2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012-2018, The GROMACS development team.
5 * Copyright (c) 2019, 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.
48 #include "thread_mpi/atomic.h"
50 #include "gromacs/math/functions.h"
51 #include "gromacs/math/utilities.h"
52 #include "gromacs/math/vec.h"
53 #include "gromacs/mdlib/gmx_omp_nthreads.h"
54 #include "gromacs/mdtypes/forcerec.h" // only for GET_CGINFO_*
55 #include "gromacs/mdtypes/mdatom.h"
56 #include "gromacs/nbnxm/nbnxm.h"
57 #include "gromacs/pbcutil/ishift.h"
58 #include "gromacs/simd/simd.h"
59 #include "gromacs/utility/exceptions.h"
60 #include "gromacs/utility/fatalerror.h"
61 #include "gromacs/utility/gmxomp.h"
62 #include "gromacs/utility/logger.h"
63 #include "gromacs/utility/smalloc.h"
64 #include "gromacs/utility/strconvert.h"
65 #include "gromacs/utility/stringutil.h"
69 #include "nbnxm_geometry.h"
72 using namespace gmx; // TODO: Remove when this file is moved into gmx namespace
74 void nbnxn_atomdata_t::resizeCoordinateBuffer(int numAtoms)
78 x_.resize(numAtoms * xstride);
81 void nbnxn_atomdata_t::resizeForceBuffers()
83 /* Force buffers need padding up to a multiple of the buffer flag size */
84 const int paddedSize =
85 (numAtoms() + NBNXN_BUFFERFLAG_SIZE - 1) / NBNXN_BUFFERFLAG_SIZE * NBNXN_BUFFERFLAG_SIZE;
87 /* Should we let each thread allocate it's own data instead? */
88 for (nbnxn_atomdata_output_t& outBuffer : out)
90 outBuffer.f.resize(paddedSize * fstride);
94 /* Initializes an nbnxn_atomdata_output_t data structure */
95 nbnxn_atomdata_output_t::nbnxn_atomdata_output_t(Nbnxm::KernelType kernelType,
97 int simdEnergyBufferStride,
98 gmx::PinningPolicy pinningPolicy) :
99 f({}, { pinningPolicy }),
100 fshift({}, { pinningPolicy }),
101 Vvdw({}, { pinningPolicy }),
102 Vc({}, { pinningPolicy })
104 fshift.resize(SHIFTS * DIM);
105 Vvdw.resize(numEnergyGroups * numEnergyGroups);
106 Vc.resize(numEnergyGroups * numEnergyGroups);
108 if (Nbnxm::kernelTypeIsSimd(kernelType))
110 int cj_size = Nbnxm::JClusterSizePerKernelType[kernelType];
112 numEnergyGroups * numEnergyGroups * simdEnergyBufferStride * (cj_size / 2) * cj_size;
113 VSvdw.resize(numElements);
114 VSc.resize(numElements);
118 static void copy_int_to_nbat_int(const int* a, int na, int na_round, const int* in, int fill, int* innb)
123 for (i = 0; i < na; i++)
125 innb[j++] = in[a[i]];
127 /* Complete the partially filled last cell with fill */
128 for (; i < na_round; i++)
134 void copy_rvec_to_nbat_real(const int* a, int na, int na_round, const rvec* x, int nbatFormat, real* xnb, int a0)
136 /* We complete partially filled cells, can only be the last one in each
137 * column, with coordinates farAway. The actual coordinate value does
138 * not influence the results, since these filler particles do not interact.
139 * Clusters with normal atoms + fillers have a bounding box based only
140 * on the coordinates of the atoms. Clusters with only fillers have as
141 * the bounding box the coordinates of the first filler. Such clusters
142 * are not considered as i-entries, but they are considered as j-entries.
143 * So for performance it is better to have their bounding boxes far away,
144 * such that filler only clusters don't end up in the pair list.
146 const real farAway = -1000000;
154 for (i = 0; i < na; i++)
156 xnb[j++] = x[a[i]][XX];
157 xnb[j++] = x[a[i]][YY];
158 xnb[j++] = x[a[i]][ZZ];
160 /* Complete the partially filled last cell with farAway elements */
161 for (; i < na_round; i++)
169 j = a0 * STRIDE_XYZQ;
170 for (i = 0; i < na; i++)
172 xnb[j++] = x[a[i]][XX];
173 xnb[j++] = x[a[i]][YY];
174 xnb[j++] = x[a[i]][ZZ];
177 /* Complete the partially filled last cell with zeros */
178 for (; i < na_round; i++)
187 j = atom_to_x_index<c_packX4>(a0);
188 c = a0 & (c_packX4 - 1);
189 for (i = 0; i < na; i++)
191 xnb[j + XX * c_packX4] = x[a[i]][XX];
192 xnb[j + YY * c_packX4] = x[a[i]][YY];
193 xnb[j + ZZ * c_packX4] = x[a[i]][ZZ];
198 j += (DIM - 1) * c_packX4;
202 /* Complete the partially filled last cell with zeros */
203 for (; i < na_round; i++)
205 xnb[j + XX * c_packX4] = farAway;
206 xnb[j + YY * c_packX4] = farAway;
207 xnb[j + ZZ * c_packX4] = farAway;
212 j += (DIM - 1) * c_packX4;
218 j = atom_to_x_index<c_packX8>(a0);
219 c = a0 & (c_packX8 - 1);
220 for (i = 0; i < na; i++)
222 xnb[j + XX * c_packX8] = x[a[i]][XX];
223 xnb[j + YY * c_packX8] = x[a[i]][YY];
224 xnb[j + ZZ * c_packX8] = x[a[i]][ZZ];
229 j += (DIM - 1) * c_packX8;
233 /* Complete the partially filled last cell with zeros */
234 for (; i < na_round; i++)
236 xnb[j + XX * c_packX8] = farAway;
237 xnb[j + YY * c_packX8] = farAway;
238 xnb[j + ZZ * c_packX8] = farAway;
243 j += (DIM - 1) * c_packX8;
248 default: gmx_incons("Unsupported nbnxn_atomdata_t format");
252 /* Stores the LJ parameter data in a format convenient for different kernels */
253 static void set_lj_parameter_data(nbnxn_atomdata_t::Params* params, gmx_bool bSIMD)
255 int nt = params->numTypes;
260 /* nbfp_aligned stores two parameters using the stride most suitable
261 * for the present SIMD architecture, as specified by the constant
262 * c_simdBestPairAlignment from the SIMD header.
263 * There's a slight inefficiency in allocating and initializing nbfp_aligned
264 * when it might not be used, but introducing the conditional code is not
267 params->nbfp_aligned.resize(nt * nt * c_simdBestPairAlignment);
269 for (int i = 0; i < nt; i++)
271 for (int j = 0; j < nt; j++)
273 params->nbfp_aligned[(i * nt + j) * c_simdBestPairAlignment + 0] =
274 params->nbfp[(i * nt + j) * 2 + 0];
275 params->nbfp_aligned[(i * nt + j) * c_simdBestPairAlignment + 1] =
276 params->nbfp[(i * nt + j) * 2 + 1];
277 if (c_simdBestPairAlignment > 2)
279 params->nbfp_aligned[(i * nt + j) * c_simdBestPairAlignment + 2] = 0;
280 params->nbfp_aligned[(i * nt + j) * c_simdBestPairAlignment + 3] = 0;
287 /* We use combination rule data for SIMD combination rule kernels
288 * and with LJ-PME kernels. We then only need parameters per atom type,
289 * not per pair of atom types.
291 params->nbfp_comb.resize(nt * 2);
292 switch (params->comb_rule)
295 params->comb_rule = ljcrGEOM;
297 for (int i = 0; i < nt; i++)
299 /* Store the sqrt of the diagonal from the nbfp matrix */
300 params->nbfp_comb[i * 2] = std::sqrt(params->nbfp[(i * nt + i) * 2]);
301 params->nbfp_comb[i * 2 + 1] = std::sqrt(params->nbfp[(i * nt + i) * 2 + 1]);
305 for (int i = 0; i < nt; i++)
307 /* Get 6*C6 and 12*C12 from the diagonal of the nbfp matrix */
308 const real c6 = params->nbfp[(i * nt + i) * 2];
309 const real c12 = params->nbfp[(i * nt + i) * 2 + 1];
310 if (c6 > 0 && c12 > 0)
312 /* We store 0.5*2^1/6*sigma and sqrt(4*3*eps),
313 * so we get 6*C6 and 12*C12 after combining.
315 params->nbfp_comb[i * 2] = 0.5 * gmx::sixthroot(c12 / c6);
316 params->nbfp_comb[i * 2 + 1] = std::sqrt(c6 * c6 / c12);
320 params->nbfp_comb[i * 2] = 0;
321 params->nbfp_comb[i * 2 + 1] = 0;
326 /* We always store the full matrix (see code above) */
328 default: gmx_incons("Unknown combination rule");
332 nbnxn_atomdata_t::SimdMasks::SimdMasks()
335 constexpr int simd_width = GMX_SIMD_REAL_WIDTH;
336 /* Set the diagonal cluster pair exclusion mask setup data.
337 * In the kernel we check 0 < j - i to generate the masks.
338 * Here we store j - i for generating the mask for the first i,
339 * we substract 0.5 to avoid rounding issues.
340 * In the kernel we can subtract 1 to generate the subsequent mask.
342 const int simd_4xn_diag_size = std::max(c_nbnxnCpuIClusterSize, simd_width);
343 diagonal_4xn_j_minus_i.resize(simd_4xn_diag_size);
344 for (int j = 0; j < simd_4xn_diag_size; j++)
346 diagonal_4xn_j_minus_i[j] = j - 0.5;
349 diagonal_2xnn_j_minus_i.resize(simd_width);
350 for (int j = 0; j < simd_width / 2; j++)
352 /* The j-cluster size is half the SIMD width */
353 diagonal_2xnn_j_minus_i[j] = j - 0.5;
354 /* The next half of the SIMD width is for i + 1 */
355 diagonal_2xnn_j_minus_i[simd_width / 2 + j] = j - 1 - 0.5;
358 /* We use up to 32 bits for exclusion masking.
359 * The same masks are used for the 4xN and 2x(N+N) kernels.
360 * The masks are read either into integer SIMD registers or into
361 * real SIMD registers (together with a cast).
362 * In single precision this means the real and integer SIMD registers
365 const int simd_excl_size = c_nbnxnCpuIClusterSize * simd_width;
366 # if GMX_DOUBLE && !GMX_SIMD_HAVE_INT32_LOGICAL
367 exclusion_filter64.resize(simd_excl_size);
369 exclusion_filter.resize(simd_excl_size);
372 for (int j = 0; j < simd_excl_size; j++)
374 /* Set the consecutive bits for masking pair exclusions */
375 # if GMX_DOUBLE && !GMX_SIMD_HAVE_INT32_LOGICAL
376 exclusion_filter64[j] = (1U << j);
378 exclusion_filter[j] = (1U << j);
382 if (!GMX_SIMD_HAVE_LOGICAL && !GMX_SIMD_HAVE_INT32_LOGICAL) // NOLINT(misc-redundant-expression)
384 // If the SIMD implementation has no bitwise logical operation support
385 // whatsoever we cannot use the normal masking. Instead,
386 // we generate a vector of all 2^4 possible ways an i atom
387 // interacts with its 4 j atoms. Each array entry contains
388 // GMX_SIMD_REAL_WIDTH values that are read with a single aligned SIMD load.
389 // Since there is no logical value representation in this case, we use
390 // any nonzero value to indicate 'true', while zero mean 'false'.
391 // This can then be converted to a SIMD boolean internally in the SIMD
392 // module by comparing to zero.
393 // Each array entry encodes how this i atom will interact with the 4 j atoms.
394 // Matching code exists in set_ci_top_excls() to generate indices into this array.
395 // Those indices are used in the kernels.
397 const int simd_excl_size = c_nbnxnCpuIClusterSize * c_nbnxnCpuIClusterSize;
398 const real simdFalse = 0.0;
399 const real simdTrue = 1.0;
401 interaction_array.resize(simd_excl_size * GMX_SIMD_REAL_WIDTH);
402 for (int j = 0; j < simd_excl_size; j++)
404 const int index = j * GMX_SIMD_REAL_WIDTH;
405 for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
407 interaction_array[index + i] = (j & (1 << i)) ? simdTrue : simdFalse;
414 nbnxn_atomdata_t::Params::Params(gmx::PinningPolicy pinningPolicy) :
416 nbfp({}, { pinningPolicy }),
417 nbfp_comb({}, { pinningPolicy }),
418 type({}, { pinningPolicy }),
419 lj_comb({}, { pinningPolicy }),
420 q({}, { pinningPolicy }),
423 energrp({}, { pinningPolicy })
427 nbnxn_atomdata_t::nbnxn_atomdata_t(gmx::PinningPolicy pinningPolicy) :
428 params_(pinningPolicy),
431 shift_vec({}, { pinningPolicy }),
432 x_({}, { pinningPolicy }),
434 bUseBufferFlags(FALSE),
435 bUseTreeReduce(FALSE)
439 /* Initializes an nbnxn_atomdata_t::Params data structure */
440 static void nbnxn_atomdata_params_init(const gmx::MDLogger& mdlog,
441 nbnxn_atomdata_t::Params* params,
442 const Nbnxm::KernelType kernelType,
443 int enbnxninitcombrule,
450 gmx_bool simple, bCombGeom, bCombLB, bSIMD;
454 fprintf(debug, "There are %d atom types in the system, adding one for nbnxn_atomdata_t\n", ntype);
456 params->numTypes = ntype + 1;
457 params->nbfp.resize(params->numTypes * params->numTypes * 2);
458 params->nbfp_comb.resize(params->numTypes * 2);
460 /* A tolerance of 1e-5 seems reasonable for (possibly hand-typed)
461 * force-field floating point parameters.
464 ptr = getenv("GMX_LJCOMB_TOL");
469 sscanf(ptr, "%lf", &dbl);
475 /* Temporarily fill params->nbfp_comb with sigma and epsilon
476 * to check for the LB rule.
478 for (int i = 0; i < ntype; i++)
480 c6 = nbfp[(i * ntype + i) * 2] / 6.0;
481 c12 = nbfp[(i * ntype + i) * 2 + 1] / 12.0;
482 if (c6 > 0 && c12 > 0)
484 params->nbfp_comb[i * 2] = gmx::sixthroot(c12 / c6);
485 params->nbfp_comb[i * 2 + 1] = 0.25 * c6 * c6 / c12;
487 else if (c6 == 0 && c12 == 0)
489 params->nbfp_comb[i * 2] = 0;
490 params->nbfp_comb[i * 2 + 1] = 0;
494 /* Can not use LB rule with only dispersion or repulsion */
499 for (int i = 0; i < params->numTypes; i++)
501 for (int j = 0; j < params->numTypes; j++)
503 if (i < ntype && j < ntype)
505 /* fr->nbfp has been updated, so that array too now stores c6/c12 including
506 * the 6.0/12.0 prefactors to save 2 flops in the most common case (force-only).
508 c6 = nbfp[(i * ntype + j) * 2];
509 c12 = nbfp[(i * ntype + j) * 2 + 1];
510 params->nbfp[(i * params->numTypes + j) * 2] = c6;
511 params->nbfp[(i * params->numTypes + j) * 2 + 1] = c12;
513 /* Compare 6*C6 and 12*C12 for geometric cobination rule */
516 && gmx_within_tol(c6 * c6,
517 nbfp[(i * ntype + i) * 2] * nbfp[(j * ntype + j) * 2], tol)
518 && gmx_within_tol(c12 * c12,
519 nbfp[(i * ntype + i) * 2 + 1] * nbfp[(j * ntype + j) * 2 + 1],
522 /* Compare C6 and C12 for Lorentz-Berthelot combination rule */
527 && ((c6 == 0 && c12 == 0
528 && (params->nbfp_comb[i * 2 + 1] == 0 || params->nbfp_comb[j * 2 + 1] == 0))
529 || (c6 > 0 && c12 > 0
531 gmx::sixthroot(c12 / c6),
532 0.5 * (params->nbfp_comb[i * 2] + params->nbfp_comb[j * 2]), tol)
533 && gmx_within_tol(0.25 * c6 * c6 / c12,
534 std::sqrt(params->nbfp_comb[i * 2 + 1]
535 * params->nbfp_comb[j * 2 + 1]),
540 /* Add zero parameters for the additional dummy atom type */
541 params->nbfp[(i * params->numTypes + j) * 2] = 0;
542 params->nbfp[(i * params->numTypes + j) * 2 + 1] = 0;
548 fprintf(debug, "Combination rules: geometric %s Lorentz-Berthelot %s\n",
549 gmx::boolToString(bCombGeom), gmx::boolToString(bCombLB));
552 simple = Nbnxm::kernelTypeUsesSimplePairlist(kernelType);
554 switch (enbnxninitcombrule)
556 case enbnxninitcombruleDETECT:
557 /* We prefer the geometic combination rule,
558 * as that gives a slightly faster kernel than the LB rule.
562 params->comb_rule = ljcrGEOM;
566 params->comb_rule = ljcrLB;
570 params->comb_rule = ljcrNONE;
572 params->nbfp_comb.clear();
577 if (params->comb_rule == ljcrNONE)
579 mesg = "Using full Lennard-Jones parameter combination matrix";
583 mesg = gmx::formatString(
584 "Using %s Lennard-Jones combination rule",
585 params->comb_rule == ljcrGEOM ? "geometric" : "Lorentz-Berthelot");
587 GMX_LOG(mdlog.info).asParagraph().appendText(mesg);
590 case enbnxninitcombruleGEOM: params->comb_rule = ljcrGEOM; break;
591 case enbnxninitcombruleLB: params->comb_rule = ljcrLB; break;
592 case enbnxninitcombruleNONE:
593 params->comb_rule = ljcrNONE;
595 params->nbfp_comb.clear();
597 default: gmx_incons("Unknown enbnxninitcombrule");
600 bSIMD = Nbnxm::kernelTypeIsSimd(kernelType);
602 set_lj_parameter_data(params, bSIMD);
604 params->nenergrp = n_energygroups;
607 // We now check for energy groups already when starting mdrun
608 GMX_RELEASE_ASSERT(n_energygroups == 1, "GPU kernels do not support energy groups");
610 /* Temporary storage goes as #grp^3*simd_width^2/2, so limit to 64 */
611 if (params->nenergrp > 64)
613 gmx_fatal(FARGS, "With NxN kernels not more than 64 energy groups are supported\n");
615 params->neg_2log = 1;
616 while (params->nenergrp > (1 << params->neg_2log))
622 /* Initializes an nbnxn_atomdata_t data structure */
623 void nbnxn_atomdata_init(const gmx::MDLogger& mdlog,
624 nbnxn_atomdata_t* nbat,
625 const Nbnxm::KernelType kernelType,
626 int enbnxninitcombrule,
632 nbnxn_atomdata_params_init(mdlog, &nbat->paramsDeprecated(), kernelType, enbnxninitcombrule,
633 ntype, nbfp, n_energygroups);
635 const bool simple = Nbnxm::kernelTypeUsesSimplePairlist(kernelType);
636 const bool bSIMD = Nbnxm::kernelTypeIsSimd(kernelType);
644 pack_x = std::max(c_nbnxnCpuIClusterSize, Nbnxm::JClusterSizePerKernelType[kernelType]);
647 case 4: nbat->XFormat = nbatX4; break;
648 case 8: nbat->XFormat = nbatX8; break;
649 default: gmx_incons("Unsupported packing width");
654 nbat->XFormat = nbatXYZ;
657 nbat->FFormat = nbat->XFormat;
661 nbat->XFormat = nbatXYZQ;
662 nbat->FFormat = nbatXYZ;
665 nbat->shift_vec.resize(SHIFTS);
667 nbat->xstride = (nbat->XFormat == nbatXYZQ ? STRIDE_XYZQ : DIM);
668 nbat->fstride = (nbat->FFormat == nbatXYZQ ? STRIDE_XYZQ : DIM);
670 /* Initialize the output data structures */
671 for (int i = 0; i < nout; i++)
673 const auto& pinningPolicy = nbat->params().type.get_allocator().pinningPolicy();
674 nbat->out.emplace_back(kernelType, nbat->params().nenergrp, 1 << nbat->params().neg_2log,
678 nbat->buffer_flags.flag = nullptr;
679 nbat->buffer_flags.flag_nalloc = 0;
681 const int nth = gmx_omp_nthreads_get(emntNonbonded);
683 const char* ptr = getenv("GMX_USE_TREEREDUCE");
686 nbat->bUseTreeReduce = (strtol(ptr, nullptr, 10) != 0);
689 else if (nth > 8) /*on the CPU we currently don't benefit even at 32*/
691 nbat->bUseTreeReduce = 1;
696 nbat->bUseTreeReduce = false;
698 if (nbat->bUseTreeReduce)
700 GMX_LOG(mdlog.info).asParagraph().appendText("Using tree force reduction");
702 nbat->syncStep = new tMPI_Atomic[nth];
706 template<int packSize>
707 static void copy_lj_to_nbat_lj_comb(gmx::ArrayRef<const real> ljparam_type, const int* type, int na, real* ljparam_at)
709 /* The LJ params follow the combination rule:
710 * copy the params for the type array to the atom array.
712 for (int is = 0; is < na; is += packSize)
714 for (int k = 0; k < packSize; k++)
717 ljparam_at[is * 2 + k] = ljparam_type[type[i] * 2];
718 ljparam_at[is * 2 + packSize + k] = ljparam_type[type[i] * 2 + 1];
723 /* Sets the atom type in nbnxn_atomdata_t */
724 static void nbnxn_atomdata_set_atomtypes(nbnxn_atomdata_t::Params* params,
725 const Nbnxm::GridSet& gridSet,
728 params->type.resize(gridSet.numGridAtomsTotal());
730 for (const Nbnxm::Grid& grid : gridSet.grids())
732 /* Loop over all columns and copy and fill */
733 for (int i = 0; i < grid.numColumns(); i++)
735 const int numAtoms = grid.paddedNumAtomsInColumn(i);
736 const int atomOffset = grid.firstAtomInColumn(i);
738 copy_int_to_nbat_int(gridSet.atomIndices().data() + atomOffset, grid.numAtomsInColumn(i),
739 numAtoms, type, params->numTypes - 1, params->type.data() + atomOffset);
744 /* Sets the LJ combination rule parameters in nbnxn_atomdata_t */
745 static void nbnxn_atomdata_set_ljcombparams(nbnxn_atomdata_t::Params* params,
747 const Nbnxm::GridSet& gridSet)
749 params->lj_comb.resize(gridSet.numGridAtomsTotal() * 2);
751 if (params->comb_rule != ljcrNONE)
753 for (const Nbnxm::Grid& grid : gridSet.grids())
755 /* Loop over all columns and copy and fill */
756 for (int i = 0; i < grid.numColumns(); i++)
758 const int numAtoms = grid.paddedNumAtomsInColumn(i);
759 const int atomOffset = grid.firstAtomInColumn(i);
761 if (XFormat == nbatX4)
763 copy_lj_to_nbat_lj_comb<c_packX4>(params->nbfp_comb,
764 params->type.data() + atomOffset, numAtoms,
765 params->lj_comb.data() + atomOffset * 2);
767 else if (XFormat == nbatX8)
769 copy_lj_to_nbat_lj_comb<c_packX8>(params->nbfp_comb,
770 params->type.data() + atomOffset, numAtoms,
771 params->lj_comb.data() + atomOffset * 2);
773 else if (XFormat == nbatXYZQ)
775 copy_lj_to_nbat_lj_comb<1>(params->nbfp_comb, params->type.data() + atomOffset,
776 numAtoms, params->lj_comb.data() + atomOffset * 2);
783 /* Sets the charges in nbnxn_atomdata_t *nbat */
784 static void nbnxn_atomdata_set_charges(nbnxn_atomdata_t* nbat, const Nbnxm::GridSet& gridSet, const real* charge)
786 if (nbat->XFormat != nbatXYZQ)
788 nbat->paramsDeprecated().q.resize(nbat->numAtoms());
791 for (const Nbnxm::Grid& grid : gridSet.grids())
793 /* Loop over all columns and copy and fill */
794 for (int cxy = 0; cxy < grid.numColumns(); cxy++)
796 const int atomOffset = grid.firstAtomInColumn(cxy);
797 const int numAtoms = grid.numAtomsInColumn(cxy);
798 const int paddedNumAtoms = grid.paddedNumAtomsInColumn(cxy);
800 if (nbat->XFormat == nbatXYZQ)
802 real* q = nbat->x().data() + atomOffset * STRIDE_XYZQ + ZZ + 1;
804 for (i = 0; i < numAtoms; i++)
806 *q = charge[gridSet.atomIndices()[atomOffset + i]];
809 /* Complete the partially filled last cell with zeros */
810 for (; i < paddedNumAtoms; i++)
818 real* q = nbat->paramsDeprecated().q.data() + atomOffset;
820 for (i = 0; i < numAtoms; i++)
822 *q = charge[gridSet.atomIndices()[atomOffset + i]];
825 /* Complete the partially filled last cell with zeros */
826 for (; i < paddedNumAtoms; i++)
836 /* Set the charges of perturbed atoms in nbnxn_atomdata_t to 0.
837 * This is to automatically remove the RF/PME self term in the nbnxn kernels.
838 * Part of the zero interactions are still calculated in the normal kernels.
839 * All perturbed interactions are calculated in the free energy kernel,
840 * using the original charge and LJ data, not nbnxn_atomdata_t.
842 static void nbnxn_atomdata_mask_fep(nbnxn_atomdata_t* nbat, const Nbnxm::GridSet& gridSet)
844 nbnxn_atomdata_t::Params& params = nbat->paramsDeprecated();
848 if (nbat->XFormat == nbatXYZQ)
850 q = nbat->x().data() + ZZ + 1;
851 stride_q = STRIDE_XYZQ;
859 for (const Nbnxm::Grid& grid : gridSet.grids())
862 if (grid.geometry().isSimple)
868 nsubc = c_gpuNumClusterPerCell;
871 int c_offset = grid.firstAtomInColumn(0);
873 /* Loop over all columns and copy and fill */
874 for (int c = 0; c < grid.numCells() * nsubc; c++)
876 /* Does this cluster contain perturbed particles? */
877 if (grid.clusterIsPerturbed(c))
879 const int numAtomsPerCluster = grid.geometry().numAtomsICluster;
880 for (int i = 0; i < numAtomsPerCluster; i++)
882 /* Is this a perturbed particle? */
883 if (grid.atomIsPerturbed(c, i))
885 int ind = c_offset + c * numAtomsPerCluster + i;
886 /* Set atom type and charge to non-interacting */
887 params.type[ind] = params.numTypes - 1;
888 q[ind * stride_q] = 0;
896 /* Copies the energy group indices to a reordered and packed array */
898 copy_egp_to_nbat_egps(const int* a, int na, int na_round, int na_c, int bit_shift, const int* in, int* innb)
904 for (i = 0; i < na; i += na_c)
906 /* Store na_c energy group numbers into one int */
908 for (int sa = 0; sa < na_c; sa++)
913 comb |= (GET_CGINFO_GID(in[at]) << (sa * bit_shift));
918 /* Complete the partially filled last cell with fill */
919 for (; i < na_round; i += na_c)
925 /* Set the energy group indices for atoms in nbnxn_atomdata_t */
926 static void nbnxn_atomdata_set_energygroups(nbnxn_atomdata_t::Params* params,
927 const Nbnxm::GridSet& gridSet,
930 if (params->nenergrp == 1)
935 params->energrp.resize(gridSet.numGridAtomsTotal());
937 for (const Nbnxm::Grid& grid : gridSet.grids())
939 /* Loop over all columns and copy and fill */
940 for (int i = 0; i < grid.numColumns(); i++)
942 const int numAtoms = grid.paddedNumAtomsInColumn(i);
943 const int atomOffset = grid.firstAtomInColumn(i);
945 copy_egp_to_nbat_egps(gridSet.atomIndices().data() + atomOffset, grid.numAtomsInColumn(i),
946 numAtoms, c_nbnxnCpuIClusterSize, params->neg_2log, atinfo,
947 params->energrp.data() + grid.atomToCluster(atomOffset));
952 /* Sets all required atom parameter data in nbnxn_atomdata_t */
953 void nbnxn_atomdata_set(nbnxn_atomdata_t* nbat,
954 const Nbnxm::GridSet& gridSet,
955 const t_mdatoms* mdatoms,
958 nbnxn_atomdata_t::Params& params = nbat->paramsDeprecated();
960 nbnxn_atomdata_set_atomtypes(¶ms, gridSet, mdatoms->typeA);
962 nbnxn_atomdata_set_charges(nbat, gridSet, mdatoms->chargeA);
964 if (gridSet.haveFep())
966 nbnxn_atomdata_mask_fep(nbat, gridSet);
969 /* This must be done after masking types for FEP */
970 nbnxn_atomdata_set_ljcombparams(¶ms, nbat->XFormat, gridSet);
972 nbnxn_atomdata_set_energygroups(¶ms, gridSet, atinfo);
975 /* Copies the shift vector array to nbnxn_atomdata_t */
976 void nbnxn_atomdata_copy_shiftvec(gmx_bool bDynamicBox, rvec* shift_vec, nbnxn_atomdata_t* nbat)
980 nbat->bDynamicBox = bDynamicBox;
981 for (i = 0; i < SHIFTS; i++)
983 copy_rvec(shift_vec[i], nbat->shift_vec[i]);
987 // This is slightly different from nbnxn_get_atom_range(...) at the end of the file
988 // TODO: Combine if possible
989 static void getAtomRanges(const Nbnxm::GridSet& gridSet,
990 const gmx::AtomLocality locality,
996 case gmx::AtomLocality::All:
998 *gridEnd = gridSet.grids().size();
1000 case gmx::AtomLocality::Local:
1004 case gmx::AtomLocality::NonLocal:
1006 *gridEnd = gridSet.grids().size();
1008 case gmx::AtomLocality::Count:
1009 GMX_ASSERT(false, "Count is invalid locality specifier");
1014 /* Copies (and reorders) the coordinates to nbnxn_atomdata_t */
1015 void nbnxn_atomdata_copy_x_to_nbat_x(const Nbnxm::GridSet& gridSet,
1016 const gmx::AtomLocality locality,
1018 const rvec* coordinates,
1019 nbnxn_atomdata_t* nbat)
1024 getAtomRanges(gridSet, locality, &gridBegin, &gridEnd);
1028 nbat->natoms_local = gridSet.grids()[0].atomIndexEnd();
1031 const int nth = gmx_omp_nthreads_get(emntPairsearch);
1032 #pragma omp parallel for num_threads(nth) schedule(static)
1033 for (int th = 0; th < nth; th++)
1037 for (int g = gridBegin; g < gridEnd; g++)
1039 const Nbnxm::Grid& grid = gridSet.grids()[g];
1040 const int numCellsXY = grid.numColumns();
1042 const int cxy0 = (numCellsXY * th + nth - 1) / nth;
1043 const int cxy1 = (numCellsXY * (th + 1) + nth - 1) / nth;
1045 for (int cxy = cxy0; cxy < cxy1; cxy++)
1047 const int na = grid.numAtomsInColumn(cxy);
1048 const int ash = grid.firstAtomInColumn(cxy);
1051 if (g == 0 && fillLocal)
1053 na_fill = grid.paddedNumAtomsInColumn(cxy);
1057 /* We fill only the real particle locations.
1058 * We assume the filling entries at the end have been
1059 * properly set before during pair-list generation.
1063 copy_rvec_to_nbat_real(gridSet.atomIndices().data() + ash, na, na_fill,
1064 coordinates, nbat->XFormat, nbat->x().data(), ash);
1068 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1072 /* Copies (and reorders) the coordinates to nbnxn_atomdata_t on the GPU*/
1073 void nbnxn_atomdata_x_to_nbat_x_gpu(const Nbnxm::GridSet& gridSet,
1074 const gmx::AtomLocality locality,
1076 gmx_nbnxn_gpu_t* gpu_nbv,
1077 DeviceBuffer<float> d_x,
1078 GpuEventSynchronizer* xReadyOnDevice)
1083 getAtomRanges(gridSet, locality, &gridBegin, &gridEnd);
1085 for (int g = gridBegin; g < gridEnd; g++)
1087 nbnxn_gpu_x_to_nbat_x(gridSet.grids()[g], fillLocal && g == 0, gpu_nbv, d_x, xReadyOnDevice,
1088 locality, g, gridSet.numColumnsMax());
1092 static void nbnxn_atomdata_clear_reals(gmx::ArrayRef<real> dest, int i0, int i1)
1094 for (int i = i0; i < i1; i++)
1100 gmx_unused static void nbnxn_atomdata_reduce_reals(real* gmx_restrict dest,
1102 const real** gmx_restrict src,
1109 /* The destination buffer contains data, add to it */
1110 for (int i = i0; i < i1; i++)
1112 for (int s = 0; s < nsrc; s++)
1114 dest[i] += src[s][i];
1120 /* The destination buffer is unitialized, set it first */
1121 for (int i = i0; i < i1; i++)
1123 dest[i] = src[0][i];
1124 for (int s = 1; s < nsrc; s++)
1126 dest[i] += src[s][i];
1132 gmx_unused static void nbnxn_atomdata_reduce_reals_simd(real gmx_unused* gmx_restrict dest,
1133 gmx_bool gmx_unused bDestSet,
1134 const gmx_unused real** gmx_restrict src,
1135 int gmx_unused nsrc,
1140 /* The SIMD width here is actually independent of that in the kernels,
1141 * but we use the same width for simplicity (usually optimal anyhow).
1143 SimdReal dest_SSE, src_SSE;
1147 for (int i = i0; i < i1; i += GMX_SIMD_REAL_WIDTH)
1149 dest_SSE = load<SimdReal>(dest + i);
1150 for (int s = 0; s < nsrc; s++)
1152 src_SSE = load<SimdReal>(src[s] + i);
1153 dest_SSE = dest_SSE + src_SSE;
1155 store(dest + i, dest_SSE);
1160 for (int i = i0; i < i1; i += GMX_SIMD_REAL_WIDTH)
1162 dest_SSE = load<SimdReal>(src[0] + i);
1163 for (int s = 1; s < nsrc; s++)
1165 src_SSE = load<SimdReal>(src[s] + i);
1166 dest_SSE = dest_SSE + src_SSE;
1168 store(dest + i, dest_SSE);
1174 /* Add part of the force array(s) from nbnxn_atomdata_t to f
1176 * Note: Adding restrict to f makes this function 50% slower with gcc 7.3
1178 static void nbnxn_atomdata_add_nbat_f_to_f_part(const Nbnxm::GridSet& gridSet,
1179 const nbnxn_atomdata_t& nbat,
1180 const nbnxn_atomdata_output_t& out,
1185 gmx::ArrayRef<const int> cell = gridSet.cells();
1186 // Note: Using ArrayRef instead makes this code 25% slower with gcc 7.3
1187 const real* fnb = out.f.data();
1189 /* Loop over all columns and copy and fill */
1190 switch (nbat.FFormat)
1194 for (int a = a0; a < a1; a++)
1196 int i = cell[a] * nbat.fstride;
1199 f[a][YY] += fnb[i + 1];
1200 f[a][ZZ] += fnb[i + 2];
1204 for (int a = a0; a < a1; a++)
1206 int i = atom_to_x_index<c_packX4>(cell[a]);
1208 f[a][XX] += fnb[i + XX * c_packX4];
1209 f[a][YY] += fnb[i + YY * c_packX4];
1210 f[a][ZZ] += fnb[i + ZZ * c_packX4];
1214 for (int a = a0; a < a1; a++)
1216 int i = atom_to_x_index<c_packX8>(cell[a]);
1218 f[a][XX] += fnb[i + XX * c_packX8];
1219 f[a][YY] += fnb[i + YY * c_packX8];
1220 f[a][ZZ] += fnb[i + ZZ * c_packX8];
1223 default: gmx_incons("Unsupported nbnxn_atomdata_t format");
1227 static inline unsigned char reverse_bits(unsigned char b)
1229 /* http://graphics.stanford.edu/~seander/bithacks.html#ReverseByteWith64BitsDiv */
1230 return (b * 0x0202020202ULL & 0x010884422010ULL) % 1023;
1233 static void nbnxn_atomdata_add_nbat_f_to_f_treereduce(nbnxn_atomdata_t* nbat, int nth)
1235 const nbnxn_buffer_flags_t* flags = &nbat->buffer_flags;
1237 int next_pow2 = 1 << (gmx::log2I(nth - 1) + 1);
1239 const int numOutputBuffers = nbat->out.size();
1240 GMX_ASSERT(numOutputBuffers == nth,
1241 "tree-reduce currently only works for numOutputBuffers==nth");
1243 memset(nbat->syncStep, 0, sizeof(*(nbat->syncStep)) * nth);
1245 #pragma omp parallel num_threads(nth)
1253 th = gmx_omp_get_thread_num();
1255 for (group_size = 2; group_size < 2 * next_pow2; group_size *= 2)
1257 int index[2], group_pos, partner_pos, wu;
1258 int partner_th = th ^ (group_size / 2);
1263 /* wait on partner thread - replaces full barrier */
1264 int sync_th, sync_group_size;
1266 tMPI_Atomic_memory_barrier(); /* gurantee data is saved before marking work as done */
1267 tMPI_Atomic_set(&(nbat->syncStep[th]), group_size / 2); /* mark previous step as completed */
1269 /* find thread to sync with. Equal to partner_th unless nth is not a power of two. */
1270 for (sync_th = partner_th, sync_group_size = group_size;
1271 sync_th >= nth && sync_group_size > 2; sync_group_size /= 2)
1273 sync_th &= ~(sync_group_size / 4);
1275 if (sync_th < nth) /* otherwise nothing to sync index[1] will be >=nout */
1277 /* wait on the thread which computed input data in previous step */
1278 while (tMPI_Atomic_get(static_cast<volatile tMPI_Atomic_t*>(&(nbat->syncStep[sync_th])))
1283 /* guarantee that no later load happens before wait loop is finisehd */
1284 tMPI_Atomic_memory_barrier();
1286 #else /* TMPI_ATOMICS */
1287 # pragma omp barrier
1291 /* Calculate buffers to sum (result goes into first buffer) */
1292 group_pos = th % group_size;
1293 index[0] = th - group_pos;
1294 index[1] = index[0] + group_size / 2;
1296 /* If no second buffer, nothing to do */
1297 if (index[1] >= numOutputBuffers && group_size > 2)
1302 #if NBNXN_BUFFERFLAG_MAX_THREADS > 256
1303 # error reverse_bits assumes max 256 threads
1305 /* Position is permuted so that one of the 2 vectors being added was computed on the same thread in the previous step.
1306 This improves locality and enables to sync with just a single thread between steps (=the levels in the btree).
1307 The permutation which allows this corresponds to reversing the bits of the group position.
1309 group_pos = reverse_bits(group_pos) / (256 / group_size);
1311 partner_pos = group_pos ^ 1;
1313 /* loop over two work-units (own and partner) */
1314 for (wu = 0; wu < 2; wu++)
1318 if (partner_th < nth)
1320 break; /* partner exists we don't have to do his work */
1324 group_pos = partner_pos;
1328 /* Calculate the cell-block range for our thread */
1329 b0 = (flags->nflag * group_pos) / group_size;
1330 b1 = (flags->nflag * (group_pos + 1)) / group_size;
1332 for (b = b0; b < b1; b++)
1334 i0 = b * NBNXN_BUFFERFLAG_SIZE * nbat->fstride;
1335 i1 = (b + 1) * NBNXN_BUFFERFLAG_SIZE * nbat->fstride;
1337 if (bitmask_is_set(flags->flag[b], index[1]) || group_size > 2)
1339 const real* fIndex1 = nbat->out[index[1]].f.data();
1341 nbnxn_atomdata_reduce_reals_simd
1343 nbnxn_atomdata_reduce_reals
1345 (nbat->out[index[0]].f.data(),
1346 bitmask_is_set(flags->flag[b], index[0]) || group_size > 2,
1347 &fIndex1, 1, i0, i1);
1349 else if (!bitmask_is_set(flags->flag[b], index[0]))
1351 nbnxn_atomdata_clear_reals(nbat->out[index[0]].f, i0, i1);
1357 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1362 static void nbnxn_atomdata_add_nbat_f_to_f_stdreduce(nbnxn_atomdata_t* nbat, int nth)
1364 #pragma omp parallel for num_threads(nth) schedule(static)
1365 for (int th = 0; th < nth; th++)
1369 const nbnxn_buffer_flags_t* flags;
1371 const real* fptr[NBNXN_BUFFERFLAG_MAX_THREADS];
1373 flags = &nbat->buffer_flags;
1375 /* Calculate the cell-block range for our thread */
1376 int b0 = (flags->nflag * th) / nth;
1377 int b1 = (flags->nflag * (th + 1)) / nth;
1379 for (int b = b0; b < b1; b++)
1381 int i0 = b * NBNXN_BUFFERFLAG_SIZE * nbat->fstride;
1382 int i1 = (b + 1) * NBNXN_BUFFERFLAG_SIZE * nbat->fstride;
1385 for (gmx::index out = 1; out < gmx::ssize(nbat->out); out++)
1387 if (bitmask_is_set(flags->flag[b], out))
1389 fptr[nfptr++] = nbat->out[out].f.data();
1395 nbnxn_atomdata_reduce_reals_simd
1397 nbnxn_atomdata_reduce_reals
1399 (nbat->out[0].f.data(), bitmask_is_set(flags->flag[b], 0), fptr, nfptr, i0, i1);
1401 else if (!bitmask_is_set(flags->flag[b], 0))
1403 nbnxn_atomdata_clear_reals(nbat->out[0].f, i0, i1);
1407 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1412 /* Add the force array(s) from nbnxn_atomdata_t to f */
1413 void reduceForces(nbnxn_atomdata_t* nbat, const gmx::AtomLocality locality, const Nbnxm::GridSet& gridSet, rvec* f)
1418 nbnxn_get_atom_range(locality, gridSet, &a0, &na);
1422 /* The are no atoms for this reduction, avoid some overhead */
1426 int nth = gmx_omp_nthreads_get(emntNonbonded);
1428 if (nbat->out.size() > 1)
1430 if (locality != gmx::AtomLocality::All)
1432 gmx_incons("add_f_to_f called with nout>1 and locality!=eatAll");
1435 /* Reduce the force thread output buffers into buffer 0, before adding
1436 * them to the, differently ordered, "real" force buffer.
1438 if (nbat->bUseTreeReduce)
1440 nbnxn_atomdata_add_nbat_f_to_f_treereduce(nbat, nth);
1444 nbnxn_atomdata_add_nbat_f_to_f_stdreduce(nbat, nth);
1447 #pragma omp parallel for num_threads(nth) schedule(static)
1448 for (int th = 0; th < nth; th++)
1452 nbnxn_atomdata_add_nbat_f_to_f_part(gridSet, *nbat, nbat->out[0], a0 + ((th + 0) * na) / nth,
1453 a0 + ((th + 1) * na) / nth, f);
1455 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1459 /* Add the force array(s) from nbnxn_atomdata_t to f */
1460 void reduceForcesGpu(const gmx::AtomLocality locality,
1461 DeviceBuffer<float> totalForcesDevice,
1462 const Nbnxm::GridSet& gridSet,
1463 void* pmeForcesDevice,
1464 gmx::ArrayRef<GpuEventSynchronizer* const> dependencyList,
1465 gmx_nbnxn_gpu_t* gpu_nbv,
1466 bool useGpuFPmeReduction,
1467 bool accumulateForce)
1472 nbnxn_get_atom_range(locality, gridSet, &atomsStart, &numAtoms);
1476 /* The are no atoms for this reduction, avoid some overhead */
1480 Nbnxm::nbnxn_gpu_add_nbat_f_to_f(locality, totalForcesDevice, gpu_nbv, pmeForcesDevice, dependencyList,
1481 atomsStart, numAtoms, useGpuFPmeReduction, accumulateForce);
1484 void nbnxn_atomdata_add_nbat_fshift_to_fshift(const nbnxn_atomdata_t& nbat, gmx::ArrayRef<gmx::RVec> fshift)
1486 gmx::ArrayRef<const nbnxn_atomdata_output_t> outputBuffers = nbat.out;
1488 for (int s = 0; s < SHIFTS; s++)
1492 for (const nbnxn_atomdata_output_t& out : outputBuffers)
1494 sum[XX] += out.fshift[s * DIM + XX];
1495 sum[YY] += out.fshift[s * DIM + YY];
1496 sum[ZZ] += out.fshift[s * DIM + ZZ];
1502 void nbnxn_get_atom_range(const gmx::AtomLocality atomLocality,
1503 const Nbnxm::GridSet& gridSet,
1508 switch (atomLocality)
1510 case gmx::AtomLocality::All:
1512 *nAtoms = gridSet.numRealAtomsTotal();
1514 case gmx::AtomLocality::Local:
1516 *nAtoms = gridSet.numRealAtomsLocal();
1518 case gmx::AtomLocality::NonLocal:
1519 *atomStart = gridSet.numRealAtomsLocal();
1520 *nAtoms = gridSet.numRealAtomsTotal() - gridSet.numRealAtomsLocal();
1522 case gmx::AtomLocality::Count:
1523 GMX_ASSERT(false, "Count is invalid locality specifier");