2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
38 #include "nbnxn_atomdata.h"
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/mdlib/nb_verlet.h"
55 #include "gromacs/mdlib/nbnxn_consts.h"
56 #include "gromacs/mdlib/nbnxn_internal.h"
57 #include "gromacs/mdlib/nbnxn_search.h"
58 #include "gromacs/mdlib/nbnxn_util.h"
59 #include "gromacs/mdtypes/forcerec.h" // only for GET_CGINFO_*
60 #include "gromacs/mdtypes/mdatom.h"
61 #include "gromacs/pbcutil/ishift.h"
62 #include "gromacs/simd/simd.h"
63 #include "gromacs/timing/wallcycle.h"
64 #include "gromacs/utility/exceptions.h"
65 #include "gromacs/utility/fatalerror.h"
66 #include "gromacs/utility/gmxomp.h"
67 #include "gromacs/utility/logger.h"
68 #include "gromacs/utility/smalloc.h"
69 #include "gromacs/utility/strconvert.h"
70 #include "gromacs/utility/stringutil.h"
72 using namespace gmx; // TODO: Remove when this file is moved into gmx namespace
74 /* Default nbnxn allocation routine, allocates NBNXN_MEM_ALIGN byte aligned */
75 void nbnxn_alloc_aligned(void **ptr, size_t nbytes)
77 *ptr = save_malloc_aligned("ptr", __FILE__, __LINE__, nbytes, 1, NBNXN_MEM_ALIGN);
80 /* Free function for memory allocated with nbnxn_alloc_aligned */
81 void nbnxn_free_aligned(void *ptr)
86 /* Reallocation wrapper function for nbnxn data structures */
87 void nbnxn_realloc_void(void **ptr,
88 int nbytes_copy, int nbytes_new,
94 ma(&ptr_new, nbytes_new);
96 if (nbytes_new > 0 && ptr_new == nullptr)
98 gmx_fatal(FARGS, "Allocation of %d bytes failed", nbytes_new);
103 if (nbytes_new < nbytes_copy)
105 gmx_incons("In nbnxn_realloc_void: new size less than copy size");
107 memcpy(ptr_new, *ptr, nbytes_copy);
116 /* Reallocate the nbnxn_atomdata_t for a size of n atoms */
117 void nbnxn_atomdata_realloc(nbnxn_atomdata_t *nbat, int n)
119 GMX_ASSERT(nbat->nalloc >= nbat->natoms, "We should have at least as many elelements allocated as there are set");
123 nbnxn_realloc_void(reinterpret_cast<void **>(&nbat->type),
124 nbat->natoms*sizeof(*nbat->type),
125 n*sizeof(*nbat->type),
126 nbat->alloc, nbat->free);
127 nbnxn_realloc_void(reinterpret_cast<void **>(&nbat->lj_comb),
128 nbat->natoms*2*sizeof(*nbat->lj_comb),
129 n*2*sizeof(*nbat->lj_comb),
130 nbat->alloc, nbat->free);
131 if (nbat->XFormat != nbatXYZQ)
133 nbnxn_realloc_void(reinterpret_cast<void **>(&nbat->q),
134 nbat->natoms*sizeof(*nbat->q),
136 nbat->alloc, nbat->free);
138 if (nbat->nenergrp > 1)
140 nbnxn_realloc_void(reinterpret_cast<void **>(&nbat->energrp),
141 nbat->natoms/nbat->na_c*sizeof(*nbat->energrp),
142 n/nbat->na_c*sizeof(*nbat->energrp),
143 nbat->alloc, nbat->free);
145 nbnxn_realloc_void(reinterpret_cast<void **>(&nbat->x),
146 nbat->natoms*nbat->xstride*sizeof(*nbat->x),
147 n*nbat->xstride*sizeof(*nbat->x),
148 nbat->alloc, nbat->free);
149 for (t = 0; t < nbat->nout; t++)
151 /* Allocate one element extra for possible signaling with GPUs */
152 nbnxn_realloc_void(reinterpret_cast<void **>(&nbat->out[t].f),
153 nbat->natoms*nbat->fstride*sizeof(*nbat->out[t].f),
154 n*nbat->fstride*sizeof(*nbat->out[t].f),
155 nbat->alloc, nbat->free);
160 /* Initializes an nbnxn_atomdata_output_t data structure */
161 static void nbnxn_atomdata_output_init(nbnxn_atomdata_output_t *out,
163 int nenergrp, int stride,
167 ma(reinterpret_cast<void **>(&out->fshift), SHIFTS*DIM*sizeof(*out->fshift));
168 out->nV = nenergrp*nenergrp;
169 ma(reinterpret_cast<void **>(&out->Vvdw), out->nV*sizeof(*out->Vvdw));
170 ma(reinterpret_cast<void **>(&out->Vc), out->nV*sizeof(*out->Vc ));
172 if (nb_kernel_type == nbnxnk4xN_SIMD_4xN ||
173 nb_kernel_type == nbnxnk4xN_SIMD_2xNN)
175 int cj_size = nbnxn_kernel_to_cluster_j_size(nb_kernel_type);
176 out->nVS = nenergrp*nenergrp*stride*(cj_size>>1)*cj_size;
177 ma(reinterpret_cast<void **>(&out->VSvdw), out->nVS*sizeof(*out->VSvdw));
178 ma(reinterpret_cast<void **>(&out->VSc), out->nVS*sizeof(*out->VSc ));
186 static void copy_int_to_nbat_int(const int *a, int na, int na_round,
187 const int *in, int fill, int *innb)
192 for (i = 0; i < na; i++)
194 innb[j++] = in[a[i]];
196 /* Complete the partially filled last cell with fill */
197 for (; i < na_round; i++)
203 void copy_rvec_to_nbat_real(const int *a, int na, int na_round,
204 const rvec *x, int nbatFormat,
207 /* We complete partially filled cells, can only be the last one in each
208 * column, with coordinates farAway. The actual coordinate value does
209 * not influence the results, since these filler particles do not interact.
210 * Clusters with normal atoms + fillers have a bounding box based only
211 * on the coordinates of the atoms. Clusters with only fillers have as
212 * the bounding box the coordinates of the first filler. Such clusters
213 * are not considered as i-entries, but they are considered as j-entries.
214 * So for performance it is better to have their bounding boxes far away,
215 * such that filler only clusters don't end up in the pair list.
217 const real farAway = -1000000;
225 for (i = 0; i < na; i++)
227 xnb[j++] = x[a[i]][XX];
228 xnb[j++] = x[a[i]][YY];
229 xnb[j++] = x[a[i]][ZZ];
231 /* Complete the partially filled last cell with farAway elements */
232 for (; i < na_round; i++)
241 for (i = 0; i < na; i++)
243 xnb[j++] = x[a[i]][XX];
244 xnb[j++] = x[a[i]][YY];
245 xnb[j++] = x[a[i]][ZZ];
248 /* Complete the partially filled last cell with zeros */
249 for (; i < na_round; i++)
258 j = atom_to_x_index<c_packX4>(a0);
259 c = a0 & (c_packX4-1);
260 for (i = 0; i < na; i++)
262 xnb[j+XX*c_packX4] = x[a[i]][XX];
263 xnb[j+YY*c_packX4] = x[a[i]][YY];
264 xnb[j+ZZ*c_packX4] = x[a[i]][ZZ];
269 j += (DIM-1)*c_packX4;
273 /* Complete the partially filled last cell with zeros */
274 for (; i < na_round; i++)
276 xnb[j+XX*c_packX4] = farAway;
277 xnb[j+YY*c_packX4] = farAway;
278 xnb[j+ZZ*c_packX4] = farAway;
283 j += (DIM-1)*c_packX4;
289 j = atom_to_x_index<c_packX8>(a0);
290 c = a0 & (c_packX8 - 1);
291 for (i = 0; i < na; i++)
293 xnb[j+XX*c_packX8] = x[a[i]][XX];
294 xnb[j+YY*c_packX8] = x[a[i]][YY];
295 xnb[j+ZZ*c_packX8] = x[a[i]][ZZ];
300 j += (DIM-1)*c_packX8;
304 /* Complete the partially filled last cell with zeros */
305 for (; i < na_round; i++)
307 xnb[j+XX*c_packX8] = farAway;
308 xnb[j+YY*c_packX8] = farAway;
309 xnb[j+ZZ*c_packX8] = farAway;
314 j += (DIM-1)*c_packX8;
320 gmx_incons("Unsupported nbnxn_atomdata_t format");
324 /* Stores the LJ parameter data in a format convenient for different kernels */
325 static void set_lj_parameter_data(nbnxn_atomdata_t *nbat, gmx_bool bSIMD)
329 int nt = nbat->ntype;
334 /* nbfp_aligned stores two parameters using the stride most suitable
335 * for the present SIMD architecture, as specified by the constant
336 * c_simdBestPairAlignment from the SIMD header.
337 * There's a slight inefficiency in allocating and initializing nbfp_aligned
338 * when it might not be used, but introducing the conditional code is not
341 nbat->alloc(reinterpret_cast<void **>(&nbat->nbfp_aligned),
342 nt*nt*c_simdBestPairAlignment*sizeof(*nbat->nbfp_aligned));
343 for (int i = 0; i < nt; i++)
345 for (int j = 0; j < nt; j++)
347 nbat->nbfp_aligned[(i*nt+j)*c_simdBestPairAlignment+0] = nbat->nbfp[(i*nt+j)*2+0];
348 nbat->nbfp_aligned[(i*nt+j)*c_simdBestPairAlignment+1] = nbat->nbfp[(i*nt+j)*2+1];
349 nbat->nbfp_aligned[(i*nt+j)*c_simdBestPairAlignment+2] = 0;
350 nbat->nbfp_aligned[(i*nt+j)*c_simdBestPairAlignment+3] = 0;
356 /* We use combination rule data for SIMD combination rule kernels
357 * and with LJ-PME kernels. We then only need parameters per atom type,
358 * not per pair of atom types.
360 switch (nbat->comb_rule)
363 nbat->comb_rule = ljcrGEOM;
365 for (int i = 0; i < nt; i++)
367 /* Store the sqrt of the diagonal from the nbfp matrix */
368 nbat->nbfp_comb[i*2 ] = std::sqrt(nbat->nbfp[(i*nt+i)*2 ]);
369 nbat->nbfp_comb[i*2+1] = std::sqrt(nbat->nbfp[(i*nt+i)*2+1]);
373 for (int i = 0; i < nt; i++)
375 /* Get 6*C6 and 12*C12 from the diagonal of the nbfp matrix */
376 c6 = nbat->nbfp[(i*nt+i)*2 ];
377 c12 = nbat->nbfp[(i*nt+i)*2+1];
378 if (c6 > 0 && c12 > 0)
380 /* We store 0.5*2^1/6*sigma and sqrt(4*3*eps),
381 * so we get 6*C6 and 12*C12 after combining.
383 nbat->nbfp_comb[i*2 ] = 0.5*gmx::sixthroot(c12/c6);
384 nbat->nbfp_comb[i*2+1] = std::sqrt(c6*c6/c12);
388 nbat->nbfp_comb[i*2 ] = 0;
389 nbat->nbfp_comb[i*2+1] = 0;
394 /* We always store the full matrix (see code above) */
397 gmx_incons("Unknown combination rule");
403 nbnxn_atomdata_init_simple_exclusion_masks(nbnxn_atomdata_t *nbat)
405 const int simd_width = GMX_SIMD_REAL_WIDTH;
407 /* Set the diagonal cluster pair exclusion mask setup data.
408 * In the kernel we check 0 < j - i to generate the masks.
409 * Here we store j - i for generating the mask for the first i,
410 * we substract 0.5 to avoid rounding issues.
411 * In the kernel we can subtract 1 to generate the subsequent mask.
413 int simd_4xn_diag_size;
415 simd_4xn_diag_size = std::max(NBNXN_CPU_CLUSTER_I_SIZE, simd_width);
416 snew_aligned(nbat->simd_4xn_diagonal_j_minus_i, simd_4xn_diag_size, NBNXN_MEM_ALIGN);
417 for (int j = 0; j < simd_4xn_diag_size; j++)
419 nbat->simd_4xn_diagonal_j_minus_i[j] = j - 0.5;
422 snew_aligned(nbat->simd_2xnn_diagonal_j_minus_i, simd_width, NBNXN_MEM_ALIGN);
423 for (int j = 0; j < simd_width/2; j++)
425 /* The j-cluster size is half the SIMD width */
426 nbat->simd_2xnn_diagonal_j_minus_i[j] = j - 0.5;
427 /* The next half of the SIMD width is for i + 1 */
428 nbat->simd_2xnn_diagonal_j_minus_i[simd_width/2+j] = j - 1 - 0.5;
431 /* We use up to 32 bits for exclusion masking.
432 * The same masks are used for the 4xN and 2x(N+N) kernels.
433 * The masks are read either into integer SIMD registers or into
434 * real SIMD registers (together with a cast).
435 * In single precision this means the real and integer SIMD registers
438 simd_excl_size = NBNXN_CPU_CLUSTER_I_SIZE*simd_width;
439 #if GMX_DOUBLE && !GMX_SIMD_HAVE_INT32_LOGICAL
440 snew_aligned(nbat->simd_exclusion_filter64, simd_excl_size, NBNXN_MEM_ALIGN);
442 snew_aligned(nbat->simd_exclusion_filter, simd_excl_size, NBNXN_MEM_ALIGN);
445 for (int j = 0; j < simd_excl_size; j++)
447 /* Set the consecutive bits for masking pair exclusions */
448 #if GMX_DOUBLE && !GMX_SIMD_HAVE_INT32_LOGICAL
449 nbat->simd_exclusion_filter64[j] = (1U << j);
451 nbat->simd_exclusion_filter[j] = (1U << j);
455 #if !GMX_SIMD_HAVE_LOGICAL && !GMX_SIMD_HAVE_INT32_LOGICAL
456 // If the SIMD implementation has no bitwise logical operation support
457 // whatsoever we cannot use the normal masking. Instead,
458 // we generate a vector of all 2^4 possible ways an i atom
459 // interacts with its 4 j atoms. Each array entry contains
460 // GMX_SIMD_REAL_WIDTH values that are read with a single aligned SIMD load.
461 // Since there is no logical value representation in this case, we use
462 // any nonzero value to indicate 'true', while zero mean 'false'.
463 // This can then be converted to a SIMD boolean internally in the SIMD
464 // module by comparing to zero.
465 // Each array entry encodes how this i atom will interact with the 4 j atoms.
466 // Matching code exists in set_ci_top_excls() to generate indices into this array.
467 // Those indices are used in the kernels.
469 simd_excl_size = NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
470 const real simdFalse = 0.0;
471 const real simdTrue = 1.0;
472 real *simd_interaction_array;
474 snew_aligned(simd_interaction_array, simd_excl_size * GMX_SIMD_REAL_WIDTH, NBNXN_MEM_ALIGN);
475 for (int j = 0; j < simd_excl_size; j++)
477 int index = j * GMX_SIMD_REAL_WIDTH;
478 for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
480 simd_interaction_array[index + i] = (j & (1 << i)) ? simdTrue : simdFalse;
483 nbat->simd_interaction_array = simd_interaction_array;
488 /* Initializes an nbnxn_atomdata_t data structure */
489 void nbnxn_atomdata_init(const gmx::MDLogger &mdlog,
490 nbnxn_atomdata_t *nbat,
492 int enbnxninitcombrule,
493 int ntype, const real *nbfp,
496 nbnxn_alloc_t *alloc,
502 gmx_bool simple, bCombGeom, bCombLB, bSIMD;
504 if (alloc == nullptr)
506 nbat->alloc = nbnxn_alloc_aligned;
514 nbat->free = nbnxn_free_aligned;
523 fprintf(debug, "There are %d atom types in the system, adding one for nbnxn_atomdata_t\n", ntype);
525 nbat->ntype = ntype + 1;
526 nbat->alloc(reinterpret_cast<void **>(&nbat->nbfp),
527 nbat->ntype*nbat->ntype*2*sizeof(*nbat->nbfp));
528 nbat->alloc(reinterpret_cast<void **>(&nbat->nbfp_comb), nbat->ntype*2*sizeof(*nbat->nbfp_comb));
530 /* A tolerance of 1e-5 seems reasonable for (possibly hand-typed)
531 * force-field floating point parameters.
534 ptr = getenv("GMX_LJCOMB_TOL");
539 sscanf(ptr, "%lf", &dbl);
545 /* Temporarily fill nbat->nbfp_comb with sigma and epsilon
546 * to check for the LB rule.
548 for (int i = 0; i < ntype; i++)
550 c6 = nbfp[(i*ntype+i)*2 ]/6.0;
551 c12 = nbfp[(i*ntype+i)*2+1]/12.0;
552 if (c6 > 0 && c12 > 0)
554 nbat->nbfp_comb[i*2 ] = gmx::sixthroot(c12/c6);
555 nbat->nbfp_comb[i*2+1] = 0.25*c6*c6/c12;
557 else if (c6 == 0 && c12 == 0)
559 nbat->nbfp_comb[i*2 ] = 0;
560 nbat->nbfp_comb[i*2+1] = 0;
564 /* Can not use LB rule with only dispersion or repulsion */
569 for (int i = 0; i < nbat->ntype; i++)
571 for (int j = 0; j < nbat->ntype; j++)
573 if (i < ntype && j < ntype)
575 /* fr->nbfp has been updated, so that array too now stores c6/c12 including
576 * the 6.0/12.0 prefactors to save 2 flops in the most common case (force-only).
578 c6 = nbfp[(i*ntype+j)*2 ];
579 c12 = nbfp[(i*ntype+j)*2+1];
580 nbat->nbfp[(i*nbat->ntype+j)*2 ] = c6;
581 nbat->nbfp[(i*nbat->ntype+j)*2+1] = c12;
583 /* Compare 6*C6 and 12*C12 for geometric cobination rule */
584 bCombGeom = bCombGeom &&
585 gmx_within_tol(c6*c6, nbfp[(i*ntype+i)*2 ]*nbfp[(j*ntype+j)*2 ], tol) &&
586 gmx_within_tol(c12*c12, nbfp[(i*ntype+i)*2+1]*nbfp[(j*ntype+j)*2+1], tol);
588 /* Compare C6 and C12 for Lorentz-Berthelot combination rule */
592 ((c6 == 0 && c12 == 0 &&
593 (nbat->nbfp_comb[i*2+1] == 0 || nbat->nbfp_comb[j*2+1] == 0)) ||
594 (c6 > 0 && c12 > 0 &&
595 gmx_within_tol(gmx::sixthroot(c12/c6),
596 0.5*(nbat->nbfp_comb[i*2]+nbat->nbfp_comb[j*2]), tol) &&
597 gmx_within_tol(0.25*c6*c6/c12, std::sqrt(nbat->nbfp_comb[i*2+1]*nbat->nbfp_comb[j*2+1]), tol)));
601 /* Add zero parameters for the additional dummy atom type */
602 nbat->nbfp[(i*nbat->ntype+j)*2 ] = 0;
603 nbat->nbfp[(i*nbat->ntype+j)*2+1] = 0;
609 fprintf(debug, "Combination rules: geometric %s Lorentz-Berthelot %s\n",
610 gmx::boolToString(bCombGeom), gmx::boolToString(bCombLB));
613 simple = nbnxn_kernel_pairlist_simple(nb_kernel_type);
615 switch (enbnxninitcombrule)
617 case enbnxninitcombruleDETECT:
618 /* We prefer the geometic combination rule,
619 * as that gives a slightly faster kernel than the LB rule.
623 nbat->comb_rule = ljcrGEOM;
627 nbat->comb_rule = ljcrLB;
631 nbat->comb_rule = ljcrNONE;
633 nbat->free(nbat->nbfp_comb);
638 if (nbat->comb_rule == ljcrNONE)
640 mesg = "Using full Lennard-Jones parameter combination matrix";
644 mesg = gmx::formatString("Using %s Lennard-Jones combination rule",
645 nbat->comb_rule == ljcrGEOM ? "geometric" : "Lorentz-Berthelot");
647 GMX_LOG(mdlog.info).asParagraph().appendText(mesg);
650 case enbnxninitcombruleGEOM:
651 nbat->comb_rule = ljcrGEOM;
653 case enbnxninitcombruleLB:
654 nbat->comb_rule = ljcrLB;
656 case enbnxninitcombruleNONE:
657 nbat->comb_rule = ljcrNONE;
659 nbat->free(nbat->nbfp_comb);
662 gmx_incons("Unknown enbnxninitcombrule");
665 bSIMD = (nb_kernel_type == nbnxnk4xN_SIMD_4xN ||
666 nb_kernel_type == nbnxnk4xN_SIMD_2xNN);
668 set_lj_parameter_data(nbat, bSIMD);
671 nbat->type = nullptr;
672 nbat->lj_comb = nullptr;
679 pack_x = std::max(NBNXN_CPU_CLUSTER_I_SIZE,
680 nbnxn_kernel_to_cluster_j_size(nb_kernel_type));
684 nbat->XFormat = nbatX4;
687 nbat->XFormat = nbatX8;
690 gmx_incons("Unsupported packing width");
695 nbat->XFormat = nbatXYZ;
698 nbat->FFormat = nbat->XFormat;
702 nbat->XFormat = nbatXYZQ;
703 nbat->FFormat = nbatXYZ;
706 nbat->nenergrp = n_energygroups;
709 // We now check for energy groups already when starting mdrun
710 GMX_RELEASE_ASSERT(n_energygroups == 1, "GPU kernels do not support energy groups");
712 /* Temporary storage goes as #grp^3*simd_width^2/2, so limit to 64 */
713 if (nbat->nenergrp > 64)
715 gmx_fatal(FARGS, "With NxN kernels not more than 64 energy groups are supported\n");
718 while (nbat->nenergrp > (1<<nbat->neg_2log))
722 nbat->energrp = nullptr;
723 nbat->alloc(reinterpret_cast<void **>(&nbat->shift_vec), SHIFTS*sizeof(*nbat->shift_vec));
724 nbat->xstride = (nbat->XFormat == nbatXYZQ ? STRIDE_XYZQ : DIM);
725 nbat->fstride = (nbat->FFormat == nbatXYZQ ? STRIDE_XYZQ : DIM);
731 nbnxn_atomdata_init_simple_exclusion_masks(nbat);
735 /* Initialize the output data structures */
737 snew(nbat->out, nbat->nout);
739 for (int i = 0; i < nbat->nout; i++)
741 nbnxn_atomdata_output_init(&nbat->out[i],
743 nbat->nenergrp, 1<<nbat->neg_2log,
746 nbat->buffer_flags.flag = nullptr;
747 nbat->buffer_flags.flag_nalloc = 0;
749 nth = gmx_omp_nthreads_get(emntNonbonded);
751 ptr = getenv("GMX_USE_TREEREDUCE");
754 nbat->bUseTreeReduce = strtol(ptr, nullptr, 10);
757 else if (nth > 8) /*on the CPU we currently don't benefit even at 32*/
759 nbat->bUseTreeReduce = 1;
764 nbat->bUseTreeReduce = 0;
766 if (nbat->bUseTreeReduce)
768 GMX_LOG(mdlog.info).asParagraph().appendText("Using tree force reduction");
770 snew(nbat->syncStep, nth);
774 template<int packSize>
775 static void copy_lj_to_nbat_lj_comb(const real *ljparam_type,
776 const int *type, int na,
779 /* The LJ params follow the combination rule:
780 * copy the params for the type array to the atom array.
782 for (int is = 0; is < na; is += packSize)
784 for (int k = 0; k < packSize; k++)
787 ljparam_at[is*2 + k] = ljparam_type[type[i]*2 ];
788 ljparam_at[is*2 + packSize + k] = ljparam_type[type[i]*2 + 1];
793 /* Sets the atom type in nbnxn_atomdata_t */
794 static void nbnxn_atomdata_set_atomtypes(nbnxn_atomdata_t *nbat,
795 const nbnxn_search *nbs,
798 for (const nbnxn_grid_t &grid : nbs->grid)
800 /* Loop over all columns and copy and fill */
801 for (int i = 0; i < grid.numCells[XX]*grid.numCells[YY]; i++)
803 int ncz = grid.cxy_ind[i+1] - grid.cxy_ind[i];
804 int ash = (grid.cell0 + grid.cxy_ind[i])*grid.na_sc;
806 copy_int_to_nbat_int(nbs->a.data() + ash, grid.cxy_na[i], ncz*grid.na_sc,
807 type, nbat->ntype-1, nbat->type+ash);
812 /* Sets the LJ combination rule parameters in nbnxn_atomdata_t */
813 static void nbnxn_atomdata_set_ljcombparams(nbnxn_atomdata_t *nbat,
814 const nbnxn_search *nbs)
816 if (nbat->comb_rule != ljcrNONE)
818 for (const nbnxn_grid_t &grid : nbs->grid)
820 /* Loop over all columns and copy and fill */
821 for (int i = 0; i < grid.numCells[XX]*grid.numCells[YY]; i++)
823 int ncz = grid.cxy_ind[i+1] - grid.cxy_ind[i];
824 int ash = (grid.cell0 + grid.cxy_ind[i])*grid.na_sc;
826 if (nbat->XFormat == nbatX4)
828 copy_lj_to_nbat_lj_comb<c_packX4>(nbat->nbfp_comb,
831 nbat->lj_comb + ash*2);
833 else if (nbat->XFormat == nbatX8)
835 copy_lj_to_nbat_lj_comb<c_packX8>(nbat->nbfp_comb,
838 nbat->lj_comb + ash*2);
840 else if (nbat->XFormat == nbatXYZQ)
842 copy_lj_to_nbat_lj_comb<1>(nbat->nbfp_comb,
845 nbat->lj_comb + ash*2);
852 /* Sets the charges in nbnxn_atomdata_t *nbat */
853 static void nbnxn_atomdata_set_charges(nbnxn_atomdata_t *nbat,
854 const nbnxn_search *nbs,
857 for (const nbnxn_grid_t &grid : nbs->grid)
859 /* Loop over all columns and copy and fill */
860 for (int cxy = 0; cxy < grid.numCells[XX]*grid.numCells[YY]; cxy++)
862 int ash = (grid.cell0 + grid.cxy_ind[cxy])*grid.na_sc;
863 int na = grid.cxy_na[cxy];
864 int na_round = (grid.cxy_ind[cxy+1] - grid.cxy_ind[cxy])*grid.na_sc;
866 if (nbat->XFormat == nbatXYZQ)
868 real *q = nbat->x + ash*STRIDE_XYZQ + ZZ + 1;
870 for (i = 0; i < na; i++)
872 *q = charge[nbs->a[ash+i]];
875 /* Complete the partially filled last cell with zeros */
876 for (; i < na_round; i++)
884 real *q = nbat->q + ash;
886 for (i = 0; i < na; i++)
888 *q = charge[nbs->a[ash+i]];
891 /* Complete the partially filled last cell with zeros */
892 for (; i < na_round; i++)
902 /* Set the charges of perturbed atoms in nbnxn_atomdata_t to 0.
903 * This is to automatically remove the RF/PME self term in the nbnxn kernels.
904 * Part of the zero interactions are still calculated in the normal kernels.
905 * All perturbed interactions are calculated in the free energy kernel,
906 * using the original charge and LJ data, not nbnxn_atomdata_t.
908 static void nbnxn_atomdata_mask_fep(nbnxn_atomdata_t *nbat,
909 const nbnxn_search *nbs)
914 if (nbat->XFormat == nbatXYZQ)
916 q = nbat->x + ZZ + 1;
917 stride_q = STRIDE_XYZQ;
925 for (const nbnxn_grid_t &grid : nbs->grid)
933 nsubc = c_gpuNumClusterPerCell;
936 int c_offset = grid.cell0*grid.na_sc;
938 /* Loop over all columns and copy and fill */
939 for (int c = 0; c < grid.nc*nsubc; c++)
941 /* Does this cluster contain perturbed particles? */
942 if (grid.fep[c] != 0)
944 for (int i = 0; i < grid.na_c; i++)
946 /* Is this a perturbed particle? */
947 if (grid.fep[c] & (1 << i))
949 int ind = c_offset + c*grid.na_c + i;
950 /* Set atom type and charge to non-interacting */
951 nbat->type[ind] = nbat->ntype - 1;
960 /* Copies the energy group indices to a reordered and packed array */
961 static void copy_egp_to_nbat_egps(const int *a, int na, int na_round,
962 int na_c, int bit_shift,
963 const int *in, int *innb)
969 for (i = 0; i < na; i += na_c)
971 /* Store na_c energy group numbers into one int */
973 for (int sa = 0; sa < na_c; sa++)
978 comb |= (GET_CGINFO_GID(in[at]) << (sa*bit_shift));
983 /* Complete the partially filled last cell with fill */
984 for (; i < na_round; i += na_c)
990 /* Set the energy group indices for atoms in nbnxn_atomdata_t */
991 static void nbnxn_atomdata_set_energygroups(nbnxn_atomdata_t *nbat,
992 const nbnxn_search *nbs,
995 if (nbat->nenergrp == 1)
1000 for (const nbnxn_grid_t &grid : nbs->grid)
1002 /* Loop over all columns and copy and fill */
1003 for (int i = 0; i < grid.numCells[XX]*grid.numCells[YY]; i++)
1005 int ncz = grid.cxy_ind[i+1] - grid.cxy_ind[i];
1006 int ash = (grid.cell0 + grid.cxy_ind[i])*grid.na_sc;
1008 copy_egp_to_nbat_egps(nbs->a.data() + ash, grid.cxy_na[i], ncz*grid.na_sc,
1009 nbat->na_c, nbat->neg_2log,
1010 atinfo, nbat->energrp+(ash>>grid.na_c_2log));
1015 /* Sets all required atom parameter data in nbnxn_atomdata_t */
1016 void nbnxn_atomdata_set(nbnxn_atomdata_t *nbat,
1017 const nbnxn_search *nbs,
1018 const t_mdatoms *mdatoms,
1021 nbnxn_atomdata_set_atomtypes(nbat, nbs, mdatoms->typeA);
1023 nbnxn_atomdata_set_charges(nbat, nbs, mdatoms->chargeA);
1027 nbnxn_atomdata_mask_fep(nbat, nbs);
1030 /* This must be done after masking types for FEP */
1031 nbnxn_atomdata_set_ljcombparams(nbat, nbs);
1033 nbnxn_atomdata_set_energygroups(nbat, nbs, atinfo);
1036 /* Copies the shift vector array to nbnxn_atomdata_t */
1037 void nbnxn_atomdata_copy_shiftvec(gmx_bool bDynamicBox,
1039 nbnxn_atomdata_t *nbat)
1043 nbat->bDynamicBox = bDynamicBox;
1044 for (i = 0; i < SHIFTS; i++)
1046 copy_rvec(shift_vec[i], nbat->shift_vec[i]);
1050 /* Copies (and reorders) the coordinates to nbnxn_atomdata_t */
1051 void nbnxn_atomdata_copy_x_to_nbat_x(const nbnxn_search *nbs,
1055 nbnxn_atomdata_t *nbat,
1056 gmx_wallcycle *wcycle)
1058 wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1059 wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1068 g1 = nbs->grid.size();
1076 g1 = nbs->grid.size();
1082 nbat->natoms_local = nbs->grid[0].nc*nbs->grid[0].na_sc;
1085 nth = gmx_omp_nthreads_get(emntPairsearch);
1087 #pragma omp parallel for num_threads(nth) schedule(static)
1088 for (th = 0; th < nth; th++)
1092 for (int g = g0; g < g1; g++)
1094 const nbnxn_grid_t &grid = nbs->grid[g];
1095 const int numCellsXY = grid.numCells[XX]*grid.numCells[YY];
1097 const int cxy0 = (numCellsXY* th + nth - 1)/nth;
1098 const int cxy1 = (numCellsXY*(th + 1) + nth - 1)/nth;
1100 for (int cxy = cxy0; cxy < cxy1; cxy++)
1102 int na, ash, na_fill;
1104 na = grid.cxy_na[cxy];
1105 ash = (grid.cell0 + grid.cxy_ind[cxy])*grid.na_sc;
1107 if (g == 0 && FillLocal)
1110 (grid.cxy_ind[cxy+1] - grid.cxy_ind[cxy])*grid.na_sc;
1114 /* We fill only the real particle locations.
1115 * We assume the filling entries at the end have been
1116 * properly set before during pair-list generation.
1120 copy_rvec_to_nbat_real(nbs->a.data() + ash, na, na_fill, x,
1121 nbat->XFormat, nbat->x, ash);
1125 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1128 wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1129 wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1133 nbnxn_atomdata_clear_reals(real * gmx_restrict dest,
1136 for (int i = i0; i < i1; i++)
1142 gmx_unused static void
1143 nbnxn_atomdata_reduce_reals(real * gmx_restrict dest,
1145 real ** gmx_restrict src,
1151 /* The destination buffer contains data, add to it */
1152 for (int i = i0; i < i1; i++)
1154 for (int s = 0; s < nsrc; s++)
1156 dest[i] += src[s][i];
1162 /* The destination buffer is unitialized, set it first */
1163 for (int i = i0; i < i1; i++)
1165 dest[i] = src[0][i];
1166 for (int s = 1; s < nsrc; s++)
1168 dest[i] += src[s][i];
1174 gmx_unused static void
1175 nbnxn_atomdata_reduce_reals_simd(real gmx_unused * gmx_restrict dest,
1176 gmx_bool gmx_unused bDestSet,
1177 real gmx_unused ** gmx_restrict src,
1178 int gmx_unused nsrc,
1179 int gmx_unused i0, int gmx_unused i1)
1182 /* The SIMD width here is actually independent of that in the kernels,
1183 * but we use the same width for simplicity (usually optimal anyhow).
1185 SimdReal dest_SSE, src_SSE;
1189 for (int i = i0; i < i1; i += GMX_SIMD_REAL_WIDTH)
1191 dest_SSE = load<SimdReal>(dest+i);
1192 for (int s = 0; s < nsrc; s++)
1194 src_SSE = load<SimdReal>(src[s]+i);
1195 dest_SSE = dest_SSE + src_SSE;
1197 store(dest+i, dest_SSE);
1202 for (int i = i0; i < i1; i += GMX_SIMD_REAL_WIDTH)
1204 dest_SSE = load<SimdReal>(src[0]+i);
1205 for (int s = 1; s < nsrc; s++)
1207 src_SSE = load<SimdReal>(src[s]+i);
1208 dest_SSE = dest_SSE + src_SSE;
1210 store(dest+i, dest_SSE);
1216 /* Add part of the force array(s) from nbnxn_atomdata_t to f */
1218 nbnxn_atomdata_add_nbat_f_to_f_part(const nbnxn_search *nbs,
1219 const nbnxn_atomdata_t *nbat,
1220 nbnxn_atomdata_output_t *out,
1225 gmx::ArrayRef<const int> cell = nbs->cell;
1227 /* Loop over all columns and copy and fill */
1228 switch (nbat->FFormat)
1234 const real *fnb = out[0].f;
1236 for (int a = a0; a < a1; a++)
1238 int i = cell[a]*nbat->fstride;
1241 f[a][YY] += fnb[i+1];
1242 f[a][ZZ] += fnb[i+2];
1247 for (int a = a0; a < a1; a++)
1249 int i = cell[a]*nbat->fstride;
1251 for (int fa = 0; fa < nfa; fa++)
1253 f[a][XX] += out[fa].f[i];
1254 f[a][YY] += out[fa].f[i+1];
1255 f[a][ZZ] += out[fa].f[i+2];
1263 const real *fnb = out[0].f;
1265 for (int a = a0; a < a1; a++)
1267 int i = atom_to_x_index<c_packX4>(cell[a]);
1269 f[a][XX] += fnb[i+XX*c_packX4];
1270 f[a][YY] += fnb[i+YY*c_packX4];
1271 f[a][ZZ] += fnb[i+ZZ*c_packX4];
1276 for (int a = a0; a < a1; a++)
1278 int i = atom_to_x_index<c_packX4>(cell[a]);
1280 for (int fa = 0; fa < nfa; fa++)
1282 f[a][XX] += out[fa].f[i+XX*c_packX4];
1283 f[a][YY] += out[fa].f[i+YY*c_packX4];
1284 f[a][ZZ] += out[fa].f[i+ZZ*c_packX4];
1292 const real *fnb = out[0].f;
1294 for (int a = a0; a < a1; a++)
1296 int i = atom_to_x_index<c_packX8>(cell[a]);
1298 f[a][XX] += fnb[i+XX*c_packX8];
1299 f[a][YY] += fnb[i+YY*c_packX8];
1300 f[a][ZZ] += fnb[i+ZZ*c_packX8];
1305 for (int a = a0; a < a1; a++)
1307 int i = atom_to_x_index<c_packX8>(cell[a]);
1309 for (int fa = 0; fa < nfa; fa++)
1311 f[a][XX] += out[fa].f[i+XX*c_packX8];
1312 f[a][YY] += out[fa].f[i+YY*c_packX8];
1313 f[a][ZZ] += out[fa].f[i+ZZ*c_packX8];
1319 gmx_incons("Unsupported nbnxn_atomdata_t format");
1323 static inline unsigned char reverse_bits(unsigned char b)
1325 /* http://graphics.stanford.edu/~seander/bithacks.html#ReverseByteWith64BitsDiv */
1326 return (b * 0x0202020202ULL & 0x010884422010ULL) % 1023;
1329 static void nbnxn_atomdata_add_nbat_f_to_f_treereduce(const nbnxn_atomdata_t *nbat,
1332 const nbnxn_buffer_flags_t *flags = &nbat->buffer_flags;
1334 int next_pow2 = 1<<(gmx::log2I(nth-1)+1);
1336 assert(nbat->nout == nth); /* tree-reduce currently only works for nout==nth */
1338 memset(nbat->syncStep, 0, sizeof(*(nbat->syncStep))*nth);
1340 #pragma omp parallel num_threads(nth)
1348 th = gmx_omp_get_thread_num();
1350 for (group_size = 2; group_size < 2*next_pow2; group_size *= 2)
1352 int index[2], group_pos, partner_pos, wu;
1353 int partner_th = th ^ (group_size/2);
1358 /* wait on partner thread - replaces full barrier */
1359 int sync_th, sync_group_size;
1361 tMPI_Atomic_memory_barrier(); /* gurantee data is saved before marking work as done */
1362 tMPI_Atomic_set(&(nbat->syncStep[th]), group_size/2); /* mark previous step as completed */
1364 /* find thread to sync with. Equal to partner_th unless nth is not a power of two. */
1365 for (sync_th = partner_th, sync_group_size = group_size; sync_th >= nth && sync_group_size > 2; sync_group_size /= 2)
1367 sync_th &= ~(sync_group_size/4);
1369 if (sync_th < nth) /* otherwise nothing to sync index[1] will be >=nout */
1371 /* wait on the thread which computed input data in previous step */
1372 while (tMPI_Atomic_get((volatile tMPI_Atomic_t*)&(nbat->syncStep[sync_th])) < group_size/2)
1376 /* guarantee that no later load happens before wait loop is finisehd */
1377 tMPI_Atomic_memory_barrier();
1379 #else /* TMPI_ATOMICS */
1384 /* Calculate buffers to sum (result goes into first buffer) */
1385 group_pos = th % group_size;
1386 index[0] = th - group_pos;
1387 index[1] = index[0] + group_size/2;
1389 /* If no second buffer, nothing to do */
1390 if (index[1] >= nbat->nout && group_size > 2)
1395 #if NBNXN_BUFFERFLAG_MAX_THREADS > 256
1396 #error reverse_bits assumes max 256 threads
1398 /* Position is permuted so that one of the 2 vectors being added was computed on the same thread in the previous step.
1399 This improves locality and enables to sync with just a single thread between steps (=the levels in the btree).
1400 The permutation which allows this corresponds to reversing the bits of the group position.
1402 group_pos = reverse_bits(group_pos)/(256/group_size);
1404 partner_pos = group_pos ^ 1;
1406 /* loop over two work-units (own and partner) */
1407 for (wu = 0; wu < 2; wu++)
1411 if (partner_th < nth)
1413 break; /* partner exists we don't have to do his work */
1417 group_pos = partner_pos;
1421 /* Calculate the cell-block range for our thread */
1422 b0 = (flags->nflag* group_pos )/group_size;
1423 b1 = (flags->nflag*(group_pos+1))/group_size;
1425 for (b = b0; b < b1; b++)
1427 i0 = b *NBNXN_BUFFERFLAG_SIZE*nbat->fstride;
1428 i1 = (b+1)*NBNXN_BUFFERFLAG_SIZE*nbat->fstride;
1430 if (bitmask_is_set(flags->flag[b], index[1]) || group_size > 2)
1433 nbnxn_atomdata_reduce_reals_simd
1435 nbnxn_atomdata_reduce_reals
1437 (nbat->out[index[0]].f,
1438 bitmask_is_set(flags->flag[b], index[0]) || group_size > 2,
1439 &(nbat->out[index[1]].f), 1, i0, i1);
1442 else if (!bitmask_is_set(flags->flag[b], index[0]))
1444 nbnxn_atomdata_clear_reals(nbat->out[index[0]].f,
1451 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1456 static void nbnxn_atomdata_add_nbat_f_to_f_stdreduce(const nbnxn_atomdata_t *nbat,
1459 #pragma omp parallel for num_threads(nth) schedule(static)
1460 for (int th = 0; th < nth; th++)
1464 const nbnxn_buffer_flags_t *flags;
1466 real *fptr[NBNXN_BUFFERFLAG_MAX_THREADS];
1468 flags = &nbat->buffer_flags;
1470 /* Calculate the cell-block range for our thread */
1471 int b0 = (flags->nflag* th )/nth;
1472 int b1 = (flags->nflag*(th+1))/nth;
1474 for (int b = b0; b < b1; b++)
1476 int i0 = b *NBNXN_BUFFERFLAG_SIZE*nbat->fstride;
1477 int i1 = (b+1)*NBNXN_BUFFERFLAG_SIZE*nbat->fstride;
1480 for (int out = 1; out < nbat->nout; out++)
1482 if (bitmask_is_set(flags->flag[b], out))
1484 fptr[nfptr++] = nbat->out[out].f;
1490 nbnxn_atomdata_reduce_reals_simd
1492 nbnxn_atomdata_reduce_reals
1495 bitmask_is_set(flags->flag[b], 0),
1499 else if (!bitmask_is_set(flags->flag[b], 0))
1501 nbnxn_atomdata_clear_reals(nbat->out[0].f,
1506 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1510 /* Add the force array(s) from nbnxn_atomdata_t to f */
1511 void nbnxn_atomdata_add_nbat_f_to_f(nbnxn_search *nbs,
1513 const nbnxn_atomdata_t *nbat,
1515 gmx_wallcycle *wcycle)
1517 wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1518 wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1522 nbs_cycle_start(&nbs->cc[enbsCCreducef]);
1528 na = nbs->natoms_nonlocal;
1532 na = nbs->natoms_local;
1535 a0 = nbs->natoms_local;
1536 na = nbs->natoms_nonlocal - nbs->natoms_local;
1540 int nth = gmx_omp_nthreads_get(emntNonbonded);
1544 if (locality != eatAll)
1546 gmx_incons("add_f_to_f called with nout>1 and locality!=eatAll");
1549 /* Reduce the force thread output buffers into buffer 0, before adding
1550 * them to the, differently ordered, "real" force buffer.
1552 if (nbat->bUseTreeReduce)
1554 nbnxn_atomdata_add_nbat_f_to_f_treereduce(nbat, nth);
1558 nbnxn_atomdata_add_nbat_f_to_f_stdreduce(nbat, nth);
1561 #pragma omp parallel for num_threads(nth) schedule(static)
1562 for (int th = 0; th < nth; th++)
1566 nbnxn_atomdata_add_nbat_f_to_f_part(nbs, nbat,
1573 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1576 nbs_cycle_stop(&nbs->cc[enbsCCreducef]);
1578 wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1579 wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1582 /* Adds the shift forces from nbnxn_atomdata_t to fshift */
1583 void nbnxn_atomdata_add_nbat_fshift_to_fshift(const nbnxn_atomdata_t *nbat,
1586 const nbnxn_atomdata_output_t * out = nbat->out;
1588 for (int s = 0; s < SHIFTS; s++)
1592 for (int th = 0; th < nbat->nout; th++)
1594 sum[XX] += out[th].fshift[s*DIM+XX];
1595 sum[YY] += out[th].fshift[s*DIM+YY];
1596 sum[ZZ] += out[th].fshift[s*DIM+ZZ];
1598 rvec_inc(fshift[s], sum);