this list will cause a lot of overhead
1787 * in the CPU FEP kernels, especially
1788 * relative to the fast GPU kernels.
1789 * So we prune the FEP list here.
1791 if (dx * dx + dy * dy + dz * dz < rlist_fep2)
1793 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
1795 fep_list_new_nri_copy(nlist);
1799 /* Add it to the FEP list */
1800 nlist->jjnr[nlist->nrj] = aj;
1801 nlist->excl_fep[nlist->nrj] =
1802 (excl.pair[excl_pair] & excl_bit) ? 1 : 0;
1806 /* Exclude it from the normal list.
1807 * Note that the charge and LJ parameters have
1808 * been set to zero, but we need to avoid 0/0,
1809 * as perturbed atoms can be on top of each other.
1811 excl.pair[excl_pair] &= ~excl_bit;
1815 /* Note that we could mask out this pair in imask
1816 * if all i- and/or all j-particles are perturbed.
1817 * But since the perturbed pairs on the CPU will
1818 * take an order of magnitude more time, the GPU
1819 * will finish before the CPU and there is no gain.
1825 if (nlist->nrj > nlist->jindex[nri])
1827 /* Actually add this new, non-empty, list */
1829 nlist->jindex[nlist->nri] = nlist->nrj;
1836 /* Set all atom-pair exclusions for a GPU type list i-entry
1838 * Sets all atom-pair exclusions from the topology stored in exclusions
1839 * as masks in the pair-list for i-super-cluster list entry iEntry.
1841 static void setExclusionsForIEntry(const Nbnxm::GridSet& gridSet,
1842 NbnxnPairlistGpu* nbl,
1843 gmx_bool diagRemoved,
1844 int gmx_unused na_cj_2log,
1845 const nbnxn_sci_t& iEntry,
1846 const ListOfLists<int>& exclusions)
1848 if (iEntry.numJClusterGroups() == 0)
1854 /* Set the search ranges using start and end j-cluster indices.
1855 * Note that here we can not use cj4_ind_end, since the last cj4
1856 * can be only partially filled, so we use cj_ind.
1858 const JListRanges ranges(iEntry.cj4_ind_start * c_nbnxnGpuJgroupSize,
1860 gmx::makeConstArrayRef(nbl->cj4));
1862 GMX_ASSERT(nbl->na_ci == c_nbnxnGpuClusterSize, "na_ci should match the GPU cluster size");
1863 constexpr int c_clusterSize = c_nbnxnGpuClusterSize;
1864 constexpr int c_superClusterSize = c_nbnxnGpuNumClusterPerSupercluster * c_nbnxnGpuClusterSize;
1866 const int iSuperCluster = iEntry.sci;
1868 gmx::ArrayRef<const int> atomIndices = gridSet.atomIndices();
1869 gmx::ArrayRef<const int> cell = gridSet.cells();
1871 /* Loop over the atoms in the i super-cluster */
1872 for (int i = 0; i < c_superClusterSize; i++)
1874 const int iIndex = iSuperCluster * c_superClusterSize + i;
1875 const int iAtom = atomIndices[iIndex];
1878 const int iCluster = i / c_clusterSize;
1880 /* Loop over the topology-based exclusions for this i-atom */
1881 for (const int jAtom : exclusions[iAtom])
1885 /* The self exclusions are already set, save some time */
1889 /* Get the index of the j-atom in the nbnxn atom data */
1890 const int jIndex = cell[jAtom];
1892 /* Without shifts we only calculate interactions j>i
1893 * for one-way pair-lists.
1895 /* NOTE: We would like to use iIndex on the right hand side,
1896 * but that makes this routine 25% slower with gcc6/7.
1897 * Even using c_superClusterSize makes it slower.
1898 * Either of these changes triggers peeling of the exclIndex
1899 * loop, which apparently leads to far less efficient code.
1901 if (diagRemoved && jIndex <= iSuperCluster * nbl->na_sc + i)
1906 const int jCluster = jIndex / c_clusterSize;
1908 /* Check whether the cluster is in our list? */
1909 if (jCluster >= ranges.cjFirst && jCluster <= ranges.cjLast)
1912 findJClusterInJList(jCluster, ranges, gmx::makeConstArrayRef(nbl->cj4));
1916 /* We found an exclusion, clear the corresponding
1919 const unsigned int pairMask =
1920 (1U << (cj_mod_cj4(index) * c_gpuNumClusterPerCell + iCluster));
1921 /* Check if the i-cluster interacts with the j-cluster */
1922 if (nbl_imask0(nbl, index) & pairMask)
1924 const int innerI = (i & (c_clusterSize - 1));
1925 const int innerJ = (jIndex & (c_clusterSize - 1));
1927 /* Determine which j-half (CUDA warp) we are in */
1928 const int jHalf = innerJ / (c_clusterSize / c_nbnxnGpuClusterpairSplit);
1930 nbnxn_excl_t& interactionMask =
1931 get_exclusion_mask(nbl, cj_to_cj4(index), jHalf);
1933 interactionMask.pair[a_mod_wj(innerJ) * c_clusterSize + innerI] &= ~pairMask;
1942 /* Make a new ci entry at the back of nbl->ci */
1943 static void addNewIEntry(NbnxnPairlistCpu* nbl, int ci, int shift, int flags)
1947 ciEntry.shift = shift;
1948 /* Store the interaction flags along with the shift */
1949 ciEntry.shift |= flags;
1950 ciEntry.cj_ind_start = nbl->cj.size();
1951 ciEntry.cj_ind_end = nbl->cj.size();
1952 nbl->ci.push_back(ciEntry);
1955 /* Make a new sci entry at index nbl->nsci */
1956 static void addNewIEntry(NbnxnPairlistGpu* nbl, int sci, int shift, int gmx_unused flags)
1958 nbnxn_sci_t sciEntry;
1960 sciEntry.shift = shift;
1961 sciEntry.cj4_ind_start = nbl->cj4.size();
1962 sciEntry.cj4_ind_end = nbl->cj4.size();
1964 nbl->sci.push_back(sciEntry);
1967 /* Sort the simple j-list cj on exclusions.
1968 * Entries with exclusions will all be sorted to the beginning of the list.
1970 static void sort_cj_excl(nbnxn_cj_t* cj, int ncj, NbnxnPairlistCpuWork* work)
1972 work->cj.resize(ncj);
1974 /* Make a list of the j-cells involving exclusions */
1976 for (int j = 0; j < ncj; j++)
1978 if (cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
1980 work->cj[jnew++] = cj[j];
1983 /* Check if there are exclusions at all or not just the first entry */
1984 if (!((jnew == 0) || (jnew == 1 && cj[0].excl != NBNXN_INTERACTION_MASK_ALL)))
1986 for (int j = 0; j < ncj; j++)
1988 if (cj[j].excl == NBNXN_INTERACTION_MASK_ALL)
1990 work->cj[jnew++] = cj[j];
1993 for (int j = 0; j < ncj; j++)
1995 cj[j] = work->cj[j];
2000 /* Close this simple list i entry */
2001 static void closeIEntry(NbnxnPairlistCpu* nbl,
2002 int gmx_unused sp_max_av,
2003 gmx_bool gmx_unused progBal,
2004 float gmx_unused nsp_tot_est,
2005 int gmx_unused thread,
2006 int gmx_unused nthread)
2008 nbnxn_ci_t& ciEntry = nbl->ci.back();
2010 /* All content of the new ci entry have already been filled correctly,
2011 * we only need to sort and increase counts or remove the entry when empty.
2013 const int jlen = ciEntry.cj_ind_end - ciEntry.cj_ind_start;
2016 sort_cj_excl(nbl->cj.data() + ciEntry.cj_ind_start, jlen, nbl->work.get());
2018 /* The counts below are used for non-bonded pair/flop counts
2019 * and should therefore match the available kernel setups.
2021 if (!(ciEntry.shift & NBNXN_CI_DO_COUL(0)))
2023 nbl->work->ncj_noq += jlen;
2025 else if ((ciEntry.shift & NBNXN_CI_HALF_LJ(0)) || !(ciEntry.shift & NBNXN_CI_DO_LJ(0)))
2027 nbl->work->ncj_hlj += jlen;
2032 /* Entry is empty: remove it */
2037 /* Split sci entry for load balancing on the GPU.
2038 * Splitting ensures we have enough lists to fully utilize the whole GPU.
2039 * With progBal we generate progressively smaller lists, which improves
2040 * load balancing. As we only know the current count on our own thread,
2041 * we will need to estimate the current total amount of i-entries.
2042 * As the lists get concatenated later, this estimate depends
2043 * both on nthread and our own thread index.
2045 static void split_sci_entry(NbnxnPairlistGpu* nbl,
2053 int nsp_max = nsp_target_av;
2057 /* Estimate the total numbers of ci's of the nblist combined
2058 * over all threads using the target number of ci's.
2060 float nsp_est = (nsp_tot_est * thread) / nthread + nbl->nci_tot;
2062 /* The first ci blocks should be larger, to avoid overhead.
2063 * The last ci blocks should be smaller, to improve load balancing.
2064 * The factor 3/2 makes the first block 3/2 times the target average
2065 * and ensures that the total number of blocks end up equal to
2066 * that of equally sized blocks of size nsp_target_av.
2068 nsp_max = static_cast<int>(nsp_target_av * (nsp_tot_est * 1.5 / (nsp_est + nsp_tot_est)));
2071 const int cj4_start = nbl->sci.back().cj4_ind_start;
2072 const int cj4_end = nbl->sci.back().cj4_ind_end;
2073 const int j4len = cj4_end - cj4_start;
2075 if (j4len > 1 && j4len * c_gpuNumClusterPerCell * c_nbnxnGpuJgroupSize > nsp_max)
2077 /* Modify the last ci entry and process the cj4's again */
2083 for (int cj4 = cj4_start; cj4 < cj4_end; cj4++)
2085 int nsp_cj4_p = nsp_cj4;
2086 /* Count the number of cluster pairs in this cj4 group */
2088 for (int p = 0; p < c_gpuNumClusterPerCell * c_nbnxnGpuJgroupSize; p++)
2090 nsp_cj4 += (nbl->cj4[cj4].imei[0].imask >> p) & 1;
2093 /* If adding the current cj4 with nsp_cj4 pairs get us further
2094 * away from our target nsp_max, split the list before this cj4.
2096 if (nsp > 0 && nsp_max - nsp < nsp + nsp_cj4 - nsp_max)
2098 /* Split the list at cj4 */
2099 nbl->sci.back().cj4_ind_end = cj4;
2100 /* Create a new sci entry */
2102 sciNew.sci = nbl->sci.back().sci;
2103 sciNew.shift = nbl->sci.back().shift;
2104 sciNew.cj4_ind_start = cj4;
2105 nbl->sci.push_back(sciNew);
2108 nsp_cj4_e = nsp_cj4_p;
2114 /* Put the remaining cj4's in the last sci entry */
2115 nbl->sci.back().cj4_ind_end = cj4_end;
2117 /* Possibly balance out the last two sci's
2118 * by moving the last cj4 of the second last sci.
2120 if (nsp_sci - nsp_cj4_e >= nsp + nsp_cj4_e)
2122 GMX_ASSERT(nbl->sci.size() >= 2, "We expect at least two elements");
2123 nbl->sci[nbl->sci.size() - 2].cj4_ind_end--;
2124 nbl->sci[nbl->sci.size() - 1].cj4_ind_start--;
2129 /* Clost this super/sub list i entry */
2130 static void closeIEntry(NbnxnPairlistGpu* nbl, int nsp_max_av, gmx_bool progBal, float nsp_tot_est, int thread, int nthread)
2132 nbnxn_sci_t& sciEntry = *getOpenIEntry(nbl);
2134 /* All content of the new ci entry have already been filled correctly,
2135 * we only need to, potentially, split or remove the entry when empty.
2137 int j4len = sciEntry.numJClusterGroups();
2140 /* We can only have complete blocks of 4 j-entries in a list,
2141 * so round the count up before closing.
2143 int ncj4 = (nbl->work->cj_ind + c_nbnxnGpuJgroupSize - 1) / c_nbnxnGpuJgroupSize;
2144 nbl->work->cj_ind = ncj4 * c_nbnxnGpuJgroupSize;
2148 /* Measure the size of the new entry and potentially split it */
2149 split_sci_entry(nbl, nsp_max_av, progBal, nsp_tot_est, thread, nthread);
2154 /* Entry is empty: remove it */
2155 nbl->sci.pop_back();
2159 /* Syncs the working array before adding another grid pair to the GPU list */
2160 static void sync_work(NbnxnPairlistCpu gmx_unused* nbl) {}
2162 /* Syncs the working array before adding another grid pair to the GPU list */
2163 static void sync_work(NbnxnPairlistGpu* nbl)
2165 nbl->work->cj_ind = nbl->cj4.size() * c_nbnxnGpuJgroupSize;
2168 /* Clears an NbnxnPairlistCpu data structure */
2169 static void clear_pairlist(NbnxnPairlistCpu* nbl)
2175 nbl->ciOuter.clear();
2176 nbl->cjOuter.clear();
2178 nbl->work->ncj_noq = 0;
2179 nbl->work->ncj_hlj = 0;
2182 /* Clears an NbnxnPairlistGpu data structure */
2183 static void clear_pairlist(NbnxnPairlistGpu* nbl)
2187 nbl->excl.resize(1);
2191 /* Clears an atom-atom-style pair list */
2192 static void clear_pairlist_fep(t_nblist* nl)
2196 if (nl->jindex.empty())
2198 nl->jindex.resize(1);
2203 /* Sets a simple list i-cell bounding box, including PBC shift */
2205 set_icell_bb_simple(gmx::ArrayRef<const BoundingBox> bb, int ci, real shx, real shy, real shz, BoundingBox* bb_ci)
2207 bb_ci->lower.x = bb[ci].lower.x + shx;
2208 bb_ci->lower.y = bb[ci].lower.y + shy;
2209 bb_ci->lower.z = bb[ci].lower.z + shz;
2210 bb_ci->upper.x = bb[ci].upper.x + shx;
2211 bb_ci->upper.y = bb[ci].upper.y + shy;
2212 bb_ci->upper.z = bb[ci].upper.z + shz;
2215 /* Sets a simple list i-cell bounding box, including PBC shift */
2216 static inline void set_icell_bb(const Grid& iGrid, int ci, real shx, real shy, real shz, NbnxnPairlistCpuWork* work)
2218 set_icell_bb_simple(iGrid.iBoundingBoxes(), ci, shx, shy, shz, &work->iClusterData.bb[0]);
2222 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
2223 static void set_icell_bbxxxx_supersub(gmx::ArrayRef<const float> bb, int ci, real shx, real shy, real shz, float* bb_ci)
2225 constexpr int cellBBStride = packedBoundingBoxesIndex(c_gpuNumClusterPerCell);
2226 constexpr int pbbStride = c_packedBoundingBoxesDimSize;
2227 const int ia = ci * cellBBStride;
2228 for (int m = 0; m < cellBBStride; m += c_packedBoundingBoxesSize)
2230 for (int i = 0; i < pbbStride; i++)
2232 bb_ci[m + 0 * pbbStride + i] = bb[ia + m + 0 * pbbStride + i] + shx;
2233 bb_ci[m + 1 * pbbStride + i] = bb[ia + m + 1 * pbbStride + i] + shy;
2234 bb_ci[m + 2 * pbbStride + i] = bb[ia + m + 2 * pbbStride + i] + shz;
2235 bb_ci[m + 3 * pbbStride + i] = bb[ia + m + 3 * pbbStride + i] + shx;
2236 bb_ci[m + 4 * pbbStride + i] = bb[ia + m + 4 * pbbStride + i] + shy;
2237 bb_ci[m + 5 * pbbStride + i] = bb[ia + m + 5 * pbbStride + i] + shz;
2243 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
2244 gmx_unused static void set_icell_bb_supersub(gmx::ArrayRef<const BoundingBox> bb,
2251 for (int i = 0; i < c_gpuNumClusterPerCell; i++)
2253 set_icell_bb_simple(bb, ci * c_gpuNumClusterPerCell + i, shx, shy, shz, &bb_ci[i]);
2257 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
2258 gmx_unused static void set_icell_bb(const Grid& iGrid, int ci, real shx, real shy, real shz, NbnxnPairlistGpuWork* work)
2261 set_icell_bbxxxx_supersub(
2262 iGrid.packedBoundingBoxes(), ci, shx, shy, shz, work->iSuperClusterData.bbPacked.data());
2264 set_icell_bb_supersub(iGrid.iBoundingBoxes(), ci, shx, shy, shz, work->iSuperClusterData.bb.data());
2268 /* Copies PBC shifted i-cell atom coordinates x,y,z to working array */
2269 static void icell_set_x_simple(int ci,
2275 NbnxnPairlistCpuWork::IClusterData* iClusterData)
2277 const int ia = ci * c_nbnxnCpuIClusterSize;
2279 for (int i = 0; i < c_nbnxnCpuIClusterSize; i++)
2281 iClusterData->x[i * STRIDE_XYZ + XX] = x[(ia + i) * stride + XX] + shx;
2282 iClusterData->x[i * STRIDE_XYZ + YY] = x[(ia + i) * stride + YY] + shy;
2283 iClusterData->x[i * STRIDE_XYZ + ZZ] = x[(ia + i) * stride + ZZ] + shz;
2287 static void icell_set_x(int ci,
2293 const ClusterDistanceKernelType kernelType,
2294 NbnxnPairlistCpuWork* work)
2299 # ifdef GMX_NBNXN_SIMD_4XN
2300 case ClusterDistanceKernelType::CpuSimd_4xM:
2301 icell_set_x_simd_4xn(ci, shx, shy, shz, stride, x, work);
2304 # ifdef GMX_NBNXN_SIMD_2XNN
2305 case ClusterDistanceKernelType::CpuSimd_2xMM:
2306 icell_set_x_simd_2xnn(ci, shx, shy, shz, stride, x, work);
2310 case ClusterDistanceKernelType::CpuPlainC:
2311 icell_set_x_simple(ci, shx, shy, shz, stride, x, &work->iClusterData);
2313 default: GMX_ASSERT(false, "Unhandled case"); break;
2317 /* Copies PBC shifted super-cell atom coordinates x,y,z to working array */
2318 static void icell_set_x(int ci,
2324 ClusterDistanceKernelType gmx_unused kernelType,
2325 NbnxnPairlistGpuWork* work)
2327 #if !GMX_SIMD4_HAVE_REAL
2329 real* x_ci = work->iSuperClusterData.x.data();
2331 int ia = ci * c_gpuNumClusterPerCell * c_nbnxnGpuClusterSize;
2332 for (int i = 0; i < c_gpuNumClusterPerCell * c_nbnxnGpuClusterSize; i++)
2334 x_ci[i * DIM + XX] = x[(ia + i) * stride + XX] + shx;
2335 x_ci[i * DIM + YY] = x[(ia + i) * stride + YY] + shy;
2336 x_ci[i * DIM + ZZ] = x[(ia + i) * stride + ZZ] + shz;
2339 #else /* !GMX_SIMD4_HAVE_REAL */
2341 real* x_ci = work->iSuperClusterData.xSimd.data();
2343 for (int si = 0; si < c_gpuNumClusterPerCell; si++)
2345 for (int i = 0; i < c_nbnxnGpuClusterSize; i += GMX_SIMD4_WIDTH)
2347 int io = si * c_nbnxnGpuClusterSize + i;
2348 int ia = ci * c_gpuNumClusterPerCell * c_nbnxnGpuClusterSize + io;
2349 for (int j = 0; j < GMX_SIMD4_WIDTH; j++)
2351 x_ci[io * DIM + j + XX * GMX_SIMD4_WIDTH] = x[(ia + j) * stride + XX] + shx;
2352 x_ci[io * DIM + j + YY * GMX_SIMD4_WIDTH] = x[(ia + j) * stride + YY] + shy;
2353 x_ci[io * DIM + j + ZZ * GMX_SIMD4_WIDTH] = x[(ia + j) * stride + ZZ] + shz;
2358 #endif /* !GMX_SIMD4_HAVE_REAL */
2361 static real minimum_subgrid_size_xy(const Grid& grid)
2363 const Grid::Dimensions& dims = grid.dimensions();
2365 if (grid.geometry().isSimple)
2367 return std::min(dims.cellSize[XX], dims.cellSize[YY]);
2371 return std::min(dims.cellSize[XX] / c_gpuNumClusterPerCellX,
2372 dims.cellSize[YY] / c_gpuNumClusterPerCellY);
2376 static real effective_buffer_1x1_vs_MxN(const Grid& iGrid, const Grid& jGrid)
2378 const real eff_1x1_buffer_fac_overest = 0.1;
2380 /* Determine an atom-pair list cut-off buffer size for atom pairs,
2381 * to be added to rlist (including buffer) used for MxN.
2382 * This is for converting an MxN list to a 1x1 list. This means we can't
2383 * use the normal buffer estimate, as we have an MxN list in which
2384 * some atom pairs beyond rlist are missing. We want to capture
2385 * the beneficial effect of buffering by extra pairs just outside rlist,
2386 * while removing the useless pairs that are further away from rlist.
2387 * (Also the buffer could have been set manually not using the estimate.)
2388 * This buffer size is an overestimate.
2389 * We add 10% of the smallest grid sub-cell dimensions.
2390 * Note that the z-size differs per cell and we don't use this,
2391 * so we overestimate.
2392 * With PME, the 10% value gives a buffer that is somewhat larger
2393 * than the effective buffer with a tolerance of 0.005 kJ/mol/ps.
2394 * Smaller tolerances or using RF lead to a smaller effective buffer,
2395 * so 10% gives a safe overestimate.
2397 return eff_1x1_buffer_fac_overest * (minimum_subgrid_size_xy(iGrid) + minimum_subgrid_size_xy(jGrid));
2400 /* Estimates the interaction volume^2 for non-local interactions */
2401 static real nonlocal_vol2(const struct gmx_domdec_zones_t* zones, const rvec ls, real r)
2403 real vol2_est_tot = 0;
2405 /* Here we simply add up the volumes of 1, 2 or 3 1D decomposition
2406 * not home interaction volume^2. As these volumes are not additive,
2407 * this is an overestimate, but it would only be significant in the limit
2408 * of small cells, where we anyhow need to split the lists into
2409 * as small parts as possible.
2412 for (int z = 0; z < zones->n; z++)
2414 if (zones->shift[z][XX] + zones->shift[z][YY] + zones->shift[z][ZZ] == 1)
2419 for (int d = 0; d < DIM; d++)
2421 if (zones->shift[z][d] == 0)
2425 za *= zones->size[z].x1[d] - zones->size[z].x0[d];
2429 /* 4 octants of a sphere */
2430 real vold_est = 0.25 * M_PI * r * r * r * r;
2431 /* 4 quarter pie slices on the edges */
2432 vold_est += 4 * cl * M_PI / 6.0 * r * r * r;
2433 /* One rectangular volume on a face */
2434 vold_est += ca * 0.5 * r * r;
2436 vol2_est_tot += vold_est * za;
2440 return vol2_est_tot;
2443 /* Estimates the average size of a full j-list for super/sub setup */
2444 static void get_nsubpair_target(const Nbnxm::GridSet& gridSet,
2445 const InteractionLocality iloc,
2447 const int min_ci_balanced,
2448 int* nsubpair_target,
2449 float* nsubpair_tot_est)
2451 /* The target value of 36 seems to be the optimum for Kepler.
2452 * Maxwell is less sensitive to the exact value.
2454 const int nsubpair_target_min = 36;
2456 const Grid& grid = gridSet.grids()[0];
2458 /* We don't need to balance list sizes if:
2459 * - We didn't request balancing.
2460 * - The number of grid cells >= the number of lists requested,
2461 * since we will always generate at least #cells lists.
2462 * - We don't have any cells, since then there won't be any lists.
2464 if (min_ci_balanced <= 0 || grid.numCells() >= min_ci_balanced || grid.numCells() == 0)
2466 /* nsubpair_target==0 signals no balancing */
2467 *nsubpair_target = 0;
2468 *nsubpair_tot_est = 0;
2474 const int numAtomsCluster = grid.geometry().numAtomsICluster;
2475 const Grid::Dimensions& dims = grid.dimensions();
2477 ls[XX] = dims.cellSize[XX] / c_gpuNumClusterPerCellX;
2478 ls[YY] = dims.cellSize[YY] / c_gpuNumClusterPerCellY;
2479 ls[ZZ] = numAtomsCluster / (dims.atomDensity * ls[XX] * ls[YY]);
2481 /* The formulas below are a heuristic estimate of the average nsj per si*/
2482 const real r_eff_sup = rlist + nbnxn_get_rlist_effective_inc(numAtomsCluster, ls);
2484 real nsp_est_nl = 0;
2485 if (gridSet.domainSetup().haveMultipleDomains && gridSet.domainSetup().zones->n != 1)
2487 nsp_est_nl = gmx::square(dims.atomDensity / numAtomsCluster)
2488 * nonlocal_vol2(gridSet.domainSetup().zones, ls, r_eff_sup);
2491 real nsp_est = nsp_est_nl;
2492 if (iloc == InteractionLocality::Local)
2494 /* Sub-cell interacts with itself */
2495 real vol_est = ls[XX] * ls[YY] * ls[ZZ];
2496 /* 6/2 rectangular volume on the faces */
2497 vol_est += (ls[XX] * ls[YY] + ls[XX] * ls[ZZ] + ls[YY] * ls[ZZ]) * r_eff_sup;
2498 /* 12/2 quarter pie slices on the edges */
2499 vol_est += 2 * (ls[XX] + ls[YY] + ls[ZZ]) * 0.25 * M_PI * gmx::square(r_eff_sup);
2500 /* 4 octants of a sphere */
2501 vol_est += 0.5 * 4.0 / 3.0 * M_PI * gmx::power3(r_eff_sup);
2503 /* Estimate the number of cluster pairs as the local number of
2504 * clusters times the volume they interact with times the density.
2506 nsp_est = grid.numClusters() * vol_est * dims.atomDensity / numAtomsCluster;
2508 /* Subtract the non-local pair count */
2509 nsp_est -= nsp_est_nl;
2511 /* For small cut-offs nsp_est will be an underestimate.
2512 * With DD nsp_est_nl is an overestimate so nsp_est can get negative.
2513 * So to avoid too small or negative nsp_est we set a minimum of
2514 * all cells interacting with all 3^3 direct neighbors (3^3-1)/2+1=14.
2515 * This might be a slight overestimate for small non-periodic groups of
2516 * atoms as will occur for a local domain with DD, but for small
2517 * groups of atoms we'll anyhow be limited by nsubpair_target_min,
2518 * so this overestimation will not matter.
2520 nsp_est = std::max(nsp_est, grid.numClusters() * 14._real);
2524 fprintf(debug, "nsp_est local %5.1f non-local %5.1f\n", nsp_est, nsp_est_nl);
2528 /* Thus the (average) maximum j-list size should be as follows.
2529 * Since there is overhead, we shouldn't make the lists too small
2530 * (and we can't chop up j-groups) so we use a minimum target size of 36.
2532 *nsubpair_target = std::max(nsubpair_target_min, roundToInt(nsp_est / min_ci_balanced));
2533 *nsubpair_tot_est = nsp_est;
2537 fprintf(debug, "nbl nsp estimate %.1f, nsubpair_target %d\n", nsp_est, *nsubpair_target);
2541 /* Debug list print function */
2542 static void print_nblist_ci_cj(FILE* fp, const NbnxnPairlistCpu& nbl)
2544 for (const nbnxn_ci_t& ciEntry : nbl.ci)
2546 fprintf(fp, "ci %4d shift %2d ncj %3d\n", ciEntry.ci, ciEntry.shift, ciEntry.cj_ind_end - ciEntry.cj_ind_start);
2548 for (int j = ciEntry.cj_ind_start; j < ciEntry.cj_ind_end; j++)
2550 fprintf(fp, " cj %5d imask %x\n", nbl.cj[j].cj, nbl.cj[j].excl);
2555 /* Debug list print function */
2556 static void print_nblist_sci_cj(FILE* fp, const NbnxnPairlistGpu& nbl)
2558 for (const nbnxn_sci_t& sci : nbl.sci)
2560 fprintf(fp, "ci %4d shift %2d ncj4 %2d\n", sci.sci, sci.shift, sci.numJClusterGroups());
2563 for (int j4 = sci.cj4_ind_start; j4 < sci.cj4_ind_end; j4++)
2565 for (int j = 0; j < c_nbnxnGpuJgroupSize; j++)
2567 fprintf(fp, " sj %5d imask %x\n", nbl.cj4[j4].cj[j], nbl.cj4[j4].imei[0].imask);
2568 for (int si = 0; si < c_gpuNumClusterPerCell; si++)
2570 if (nbl.cj4[j4].imei[0].imask & (1U << (j * c_gpuNumClusterPerCell + si)))
2577 fprintf(fp, "ci %4d shift %2d ncj4 %2d ncp %3d\n", sci.sci, sci.shift, sci.numJClusterGroups(), ncp);
2581 /* Combine pair lists *nbl generated on multiple threads nblc */
2582 static void combine_nblists(gmx::ArrayRef<const NbnxnPairlistGpu> nbls, NbnxnPairlistGpu* nblc)
2584 int nsci = nblc->sci.size();
2585 int ncj4 = nblc->cj4.size();
2586 int nexcl = nblc->excl.size();
2587 for (const auto& nbl : nbls)
2589 nsci += nbl.sci.size();
2590 ncj4 += nbl.cj4.size();
2591 nexcl += nbl.excl.size();
2594 /* Resize with the final, combined size, so we can fill in parallel */
2595 /* NOTE: For better performance we should use default initialization */
2596 nblc->sci.resize(nsci);
2597 nblc->cj4.resize(ncj4);
2598 nblc->excl.resize(nexcl);
2600 /* Each thread should copy its own data to the combined arrays,
2601 * as otherwise data will go back and forth between different caches.
2603 const int gmx_unused nthreads = gmx_omp_nthreads_get(emntPairsearch);
2605 #pragma omp parallel for num_threads(nthreads) schedule(static)
2606 for (gmx::index n = 0; n < nbls.ssize(); n++)
2610 /* Determine the offset in the combined data for our thread.
2611 * Note that the original sizes in nblc are lost.
2613 int sci_offset = nsci;
2614 int cj4_offset = ncj4;
2615 int excl_offset = nexcl;
2617 for (gmx::index i = n; i < nbls.ssize(); i++)
2619 sci_offset -= nbls[i].sci.size();
2620 cj4_offset -= nbls[i].cj4.size();
2621 excl_offset -= nbls[i].excl.size();
2624 const NbnxnPairlistGpu& nbli = nbls[n];
2626 for (size_t i = 0; i < nbli.sci.size(); i++)
2628 nblc->sci[sci_offset + i] = nbli.sci[i];
2629 nblc->sci[sci_offset + i].cj4_ind_start += cj4_offset;
2630 nblc->sci[sci_offset + i].cj4_ind_end += cj4_offset;
2633 for (size_t j4 = 0; j4 < nbli.cj4.size(); j4++)
2635 nblc->cj4[cj4_offset + j4] = nbli.cj4[j4];
2636 nblc->cj4[cj4_offset + j4].imei[0].excl_ind += excl_offset;
2637 nblc->cj4[cj4_offset + j4].imei[1].excl_ind += excl_offset;
2640 for (size_t j4 = 0; j4 < nbli.excl.size(); j4++)
2642 nblc->excl[excl_offset + j4] = nbli.excl[j4];
2645 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
2648 for (const auto& nbl : nbls)
2650 nblc->nci_tot += nbl.nci_tot;
2654 static void balance_fep_lists(gmx::ArrayRef<std::unique_ptr<t_nblist>> fepLists,
2655 gmx::ArrayRef<PairsearchWork> work)
2657 const int numLists = fepLists.ssize();
2661 /* Nothing to balance */
2665 /* Count the total i-lists and pairs */
2668 for (const auto& list : fepLists)
2670 nri_tot += list->nri;
2671 nrj_tot += list->nrj;
2674 const int nrj_target = (nrj_tot + numLists - 1) / numLists;
2676 GMX_ASSERT(gmx_omp_nthreads_get(emntNonbonded) == numLists,
2677 "We should have as many work objects as FEP lists");
2679 #pragma omp parallel for schedule(static) num_threads(numLists)
2680 for (int th = 0; th < numLists; th++)
2684 t_nblist* nbl = work[th].nbl_fep.get();
2686 /* Note that here we allocate for the total size, instead of
2687 * a per-thread esimate (which is hard to obtain).
2689 if (nri_tot > nbl->maxnri)
2691 nbl->maxnri = over_alloc_large(nri_tot);
2692 reallocate_nblist(nbl);
2694 if (nri_tot > nbl->maxnri || nrj_tot > nbl->maxnrj)
2696 nbl->maxnrj = over_alloc_small(nrj_tot);
2697 nbl->jjnr.resize(nbl->maxnrj);
2698 nbl->excl_fep.resize(nbl->maxnrj);
2701 clear_pairlist_fep(nbl);
2703 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
2706 /* Loop over the source lists and assign and copy i-entries */
2708 t_nblist* nbld = work[th_dest].nbl_fep.get();
2709 for (int th = 0; th < numLists; th++)
2711 const t_nblist* nbls = fepLists[th].get();
2713 for (int i = 0; i < nbls->nri; i++)
2715 /* The number of pairs in this i-entry */
2716 const int nrj = nbls->jindex[i + 1] - nbls->jindex[i];
2718 /* Decide if list th_dest is too large and we should procede
2719 * to the next destination list.
2721 if (th_dest + 1 < numLists && nbld->nrj > 0
2722 && nbld->nrj + nrj - nrj_target > nrj_target - nbld->nrj)
2725 nbld = work[th_dest].nbl_fep.get();
2728 nbld->iinr[nbld->nri] = nbls->iinr[i];
2729 nbld->gid[nbld->nri] = nbls->gid[i];
2730 nbld->shift[nbld->nri] = nbls->shift[i];
2732 for (int j = nbls->jindex[i]; j < nbls->jindex[i + 1]; j++)
2734 nbld->jjnr[nbld->nrj] = nbls->jjnr[j];
2735 nbld->excl_fep[nbld->nrj] = nbls->excl_fep[j];
2739 nbld->jindex[nbld->nri] = nbld->nrj;
2743 /* Swap the list pointers */
2744 for (int th = 0; th < numLists; th++)
2746 fepLists[th].swap(work[th].nbl_fep);
2750 fprintf(debug, "nbl_fep[%d] nri %4d nrj %4d\n", th, fepLists[th]->nri, fepLists[th]->nrj);
2755 /* Returns the next ci to be processes by our thread */
2756 static gmx_bool next_ci(const Grid& grid, int nth, int ci_block, int* ci_x, int* ci_y, int* ci_b, int* ci)
2761 if (*ci_b == ci_block)
2763 /* Jump to the next block assigned to this task */
2764 *ci += (nth - 1) * ci_block;
2768 if (*ci >= grid.numCells())
2773 while (*ci >= grid.firstCellInColumn(*ci_x * grid.dimensions().numCells[YY] + *ci_y + 1))
2776 if (*ci_y == grid.dimensions().numCells[YY])
2786 /* Returns the distance^2 for which we put cell pairs in the list
2787 * without checking atom pair distances. This is usually < rlist^2.
2789 static float boundingbox_only_distance2(const Grid::Dimensions& iGridDims,
2790 const Grid::Dimensions& jGridDims,
2794 /* If the distance between two sub-cell bounding boxes is less
2795 * than this distance, do not check the distance between
2796 * all particle pairs in the sub-cell, since then it is likely
2797 * that the box pair has atom pairs within the cut-off.
2798 * We use the nblist cut-off minus 0.5 times the average x/y diagonal
2799 * spacing of the sub-cells. Around 40% of the checked pairs are pruned.
2800 * Using more than 0.5 gains at most 0.5%.
2801 * If forces are calculated more than twice, the performance gain
2802 * in the force calculation outweighs the cost of checking.
2803 * Note that with subcell lists, the atom-pair distance check
2804 * is only performed when only 1 out of 8 sub-cells in within range,
2805 * this is because the GPU is much faster than the cpu.
2808 real bbx = 0.5 * (iGridDims.cellSize[XX] + jGridDims.cellSize[XX]);
2809 real bby = 0.5 * (iGridDims.cellSize[YY] + jGridDims.cellSize[YY]);
2812 bbx /= c_gpuNumClusterPerCellX;
2813 bby /= c_gpuNumClusterPerCellY;
2816 real rbb2 = std::max(0.0, rlist - 0.5 * std::sqrt(bbx * bbx + bby * bby));
2822 return static_cast<float>((1 + GMX_FLOAT_EPS) * rbb2);
2826 static int get_ci_block_size(const Grid& iGrid, const bool haveMultipleDomains, const int numLists)
2828 const int ci_block_enum = 5;
2829 const int ci_block_denom = 11;
2830 const int ci_block_min_atoms = 16;
2832 /* Here we decide how to distribute the blocks over the threads.
2833 * We use prime numbers to try to avoid that the grid size becomes
2834 * a multiple of the number of threads, which would lead to some
2835 * threads getting "inner" pairs and others getting boundary pairs,
2836 * which in turns will lead to load imbalance between threads.
2837 * Set the block size as 5/11/ntask times the average number of cells
2838 * in a y,z slab. This should ensure a quite uniform distribution
2839 * of the grid parts of the different thread along all three grid
2840 * zone boundaries with 3D domain decomposition. At the same time
2841 * the blocks will not become too small.
2843 GMX_ASSERT(iGrid.dimensions().numCells[XX] > 0, "Grid can't be empty");
2844 GMX_ASSERT(numLists > 0, "We need at least one list");
2845 int ci_block = (iGrid.numCells() * ci_block_enum)
2846 / (ci_block_denom * iGrid.dimensions().numCells[XX] * numLists);
2848 const int numAtomsPerCell = iGrid.geometry().numAtomsPerCell;
2850 /* Ensure the blocks are not too small: avoids cache invalidation */
2851 if (ci_block * numAtomsPerCell < ci_block_min_atoms)
2853 ci_block = (ci_block_min_atoms + numAtomsPerCell - 1) / numAtomsPerCell;
2856 /* Without domain decomposition
2857 * or with less than 3 blocks per task, divide in nth blocks.
2859 if (!haveMultipleDomains || numLists * 3 * ci_block > iGrid.numCells())
2861 ci_block = (iGrid.numCells() + numLists - 1) / numLists;
2864 if (ci_block > 1 && (numLists - 1) * ci_block >= iGrid.numCells())
2866 /* Some threads have no work. Although reducing the block size
2867 * does not decrease the block count on the first few threads,
2868 * with GPUs better mixing of "upper" cells that have more empty
2869 * clusters results in a somewhat lower max load over all threads.
2870 * Without GPUs the regime of so few atoms per thread is less
2871 * performance relevant, but with 8-wide SIMD the same reasoning
2872 * applies, since the pair list uses 4 i-atom "sub-clusters".
2880 /* Returns the number of bits to right-shift a cluster index to obtain
2881 * the corresponding force buffer flag index.
2883 static int getBufferFlagShift(int numAtomsPerCluster)
2885 int bufferFlagShift = 0;
2886 while ((numAtomsPerCluster << bufferFlagShift) < NBNXN_BUFFERFLAG_SIZE)
2891 return bufferFlagShift;
2894 static bool pairlistIsSimple(const NbnxnPairlistCpu gmx_unused& pairlist)
2899 static bool pairlistIsSimple(const NbnxnPairlistGpu gmx_unused& pairlist)
2904 static void makeClusterListWrapper(NbnxnPairlistCpu* nbl,
2905 const Grid gmx_unused& iGrid,
2908 const int firstCell,
2910 const bool excludeSubDiagonal,
2911 const nbnxn_atomdata_t* nbat,
2914 const ClusterDistanceKernelType kernelType,
2915 int* numDistanceChecks)
2919 case ClusterDistanceKernelType::CpuPlainC:
2920 makeClusterListSimple(
2921 jGrid, nbl, ci, firstCell, lastCell, excludeSubDiagonal, nbat->x().data(), rlist2, rbb2, numDistanceChecks);
2923 #ifdef GMX_NBNXN_SIMD_4XN
2924 case ClusterDistanceKernelType::CpuSimd_4xM:
2925 makeClusterListSimd4xn(
2926 jGrid, nbl, ci, firstCell, lastCell, excludeSubDiagonal, nbat->x().data(), rlist2, rbb2, numDistanceChecks);
2929 #ifdef GMX_NBNXN_SIMD_2XNN
2930 case ClusterDistanceKernelType::CpuSimd_2xMM:
2931 makeClusterListSimd2xnn(
2932 jGrid, nbl, ci, firstCell, lastCell, excludeSubDiagonal, nbat->x().data(), rlist2, rbb2, numDistanceChecks);
2935 default: GMX_ASSERT(false, "Unhandled kernel type");
2939 static void makeClusterListWrapper(NbnxnPairlistGpu* nbl,
2940 const Grid& gmx_unused iGrid,
2943 const int firstCell,
2945 const bool excludeSubDiagonal,
2946 const nbnxn_atomdata_t* nbat,
2949 ClusterDistanceKernelType gmx_unused kernelType,
2950 int* numDistanceChecks)
2952 for (int cj = firstCell; cj <= lastCell; cj++)
2954 make_cluster_list_supersub(
2955 iGrid, jGrid, nbl, ci, cj, excludeSubDiagonal, nbat->xstride, nbat->x().data(), rlist2, rbb2, numDistanceChecks);
2959 static int getNumSimpleJClustersInList(const NbnxnPairlistCpu& nbl)
2961 return nbl.cj.size();
2964 static int getNumSimpleJClustersInList(const gmx_unused NbnxnPairlistGpu& nbl)
2969 static void incrementNumSimpleJClustersInList(NbnxnPairlistCpu* nbl, int ncj_old_j)
2971 nbl->ncjInUse += nbl->cj.size();
2972 nbl->ncjInUse -= ncj_old_j;
2975 static void incrementNumSimpleJClustersInList(NbnxnPairlistGpu gmx_unused* nbl, int gmx_unused ncj_old_j)
2979 static void checkListSizeConsistency(const NbnxnPairlistCpu& nbl, const bool haveFreeEnergy)
2981 GMX_RELEASE_ASSERT(static_cast<size_t>(nbl.ncjInUse) == nbl.cj.size() || haveFreeEnergy,
2982 "Without free-energy all cj pair-list entries should be in use. "
2983 "Note that subsequent code does not make use of the equality, "
2984 "this check is only here to catch bugs");
2987 static void checkListSizeConsistency(const NbnxnPairlistGpu gmx_unused& nbl, bool gmx_unused haveFreeEnergy)
2989 /* We currently can not check consistency here */
2992 /* Set the buffer flags for newly added entries in the list */
2993 static void setBufferFlags(const NbnxnPairlistCpu& nbl,
2994 const int ncj_old_j,
2995 const int gridj_flag_shift,
2996 gmx_bitmask_t* gridj_flag,
2999 if (gmx::ssize(nbl.cj) > ncj_old_j)
3001 int cbFirst = nbl.cj[ncj_old_j].cj >> gridj_flag_shift;
3002 int cbLast = nbl.cj.back().cj >> gridj_flag_shift;
3003 for (int cb = cbFirst; cb <= cbLast; cb++)
3005 bitmask_init_bit(&gridj_flag[cb], th);
3010 static void setBufferFlags(const NbnxnPairlistGpu gmx_unused& nbl,
3011 int gmx_unused ncj_old_j,
3012 int gmx_unused gridj_flag_shift,
3013 gmx_bitmask_t gmx_unused* gridj_flag,
3016 GMX_ASSERT(false, "This function should never be called");
3019 /* Generates the part of pair-list nbl assigned to our thread */
3020 template<typename T>
3021 static void nbnxn_make_pairlist_part(const Nbnxm::GridSet& gridSet,
3024 PairsearchWork* work,
3025 const nbnxn_atomdata_t* nbat,
3026 const ListOfLists<int>& exclusions,
3028 const PairlistType pairlistType,
3030 gmx_bool bFBufferFlag,
3033 float nsubpair_tot_est,
3042 int gridi_flag_shift = 0, gridj_flag_shift = 0;
3043 gmx_bitmask_t* gridj_flag = nullptr;
3045 if (jGrid.geometry().isSimple != pairlistIsSimple(*nbl)
3046 || iGrid.geometry().isSimple != pairlistIsSimple(*nbl))
3048 gmx_incons("Grid incompatible with pair-list");
3052 GMX_ASSERT(nbl->na_ci == jGrid.geometry().numAtomsICluster,
3053 "The cluster sizes in the list and grid should match");
3054 nbl->na_cj = JClusterSizePerListType[pairlistType];
3055 const int na_cj_2log = get_2log(nbl->na_cj);
3061 /* Determine conversion of clusters to flag blocks */
3062 gridi_flag_shift = getBufferFlagShift(nbl->na_ci);
3063 gridj_flag_shift = getBufferFlagShift(nbl->na_cj);
3065 gridj_flag = work->buffer_flags.data();
3068 gridSet.getBox(box);
3070 const bool haveFep = gridSet.haveFep();
3072 const real rlist2 = nbl->rlist * nbl->rlist;
3074 // Select the cluster pair distance kernel type
3075 const ClusterDistanceKernelType kernelType = getClusterDistanceKernelType(pairlistType, *nbat);
3077 if (haveFep && !pairlistIsSimple(*nbl))
3079 /* Determine an atom-pair list cut-off distance for FEP atom pairs.
3080 * We should not simply use rlist, since then we would not have
3081 * the small, effective buffering of the NxN lists.
3082 * The buffer is on overestimate, but the resulting cost for pairs
3083 * beyond rlist is negligible compared to the FEP pairs within rlist.
3085 rl_fep2 = nbl->rlist + effective_buffer_1x1_vs_MxN(iGrid, jGrid);
3089 fprintf(debug, "nbl_fep atom-pair rlist %f\n", rl_fep2);
3091 rl_fep2 = rl_fep2 * rl_fep2;
3094 const Grid::Dimensions& iGridDims = iGrid.dimensions();
3095 const Grid::Dimensions& jGridDims = jGrid.dimensions();
3098 boundingbox_only_distance2(iGridDims, jGridDims, nbl->rlist, pairlistIsSimple(*nbl));
3102 fprintf(debug, "nbl bounding box only distance %f\n", std::sqrt(rbb2));
3105 const bool isIntraGridList = (&iGrid == &jGrid);
3107 /* Set the shift range */
3108 for (int d = 0; d < DIM; d++)
3110 /* Check if we need periodicity shifts.
3111 * Without PBC or with domain decomposition we don't need them.
3113 if (d >= numPbcDimensions(gridSet.domainSetup().pbcType)
3114 || gridSet.domainSetup().haveMultipleDomainsPerDim[d])
3120 const real listRangeCellToCell =
3121 listRangeForGridCellToGridCell(rlist, iGrid.dimensions(), jGrid.dimensions());
3122 if (d == XX && box[XX][XX] - fabs(box[YY][XX]) - fabs(box[ZZ][XX]) < listRangeCellToCell)
3132 const bool bSimple = pairlistIsSimple(*nbl);
3133 gmx::ArrayRef<const BoundingBox> bb_i;
3135 gmx::ArrayRef<const float> pbb_i;
3138 bb_i = iGrid.iBoundingBoxes();
3142 pbb_i = iGrid.packedBoundingBoxes();
3145 /* We use the normal bounding box format for both grid types */
3146 bb_i = iGrid.iBoundingBoxes();
3148 gmx::ArrayRef<const BoundingBox1D> bbcz_i = iGrid.zBoundingBoxes();
3149 gmx::ArrayRef<const int> flags_i = iGrid.clusterFlags();
3150 gmx::ArrayRef<const BoundingBox1D> bbcz_j = jGrid.zBoundingBoxes();
3151 int cell0_i = iGrid.cellOffset();
3156 "nbl nc_i %d col.av. %.1f ci_block %d\n",
3158 iGrid.numCells() / static_cast<double>(iGrid.numColumns()),
3162 int numDistanceChecks = 0;
3164 const real listRangeBBToJCell2 =
3165 gmx::square(listRangeForBoundingBoxToGridCell(rlist, jGrid.dimensions()));
3167 /* Initially ci_b and ci to 1 before where we want them to start,
3168 * as they will both be incremented in next_ci.
3171 int ci = th * ci_block - 1;
3174 while (next_ci(iGrid, nth, ci_block, &ci_x, &ci_y, &ci_b, &ci))
3176 if (bSimple && flags_i[ci] == 0)
3180 const int ncj_old_i = getNumSimpleJClustersInList(*nbl);
3183 if (!isIntraGridList && shp[XX] == 0)
3186 bSimple ? bb_i[ci].upper.x
3187 : iGridDims.lowerCorner[XX] + (real(ci_x) + 1) * iGridDims.cellSize[XX];
3188 if (bx1 < jGridDims.lowerCorner[XX])
3190 d2cx = gmx::square(jGridDims.lowerCorner[XX] - bx1);
3192 if (d2cx >= listRangeBBToJCell2)
3199 int ci_xy = ci_x * iGridDims.numCells[YY] + ci_y;
3201 /* Loop over shift vectors in three dimensions */
3202 for (int tz = -shp[ZZ]; tz <= shp[ZZ]; tz++)
3204 const real shz = real(tz) * box[ZZ][ZZ];
3206 real bz0 = bbcz_i[ci].lower + shz;
3207 real bz1 = bbcz_i[ci].upper + shz;
3212 d2z = gmx::square(bz1);
3216 d2z = gmx::square(bz0 - box[ZZ][ZZ]);
3219 const real d2z_cx = d2z + d2cx;
3221 if (d2z_cx >= rlist2)
3226 real bz1_frac = bz1 / real(iGrid.numCellsInColumn(ci_xy));
3231 /* The check with bz1_frac close to or larger than 1 comes later */
3233 for (int ty = -shp[YY]; ty <= shp[YY]; ty++)
3235 const real shy = real(ty) * box[YY][YY] + real(tz) * box[ZZ][YY];
3237 const real by0 = bSimple ? bb_i[ci].lower.y + shy
3238 : iGridDims.lowerCorner[YY]
3239 + (real(ci_y)) * iGridDims.cellSize[YY] + shy;
3240 const real by1 = bSimple ? bb_i[ci].upper.y + shy
3241 : iGridDims.lowerCorner[YY]
3242 + (real(ci_y) + 1) * iGridDims.cellSize[YY] + shy;
3244 int cyf, cyl; //NOLINT(cppcoreguidelines-init-variables)
3245 get_cell_range<YY>(by0, by1, jGridDims, d2z_cx, rlist, &cyf, &cyl);
3253 if (by1 < jGridDims.lowerCorner[YY])
3255 d2z_cy += gmx::square(jGridDims.lowerCorner[YY] - by1);
3257 else if (by0 > jGridDims.upperCorner[YY])
3259 d2z_cy += gmx::square(by0 - jGridDims.upperCorner[YY]);
3262 for (int tx = -shp[XX]; tx <= shp[XX]; tx++)
3264 const int shift = XYZ2IS(tx, ty, tz);
3266 const bool excludeSubDiagonal = (isIntraGridList && shift == CENTRAL);
3268 if (c_pbcShiftBackward && isIntraGridList && shift > CENTRAL)
3274 real(tx) * box[XX][XX] + real(ty) * box[YY][XX] + real(tz) * box[ZZ][XX];
3276 const real bx0 = bSimple ? bb_i[ci].lower.x + shx
3277 : iGridDims.lowerCorner[XX]
3278 + (real(ci_x)) * iGridDims.cellSize[XX] + shx;
3279 const real bx1 = bSimple ? bb_i[ci].upper.x + shx
3280 : iGridDims.lowerCorner[XX]
3281 + (real(ci_x) + 1) * iGridDims.cellSize[XX] + shx;
3283 int cxf, cxl; //NOLINT(cppcoreguidelines-init-variables)
3284 get_cell_range<XX>(bx0, bx1, jGridDims, d2z_cy, rlist, &cxf, &cxl);
3291 addNewIEntry(nbl, cell0_i + ci, shift, flags_i[ci]);
3293 if ((!c_pbcShiftBackward || excludeSubDiagonal) && cxf < ci_x)
3295 /* Leave the pairs with i > j.
3296 * x is the major index, so skip half of it.
3301 set_icell_bb(iGrid, ci, shx, shy, shz, nbl->work.get());
3303 icell_set_x(cell0_i + ci,
3312 for (int cx = cxf; cx <= cxl; cx++)
3314 const real cx_real = cx;
3316 if (jGridDims.lowerCorner[XX] + cx_real * jGridDims.cellSize[XX] > bx1)
3318 d2zx += gmx::square(jGridDims.lowerCorner[XX]
3319 + cx_real * jGridDims.cellSize[XX] - bx1);
3321 else if (jGridDims.lowerCorner[XX] + (cx_real + 1) * jGridDims.cellSize[XX] < bx0)
3323 d2zx += gmx::square(jGridDims.lowerCorner[XX]
3324 + (cx_real + 1) * jGridDims.cellSize[XX] - bx0);
3327 /* When true, leave the pairs with i > j.
3328 * Skip half of y when i and j have the same x.
3330 const bool skipHalfY =
3331 (isIntraGridList && cx == 0
3332 && (!c_pbcShiftBackward || shift == CENTRAL) && cyf < ci_y);
3333 const int cyf_x = skipHalfY ? ci_y : cyf;
3335 for (int cy = cyf_x; cy <= cyl; cy++)
3337 const int columnStart =
3338 jGrid.firstCellInColumn(cx * jGridDims.numCells[YY] + cy);
3339 const int columnEnd =
3340 jGrid.firstCellInColumn(cx * jGridDims.numCells[YY] + cy + 1);
3342 const real cy_real = cy;
3344 if (jGridDims.lowerCorner[YY] + cy_real * jGridDims.cellSize[YY] > by1)
3346 d2zxy += gmx::square(jGridDims.lowerCorner[YY]
3347 + cy_real * jGridDims.cellSize[YY] - by1);
3349 else if (jGridDims.lowerCorner[YY] + (cy_real + 1) * jGridDims.cellSize[YY] < by0)
3351 d2zxy += gmx::square(jGridDims.lowerCorner[YY]
3352 + (cy_real + 1) * jGridDims.cellSize[YY] - by0);
3354 if (columnStart < columnEnd && d2zxy < listRangeBBToJCell2)
3356 /* To improve efficiency in the common case
3357 * of a homogeneous particle distribution,
3358 * we estimate the index of the middle cell
3359 * in range (midCell). We search down and up
3360 * starting from this index.
3362 * Note that the bbcz_j array contains bounds
3363 * for i-clusters, thus for clusters of 4 atoms.
3364 * For the common case where the j-cluster size
3365 * is 8, we could step with a stride of 2,
3366 * but we do not do this because it would
3367 * complicate this code even more.
3371 + static_cast<int>(bz1_frac
3372 * static_cast<real>(columnEnd - columnStart));
3373 if (midCell >= columnEnd)
3375 midCell = columnEnd - 1;
3378 const real d2xy = d2zxy - d2z;
3380 /* Find the lowest cell that can possibly
3382 * Check if we hit the bottom of the grid,
3383 * if the j-cell is below the i-cell and if so,
3384 * if it is within range.
3386 int downTestCell = midCell;
3387 while (downTestCell >= columnStart
3388 && (bbcz_j[downTestCell].upper >= bz0
3389 || d2xy + gmx::square(bbcz_j[downTestCell].upper - bz0) < rlist2))
3393 int firstCell = downTestCell + 1;
3395 /* Find the highest cell that can possibly
3397 * Check if we hit the top of the grid,
3398 * if the j-cell is above the i-cell and if so,
3399 * if it is within range.
3401 int upTestCell = midCell + 1;
3402 while (upTestCell < columnEnd
3403 && (bbcz_j[upTestCell].lower <= bz1
3404 || d2xy + gmx::square(bbcz_j[upTestCell].lower - bz1) < rlist2))
3408 int lastCell = upTestCell - 1;
3410 #define NBNXN_REFCODE 0
3413 /* Simple reference code, for debugging,
3414 * overrides the more complex code above.
3416 firstCell = columnEnd;
3418 for (int k = columnStart; k < columnEnd; k++)
3420 if (d2xy + gmx::square(bbcz_j[k * NNBSBB_D + 1] - bz0) < rlist2
3425 if (d2xy + gmx::square(bbcz_j[k * NNBSBB_D] - bz1) < rlist2
3434 if (isIntraGridList)
3436 /* We want each atom/cell pair only once,
3437 * only use cj >= ci.
3439 if (!c_pbcShiftBackward || shift == CENTRAL)
3441 firstCell = std::max(firstCell, ci);
3445 if (firstCell <= lastCell)
3447 GMX_ASSERT(firstCell >= columnStart && lastCell < columnEnd,
3448 "The range should reside within the current grid "
3451 /* For f buffer flags with simple lists */
3452 const int ncj_old_j = getNumSimpleJClustersInList(*nbl);
3454 makeClusterListWrapper(nbl,
3465 &numDistanceChecks);
3469 setBufferFlags(*nbl, ncj_old_j, gridj_flag_shift, gridj_flag, th);
3472 incrementNumSimpleJClustersInList(nbl, ncj_old_j);
3478 if (!exclusions.empty())
3480 /* Set the exclusions for this ci list */
3481 setExclusionsForIEntry(
3482 gridSet, nbl, excludeSubDiagonal, na_cj_2log, *getOpenIEntry(nbl), exclusions);
3487 make_fep_list(gridSet.atomIndices(),
3501 /* Close this ci list */
3502 closeIEntry(nbl, nsubpair_max, progBal, nsubpair_tot_est, th, nth);
3507 if (bFBufferFlag && getNumSimpleJClustersInList(*nbl) > ncj_old_i)
3509 bitmask_init_bit(&(work->buffer_flags[(iGrid.cellOffset() + ci) >> gridi_flag_shift]), th);
3513 work->ndistc = numDistanceChecks;
3515 checkListSizeConsistency(*nbl, haveFep);
3519 fprintf(debug, "number of distance checks %d\n", numDistanceChecks);
3521 print_nblist_statistics(debug, *nbl, gridSet, rlist);
3525 fprintf(debug, "nbl FEP list pairs: %d\n", nbl_fep->nrj);
3530 static void reduce_buffer_flags(gmx::ArrayRef<PairsearchWork> searchWork,
3532 gmx::ArrayRef<gmx_bitmask_t> dest)
3534 for (int s = 0; s < nsrc; s++)
3536 gmx::ArrayRef<gmx_bitmask_t> flags(searchWork[s].buffer_flags);
3538 for (size_t b = 0; b < dest.size(); b++)
3540 gmx_bitmask_t& flag = dest[b];
3541 bitmask_union(&flag, flags[b]);
3546 static void print_reduction_cost(gmx::ArrayRef<const gmx_bitmask_t> flags, int nout)
3553 gmx_bitmask_t mask_0; // NOLINT(cppcoreguidelines-init-variables)
3554 bitmask_init_bit(&mask_0, 0);
3555 for (const gmx_bitmask_t& flag_mask : flags)
3557 if (bitmask_is_equal(flag_mask, mask_0))
3559 /* Only flag 0 is set, no copy of reduction required */
3563 else if (!bitmask_is_zero(flag_mask))
3566 for (int out = 0; out < nout; out++)
3568 if (bitmask_is_set(flag_mask, out))
3584 const auto numFlags = static_cast<double>(flags.size());
3586 "nbnxn reduction: #flag %zu #list %d elem %4.2f, keep %4.2f copy %4.2f red %4.2f\n",
3595 /* Copies the list entries from src to dest when cjStart <= *cjGlobal < cjEnd.
3596 * *cjGlobal is updated with the cj count in src.
3597 * When setFlags==true, flag bit t is set in flag for all i and j clusters.
3599 template<bool setFlags>
3600 static void copySelectedListRange(const nbnxn_ci_t* gmx_restrict srcCi,
3601 const NbnxnPairlistCpu* gmx_restrict src,
3602 NbnxnPairlistCpu* gmx_restrict dest,
3603 gmx_bitmask_t* flag,
3608 const int ncj = srcCi->cj_ind_end - srcCi->cj_ind_start;
3610 dest->ci.push_back(*srcCi);
3611 dest->ci.back().cj_ind_start = dest->cj.size();
3612 dest->ci.back().cj_ind_end = dest->ci.back().cj_ind_start + ncj;
3616 bitmask_init_bit(&flag[srcCi->ci >> iFlagShift], t);
3619 for (int j = srcCi->cj_ind_start; j < srcCi->cj_ind_end; j++)
3621 dest->cj.push_back(src->cj[j]);
3625 /* NOTE: This is relatively expensive, since this
3626 * operation is done for all elements in the list,
3627 * whereas at list generation this is done only
3628 * once for each flag entry.
3630 bitmask_init_bit(&flag[src->cj[j].cj >> jFlagShift], t);
3635 #if defined(__GNUC__) && !defined(__clang__) && __GNUC__ == 7
3636 /* Avoid gcc 7 avx512 loop vectorization bug (actually only needed with -mavx512f) */
3637 # pragma GCC push_options
3638 # pragma GCC optimize("no-tree-vectorize")
3641 /* Returns the number of cluster pairs that are in use summed over all lists */
3642 static int countClusterpairs(gmx::ArrayRef<const NbnxnPairlistCpu> pairlists)
3644 /* gcc 7 with -mavx512f can miss the contributions of 16 consecutive
3645 * elements to the sum calculated in this loop. Above we have disabled
3646 * loop vectorization to avoid this bug.
3649 for (const auto& pairlist : pairlists)
3651 ncjTotal += pairlist.ncjInUse;
3656 #if defined(__GNUC__) && !defined(__clang__) && __GNUC__ == 7
3657 # pragma GCC pop_options
3660 /* This routine re-balances the pairlists such that all are nearly equally
3661 * sized. Only whole i-entries are moved between lists. These are moved
3662 * between the ends of the lists, such that the buffer reduction cost should
3663 * not change significantly.
3664 * Note that all original reduction flags are currently kept. This can lead
3665 * to reduction of parts of the force buffer that could be avoided. But since
3666 * the original lists are quite balanced, this will only give minor overhead.
3668 static void rebalanceSimpleLists(gmx::ArrayRef<const NbnxnPairlistCpu> srcSet,
3669 gmx::ArrayRef<NbnxnPairlistCpu> destSet,
3670 gmx::ArrayRef<PairsearchWork> searchWork)
3672 const int ncjTotal = countClusterpairs(srcSet);
3673 const int numLists = srcSet.ssize();
3674 const int ncjTarget = (ncjTotal + numLists - 1) / numLists;
3676 #pragma omp parallel num_threads(numLists)
3678 int t = gmx_omp_get_thread_num();
3680 int cjStart = ncjTarget * t;
3681 int cjEnd = ncjTarget * (t + 1);
3683 /* The destination pair-list for task/thread t */
3684 NbnxnPairlistCpu& dest = destSet[t];
3686 clear_pairlist(&dest);
3687 dest.na_cj = srcSet[0].na_cj;
3689 /* Note that the flags in the work struct (still) contain flags
3690 * for all entries that are present in srcSet->nbl[t].
3692 gmx_bitmask_t* flag = &searchWork[t].buffer_flags[0];
3694 int iFlagShift = getBufferFlagShift(dest.na_ci);
3695 int jFlagShift = getBufferFlagShift(dest.na_cj);
3698 for (int s = 0; s < numLists && cjGlobal < cjEnd; s++)
3700 const NbnxnPairlistCpu* src = &srcSet[s];
3702 if (cjGlobal + src->ncjInUse > cjStart)
3704 for (gmx::index i = 0; i < gmx::ssize(src->ci) && cjGlobal < cjEnd; i++)
3706 const nbnxn_ci_t* srcCi = &src->ci[i];
3707 int ncj = srcCi->cj_ind_end - srcCi->cj_ind_start;
3708 if (cjGlobal >= cjStart)
3710 /* If the source list is not our own, we need to set
3711 * extra flags (the template bool parameter).
3715 copySelectedListRange<true>(srcCi, src, &dest, flag, iFlagShift, jFlagShift, t);
3719 copySelectedListRange<false>(
3720 srcCi, src, &dest, flag, iFlagShift, jFlagShift, t);
3728 cjGlobal += src->ncjInUse;
3732 dest.ncjInUse = dest.cj.size();
3736 const int ncjTotalNew = countClusterpairs(destSet);
3737 GMX_RELEASE_ASSERT(ncjTotalNew == ncjTotal,
3738 "The total size of the lists before and after rebalancing should match");
3742 /* Returns if the pairlists are so imbalanced that it is worth rebalancing. */
3743 static bool checkRebalanceSimpleLists(gmx::ArrayRef<const NbnxnPairlistCpu> lists)
3745 int numLists = lists.ssize();
3748 for (int s = 0; s < numLists; s++)
3750 ncjMax = std::max(ncjMax, lists[s].ncjInUse);
3751 ncjTotal += lists[s].ncjInUse;
3755 fprintf(debug, "Pair-list ncjMax %d ncjTotal %d\n", ncjMax, ncjTotal);
3757 /* The rebalancing adds 3% extra time to the search. Heuristically we
3758 * determined that under common conditions the non-bonded kernel balance
3759 * improvement will outweigh this when the imbalance is more than 3%.
3760 * But this will, obviously, depend on search vs kernel time and nstlist.
3762 const real rebalanceTolerance = 1.03;
3764 return real(numLists * ncjMax) > real(ncjTotal) * rebalanceTolerance;
3767 /* Perform a count (linear) sort to sort the smaller lists to the end.
3768 * This avoids load imbalance on the GPU, as large lists will be
3769 * scheduled and executed first and the smaller lists later.
3770 * Load balancing between multi-processors only happens at the end
3771 * and there smaller lists lead to more effective load balancing.
3772 * The sorting is done on the cj4 count, not on the actual pair counts.
3773 * Not only does this make the sort faster, but it also results in
3774 * better load balancing than using a list sorted on exact load.
3775 * This function swaps the pointer in the pair list to avoid a copy operation.
3777 static void sort_sci(NbnxnPairlistGpu* nbl)
3779 if (nbl->cj4.size() <= nbl->sci.size())
3781 /* nsci = 0 or all sci have size 1, sorting won't change the order */
3785 NbnxnPairlistGpuWork& work = *nbl->work;
3787 /* We will distinguish differences up to double the average */
3788 const int m = static_cast<int>((2 * ssize(nbl->cj4)) / ssize(nbl->sci));
3790 /* Resize work.sci_sort so we can sort into it */
3791 work.sci_sort.resize(nbl->sci.size());
3793 std::vector<int>& sort = work.sortBuffer;
3794 /* Set up m + 1 entries in sort, initialized at 0 */
3796 sort.resize(m + 1, 0);
3797 /* Count the entries of each size */
3798 for (const nbnxn_sci_t& sci : nbl->sci)
3800 int i = std::min(m, sci.numJClusterGroups());
3803 /* Calculate the offset for each count */
3806 for (gmx::index i = m - 1; i >= 0; i--)
3809 sort[i] = sort[i + 1] + s0;
3813 /* Sort entries directly into place */
3814 gmx::ArrayRef<nbnxn_sci_t> sci_sort = work.sci_sort;
3815 for (const nbnxn_sci_t& sci : nbl->sci)
3817 int i = std::min(m, sci.numJClusterGroups());
3818 sci_sort[sort[i]++] = sci;
3821 /* Swap the sci pointers so we use the new, sorted list */
3822 std::swap(nbl->sci, work.sci_sort);
3825 /* Returns the i-zone range for pairlist construction for the give locality */
3826 static Range<int> getIZoneRange(const Nbnxm::GridSet::DomainSetup& domainSetup,
3827 const InteractionLocality locality)
3829 if (domainSetup.doTestParticleInsertion)
3831 /* With TPI we do grid 1, the inserted molecule, versus grid 0, the rest */
3834 else if (locality == InteractionLocality::Local)
3836 /* Local: only zone (grid) 0 vs 0 */
3841 /* Non-local: we need all i-zones */
3842 return { 0, int(domainSetup.zones->iZones.size()) };
3846 /* Returns the j-zone range for pairlist construction for the give locality and i-zone */
3847 static Range<int> getJZoneRange(const gmx_domdec_zones_t* ddZones,
3848 const InteractionLocality locality,
3851 if (locality == InteractionLocality::Local)
3853 /* Local: zone 0 vs 0 or with TPI 1 vs 0 */
3856 else if (iZone == 0)
3858 /* Non-local: we need to avoid the local (zone 0 vs 0) interactions */
3859 return { 1, *ddZones->iZones[iZone].jZoneRange.end() };
3863 /* Non-local with non-local i-zone: use all j-zones */
3864 return ddZones->iZones[iZone].jZoneRange;
3868 //! Prepares CPU lists produced by the search for dynamic pruning
3869 static void prepareListsForDynamicPruning(gmx::ArrayRef<NbnxnPairlistCpu> lists);
3871 void PairlistSet::constructPairlists(const Nbnxm::GridSet& gridSet,
3872 gmx::ArrayRef<PairsearchWork> searchWork,
3873 nbnxn_atomdata_t* nbat,
3874 const ListOfLists<int>& exclusions,
3875 const int minimumIlistCountForGpuBalancing,
3877 SearchCycleCounting* searchCycleCounting)
3879 const real rlist = params_.rlistOuter;
3881 const int numLists = (isCpuType_ ? cpuLists_.size() : gpuLists_.size());
3885 fprintf(debug, "ns making %d nblists\n", numLists);
3888 nbat->bUseBufferFlags = (nbat->out.size() > 1);
3889 /* We should re-init the flags before making the first list */
3890 if (nbat->bUseBufferFlags && locality_ == InteractionLocality::Local)
3892 resizeAndZeroBufferFlags(&nbat->buffer_flags, nbat->numAtoms());
3895 int nsubpair_target = 0;
3896 float nsubpair_tot_est = 0.0F;
3897 if (!isCpuType_ && minimumIlistCountForGpuBalancing > 0)
3899 get_nsubpair_target(
3900 gridSet, locality_, rlist, minimumIlistCountForGpuBalancing, &nsubpair_target, &nsubpair_tot_est);
3903 /* Clear all pair-lists */
3904 for (int th = 0; th < numLists; th++)
3908 clear_pairlist(&cpuLists_[th]);
3912 clear_pairlist(&gpuLists_[th]);
3915 if (params_.haveFep)
3917 clear_pairlist_fep(fepLists_[th].get());
3921 const gmx_domdec_zones_t* ddZones = gridSet.domainSetup().zones;
3922 GMX_ASSERT(locality_ == InteractionLocality::Local || ddZones != nullptr,
3923 "Nonlocal interaction locality with null ddZones.");
3925 const auto iZoneRange = getIZoneRange(gridSet.domainSetup(), locality_);
3927 for (const int iZone : iZoneRange)
3929 const Grid& iGrid = gridSet.grids()[iZone];
3931 const auto jZoneRange = getJZoneRange(ddZones, locality_, iZone);
3933 for (int jZone : jZoneRange)
3935 const Grid& jGrid = gridSet.grids()[jZone];
3939 fprintf(debug, "ns search grid %d vs %d\n", iZone, jZone);
3942 searchCycleCounting->start(enbsCCsearch);
3944 const int ci_block =
3945 get_ci_block_size(iGrid, gridSet.domainSetup().haveMultipleDomains, numLists);
3947 /* With GPU: generate progressively smaller lists for
3948 * load balancing for local only or non-local with 2 zones.
3950 const bool progBal = (locality_ == InteractionLocality::Local || ddZones->n <= 2);
3952 #pragma omp parallel for num_threads(numLists) schedule(static)
3953 for (int th = 0; th < numLists; th++)
3957 /* Re-init the thread-local work flag data before making
3958 * the first list (not an elegant conditional).
3960 if (nbat->bUseBufferFlags && (iZone == 0 && jZone == 0))
3962 resizeAndZeroBufferFlags(&searchWork[th].buffer_flags, nbat->numAtoms());
3965 if (combineLists_ && th > 0)
3967 GMX_ASSERT(!isCpuType_, "Can only combine GPU lists");
3969 clear_pairlist(&gpuLists_[th]);
3972 PairsearchWork& work = searchWork[th];
3974 work.cycleCounter.start();
3976 t_nblist* fepListPtr = (fepLists_.empty() ? nullptr : fepLists_[th].get());
3978 /* Divide the i cells equally over the pairlists */
3981 nbnxn_make_pairlist_part(gridSet,
3988 params_.pairlistType,
3990 nbat->bUseBufferFlags,
4001 nbnxn_make_pairlist_part(gridSet,
4008 params_.pairlistType,
4010 nbat->bUseBufferFlags,
4020 work.cycleCounter.stop();
4022 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
4024 searchCycleCounting->stop(enbsCCsearch);
4029 for (int th = 0; th < numLists; th++)
4031 inc_nrnb(nrnb, eNR_NBNXN_DIST2, searchWork[th].ndistc);
4035 const NbnxnPairlistCpu& nbl = cpuLists_[th];
4036 np_tot += nbl.cj.size();
4037 np_noq += nbl.work->ncj_noq;
4038 np_hlj += nbl.work->ncj_hlj;
4042 const NbnxnPairlistGpu& nbl = gpuLists_[th];
4043 /* This count ignores potential subsequent pair pruning */
4044 np_tot += nbl.nci_tot;
4047 const int nap = isCpuType_ ? cpuLists_[0].na_ci * cpuLists_[0].na_cj
4048 : gmx::square(gpuLists_[0].na_ci);
4050 natpair_ljq_ = (np_tot - np_noq) * nap - np_hlj * nap / 2;
4051 natpair_lj_ = np_noq * nap;
4052 natpair_q_ = np_hlj * nap / 2;
4054 if (combineLists_ && numLists > 1)
4056 GMX_ASSERT(!isCpuType_, "Can only combine GPU lists");
4058 searchCycleCounting->start(enbsCCcombine);
4060 combine_nblists(gmx::constArrayRefFromArray(&gpuLists_[1], numLists - 1), &gpuLists_[0]);
4062 searchCycleCounting->stop(enbsCCcombine);
4069 if (numLists > 1 && checkRebalanceSimpleLists(cpuLists_))
4071 rebalanceSimpleLists(cpuLists_, cpuListsWork_, searchWork);
4073 /* Swap the sets of pair lists */
4074 cpuLists_.swap(cpuListsWork_);
4079 /* Sort the entries on size, large ones first */
4080 if (combineLists_ || gpuLists_.size() == 1)
4082 sort_sci(&gpuLists_[0]);
4086 #pragma omp parallel for num_threads(numLists) schedule(static)
4087 for (int th = 0; th < numLists; th++)
4091 sort_sci(&gpuLists_[th]);
4093 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
4098 if (nbat->bUseBufferFlags)
4100 reduce_buffer_flags(searchWork, numLists, nbat->buffer_flags);
4103 if (gridSet.haveFep())
4105 /* Balance the free-energy lists over all the threads */
4106 balance_fep_lists(fepLists_, searchWork);
4111 /* This is a fresh list, so not pruned, stored using ci.
4112 * ciOuter is invalid at this point.
4114 GMX_ASSERT(cpuLists_[0].ciOuter.empty(), "ciOuter is invalid so it should be empty");
4117 /* If we have more than one list, they either got rebalancing (CPU)
4118 * or combined (GPU), so we should dump the final result to debug.
4122 if (isCpuType_ && cpuLists_.size() > 1)
4124 for (auto& cpuList : cpuLists_)
4126 print_nblist_statistics(debug, cpuList, gridSet, rlist);
4129 else if (!isCpuType_ && gpuLists_.size() > 1)
4131 print_nblist_statistics(debug, gpuLists_[0], gridSet, rlist);
4141 for (auto& cpuList : cpuLists_)
4143 print_nblist_ci_cj(debug, cpuList);
4148 print_nblist_sci_cj(debug, gpuLists_[0]);
4152 if (nbat->bUseBufferFlags)
4154 print_reduction_cost(nbat->buffer_flags, numLists);
4158 if (params_.useDynamicPruning && isCpuType_)
4160 prepareListsForDynamicPruning(cpuLists_);
4164 void PairlistSets::construct(const InteractionLocality iLocality,
4165 PairSearch* pairSearch,
4166 nbnxn_atomdata_t* nbat,
4167 const ListOfLists<int>& exclusions,
4171 const auto& gridSet = pairSearch->gridSet();
4172 const auto* ddZones = gridSet.domainSetup().zones;
4174 /* The Nbnxm code can also work with more exclusions than those in i-zones only
4175 * when using DD, but the equality check can catch more issues.
4178 exclusions.empty() || (!ddZones && exclusions.ssize() == gridSet.numRealAtomsTotal())
4179 || (ddZones && exclusions.ssize() == ddZones->cg_range[ddZones->iZones.size()]),
4180 "exclusions should either be empty or the number of lists should match the number of "
4183 pairlistSet(iLocality).constructPairlists(gridSet,
4187 minimumIlistCountForGpuBalancing_,
4189 &pairSearch->cycleCounting_);
4191 if (iLocality == InteractionLocality::Local)
4193 outerListCreationStep_ = step;
4197 GMX_RELEASE_ASSERT(outerListCreationStep_ == step,
4198 "Outer list should be created at the same step as the inner list");
4201 /* Special performance logging stuff (env.var. GMX_NBNXN_CYCLE) */
4202 if (iLocality == InteractionLocality::Local)
4204 pairSearch->cycleCounting_.searchCount_++;
4206 if (pairSearch->cycleCounting_.recordCycles_
4207 && (!pairSearch->gridSet().domainSetup().haveMultipleDomains || iLocality == InteractionLocality::NonLocal)
4208 && pairSearch->cycleCounting_.searchCount_ % 100 == 0)
4210 pairSearch->cycleCounting_.printCycles(stderr, pairSearch->work());
4214 void nonbonded_verlet_t::constructPairlist(const InteractionLocality iLocality,
4215 const ListOfLists<int>& exclusions,
4219 pairlistSets_->construct(iLocality, pairSearch_.get(), nbat.get(), exclusions, step, nrnb);
4223 /* Launch the transfer of the pairlist to the GPU.
4225 * NOTE: The launch overhead is currently not timed separately
4227 Nbnxm::gpu_init_pairlist(gpu_nbv, pairlistSets().pairlistSet(iLocality).gpuList(), iLocality);
4231 static void prepareListsForDynamicPruning(gmx::ArrayRef<NbnxnPairlistCpu> lists)
4233 /* TODO: Restructure the lists so we have actual outer and inner
4234 * list objects so we can set a single pointer instead of
4235 * swapping several pointers.
4238 for (auto& list : lists)
4240 /* The search produced a list in ci/cj.
4241 * Swap the list pointers so we get the outer list is ciOuter,cjOuter
4242 * and we can prune that to get an inner list in ci/cj.
4244 GMX_RELEASE_ASSERT(list.ciOuter.empty() && list.cjOuter.empty(),
4245 "The outer lists should be empty before preparation");
4247 std::swap(list.ci, list.ciOuter);
4248 std::swap(list.cj, list.cjOuter);