2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017, 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/pbcutil/ishift.h"
60 #include "gromacs/simd/simd.h"
61 #include "gromacs/utility/exceptions.h"
62 #include "gromacs/utility/fatalerror.h"
63 #include "gromacs/utility/gmxomp.h"
64 #include "gromacs/utility/logger.h"
65 #include "gromacs/utility/smalloc.h"
66 #include "gromacs/utility/stringutil.h"
68 using namespace gmx; // TODO: Remove when this file is moved into gmx namespace
70 /* Default nbnxn allocation routine, allocates NBNXN_MEM_ALIGN byte aligned */
71 void nbnxn_alloc_aligned(void **ptr, size_t nbytes)
73 *ptr = save_malloc_aligned("ptr", __FILE__, __LINE__, nbytes, 1, NBNXN_MEM_ALIGN);
76 /* Free function for memory allocated with nbnxn_alloc_aligned */
77 void nbnxn_free_aligned(void *ptr)
82 /* Reallocation wrapper function for nbnxn data structures */
83 void nbnxn_realloc_void(void **ptr,
84 int nbytes_copy, int nbytes_new,
90 ma(&ptr_new, nbytes_new);
92 if (nbytes_new > 0 && ptr_new == nullptr)
94 gmx_fatal(FARGS, "Allocation of %d bytes failed", nbytes_new);
99 if (nbytes_new < nbytes_copy)
101 gmx_incons("In nbnxn_realloc_void: new size less than copy size");
103 memcpy(ptr_new, *ptr, nbytes_copy);
112 /* Reallocate the nbnxn_atomdata_t for a size of n atoms */
113 void nbnxn_atomdata_realloc(nbnxn_atomdata_t *nbat, int n)
117 nbnxn_realloc_void((void **)&nbat->type,
118 nbat->natoms*sizeof(*nbat->type),
119 n*sizeof(*nbat->type),
120 nbat->alloc, nbat->free);
121 nbnxn_realloc_void((void **)&nbat->lj_comb,
122 nbat->natoms*2*sizeof(*nbat->lj_comb),
123 n*2*sizeof(*nbat->lj_comb),
124 nbat->alloc, nbat->free);
125 if (nbat->XFormat != nbatXYZQ)
127 nbnxn_realloc_void((void **)&nbat->q,
128 nbat->natoms*sizeof(*nbat->q),
130 nbat->alloc, nbat->free);
132 if (nbat->nenergrp > 1)
134 nbnxn_realloc_void((void **)&nbat->energrp,
135 nbat->natoms/nbat->na_c*sizeof(*nbat->energrp),
136 n/nbat->na_c*sizeof(*nbat->energrp),
137 nbat->alloc, nbat->free);
139 nbnxn_realloc_void((void **)&nbat->x,
140 nbat->natoms*nbat->xstride*sizeof(*nbat->x),
141 n*nbat->xstride*sizeof(*nbat->x),
142 nbat->alloc, nbat->free);
143 for (t = 0; t < nbat->nout; t++)
145 /* Allocate one element extra for possible signaling with GPUs */
146 nbnxn_realloc_void((void **)&nbat->out[t].f,
147 nbat->natoms*nbat->fstride*sizeof(*nbat->out[t].f),
148 n*nbat->fstride*sizeof(*nbat->out[t].f),
149 nbat->alloc, nbat->free);
154 /* Initializes an nbnxn_atomdata_output_t data structure */
155 static void nbnxn_atomdata_output_init(nbnxn_atomdata_output_t *out,
157 int nenergrp, int stride,
161 ma((void **)&out->fshift, SHIFTS*DIM*sizeof(*out->fshift));
162 out->nV = nenergrp*nenergrp;
163 ma((void **)&out->Vvdw, out->nV*sizeof(*out->Vvdw));
164 ma((void **)&out->Vc, out->nV*sizeof(*out->Vc ));
166 if (nb_kernel_type == nbnxnk4xN_SIMD_4xN ||
167 nb_kernel_type == nbnxnk4xN_SIMD_2xNN)
169 int cj_size = nbnxn_kernel_to_cluster_j_size(nb_kernel_type);
170 out->nVS = nenergrp*nenergrp*stride*(cj_size>>1)*cj_size;
171 ma((void **)&out->VSvdw, out->nVS*sizeof(*out->VSvdw));
172 ma((void **)&out->VSc, out->nVS*sizeof(*out->VSc ));
180 static void copy_int_to_nbat_int(const int *a, int na, int na_round,
181 const int *in, int fill, int *innb)
186 for (i = 0; i < na; i++)
188 innb[j++] = in[a[i]];
190 /* Complete the partially filled last cell with fill */
191 for (; i < na_round; i++)
197 void copy_rvec_to_nbat_real(const int *a, int na, int na_round,
198 const rvec *x, int nbatFormat,
201 /* We complete partially filled cells, can only be the last one in each
202 * column, with coordinates farAway. The actual coordinate value does
203 * not influence the results, since these filler particles do not interact.
204 * Clusters with normal atoms + fillers have a bounding box based only
205 * on the coordinates of the atoms. Clusters with only fillers have as
206 * the bounding box the coordinates of the first filler. Such clusters
207 * are not considered as i-entries, but they are considered as j-entries.
208 * So for performance it is better to have their bounding boxes far away,
209 * such that filler only clusters don't end up in the pair list.
211 const real farAway = -1000000;
219 for (i = 0; i < na; i++)
221 xnb[j++] = x[a[i]][XX];
222 xnb[j++] = x[a[i]][YY];
223 xnb[j++] = x[a[i]][ZZ];
225 /* Complete the partially filled last cell with farAway elements */
226 for (; i < na_round; i++)
235 for (i = 0; i < na; i++)
237 xnb[j++] = x[a[i]][XX];
238 xnb[j++] = x[a[i]][YY];
239 xnb[j++] = x[a[i]][ZZ];
242 /* Complete the partially filled last cell with zeros */
243 for (; i < na_round; i++)
252 j = atom_to_x_index<c_packX4>(a0);
253 c = a0 & (c_packX4-1);
254 for (i = 0; i < na; i++)
256 xnb[j+XX*c_packX4] = x[a[i]][XX];
257 xnb[j+YY*c_packX4] = x[a[i]][YY];
258 xnb[j+ZZ*c_packX4] = x[a[i]][ZZ];
263 j += (DIM-1)*c_packX4;
267 /* Complete the partially filled last cell with zeros */
268 for (; i < na_round; i++)
270 xnb[j+XX*c_packX4] = farAway;
271 xnb[j+YY*c_packX4] = farAway;
272 xnb[j+ZZ*c_packX4] = farAway;
277 j += (DIM-1)*c_packX4;
283 j = atom_to_x_index<c_packX8>(a0);
284 c = a0 & (c_packX8 - 1);
285 for (i = 0; i < na; i++)
287 xnb[j+XX*c_packX8] = x[a[i]][XX];
288 xnb[j+YY*c_packX8] = x[a[i]][YY];
289 xnb[j+ZZ*c_packX8] = x[a[i]][ZZ];
294 j += (DIM-1)*c_packX8;
298 /* Complete the partially filled last cell with zeros */
299 for (; i < na_round; i++)
301 xnb[j+XX*c_packX8] = farAway;
302 xnb[j+YY*c_packX8] = farAway;
303 xnb[j+ZZ*c_packX8] = farAway;
308 j += (DIM-1)*c_packX8;
314 gmx_incons("Unsupported nbnxn_atomdata_t format");
318 /* Stores the LJ parameter data in a format convenient for different kernels */
319 static void set_lj_parameter_data(nbnxn_atomdata_t *nbat, gmx_bool bSIMD)
323 int nt = nbat->ntype;
328 /* nbfp_aligned stores two parameters using the stride most suitable
329 * for the present SIMD architecture, as specified by the constant
330 * c_simdBestPairAlignment from the SIMD header.
331 * There's a slight inefficiency in allocating and initializing nbfp_aligned
332 * when it might not be used, but introducing the conditional code is not
335 nbat->alloc((void **)&nbat->nbfp_aligned,
336 nt*nt*c_simdBestPairAlignment*sizeof(*nbat->nbfp_aligned));
337 for (int i = 0; i < nt; i++)
339 for (int j = 0; j < nt; j++)
341 nbat->nbfp_aligned[(i*nt+j)*c_simdBestPairAlignment+0] = nbat->nbfp[(i*nt+j)*2+0];
342 nbat->nbfp_aligned[(i*nt+j)*c_simdBestPairAlignment+1] = nbat->nbfp[(i*nt+j)*2+1];
343 nbat->nbfp_aligned[(i*nt+j)*c_simdBestPairAlignment+2] = 0;
344 nbat->nbfp_aligned[(i*nt+j)*c_simdBestPairAlignment+3] = 0;
350 /* We use combination rule data for SIMD combination rule kernels
351 * and with LJ-PME kernels. We then only need parameters per atom type,
352 * not per pair of atom types.
354 switch (nbat->comb_rule)
357 nbat->comb_rule = ljcrGEOM;
359 for (int i = 0; i < nt; i++)
361 /* Store the sqrt of the diagonal from the nbfp matrix */
362 nbat->nbfp_comb[i*2 ] = std::sqrt(nbat->nbfp[(i*nt+i)*2 ]);
363 nbat->nbfp_comb[i*2+1] = std::sqrt(nbat->nbfp[(i*nt+i)*2+1]);
367 for (int i = 0; i < nt; i++)
369 /* Get 6*C6 and 12*C12 from the diagonal of the nbfp matrix */
370 c6 = nbat->nbfp[(i*nt+i)*2 ];
371 c12 = nbat->nbfp[(i*nt+i)*2+1];
372 if (c6 > 0 && c12 > 0)
374 /* We store 0.5*2^1/6*sigma and sqrt(4*3*eps),
375 * so we get 6*C6 and 12*C12 after combining.
377 nbat->nbfp_comb[i*2 ] = 0.5*gmx::sixthroot(c12/c6);
378 nbat->nbfp_comb[i*2+1] = std::sqrt(c6*c6/c12);
382 nbat->nbfp_comb[i*2 ] = 0;
383 nbat->nbfp_comb[i*2+1] = 0;
388 /* We always store the full matrix (see code above) */
391 gmx_incons("Unknown combination rule");
398 nbnxn_atomdata_init_simple_exclusion_masks(nbnxn_atomdata_t *nbat)
400 const int simd_width = GMX_SIMD_REAL_WIDTH;
402 /* Set the diagonal cluster pair exclusion mask setup data.
403 * In the kernel we check 0 < j - i to generate the masks.
404 * Here we store j - i for generating the mask for the first i,
405 * we substract 0.5 to avoid rounding issues.
406 * In the kernel we can subtract 1 to generate the subsequent mask.
408 int simd_4xn_diag_size;
410 simd_4xn_diag_size = std::max(NBNXN_CPU_CLUSTER_I_SIZE, simd_width);
411 snew_aligned(nbat->simd_4xn_diagonal_j_minus_i, simd_4xn_diag_size, NBNXN_MEM_ALIGN);
412 for (int j = 0; j < simd_4xn_diag_size; j++)
414 nbat->simd_4xn_diagonal_j_minus_i[j] = j - 0.5;
417 snew_aligned(nbat->simd_2xnn_diagonal_j_minus_i, simd_width, NBNXN_MEM_ALIGN);
418 for (int j = 0; j < simd_width/2; j++)
420 /* The j-cluster size is half the SIMD width */
421 nbat->simd_2xnn_diagonal_j_minus_i[j] = j - 0.5;
422 /* The next half of the SIMD width is for i + 1 */
423 nbat->simd_2xnn_diagonal_j_minus_i[simd_width/2+j] = j - 1 - 0.5;
426 /* We use up to 32 bits for exclusion masking.
427 * The same masks are used for the 4xN and 2x(N+N) kernels.
428 * The masks are read either into integer SIMD registers or into
429 * real SIMD registers (together with a cast).
430 * In single precision this means the real and integer SIMD registers
433 simd_excl_size = NBNXN_CPU_CLUSTER_I_SIZE*simd_width;
434 #if GMX_DOUBLE && !GMX_SIMD_HAVE_INT32_LOGICAL
435 snew_aligned(nbat->simd_exclusion_filter64, simd_excl_size, NBNXN_MEM_ALIGN);
437 snew_aligned(nbat->simd_exclusion_filter, simd_excl_size, NBNXN_MEM_ALIGN);
440 for (int j = 0; j < simd_excl_size; j++)
442 /* Set the consecutive bits for masking pair exclusions */
443 #if GMX_DOUBLE && !GMX_SIMD_HAVE_INT32_LOGICAL
444 nbat->simd_exclusion_filter64[j] = (1U << j);
446 nbat->simd_exclusion_filter[j] = (1U << j);
450 #if !GMX_SIMD_HAVE_LOGICAL && !GMX_SIMD_HAVE_INT32_LOGICAL
451 // If the SIMD implementation has no bitwise logical operation support
452 // whatsoever we cannot use the normal masking. Instead,
453 // we generate a vector of all 2^4 possible ways an i atom
454 // interacts with its 4 j atoms. Each array entry contains
455 // GMX_SIMD_REAL_WIDTH values that are read with a single aligned SIMD load.
456 // Since there is no logical value representation in this case, we use
457 // any nonzero value to indicate 'true', while zero mean 'false'.
458 // This can then be converted to a SIMD boolean internally in the SIMD
459 // module by comparing to zero.
460 // Each array entry encodes how this i atom will interact with the 4 j atoms.
461 // Matching code exists in set_ci_top_excls() to generate indices into this array.
462 // Those indices are used in the kernels.
464 simd_excl_size = NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
465 const real simdFalse = 0.0;
466 const real simdTrue = 1.0;
467 real *simd_interaction_array;
469 snew_aligned(simd_interaction_array, simd_excl_size * GMX_SIMD_REAL_WIDTH, NBNXN_MEM_ALIGN);
470 for (int j = 0; j < simd_excl_size; j++)
472 int index = j * GMX_SIMD_REAL_WIDTH;
473 for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
475 simd_interaction_array[index + i] = (j & (1 << i)) ? simdTrue : simdFalse;
478 nbat->simd_interaction_array = simd_interaction_array;
483 /* Initializes an nbnxn_atomdata_t data structure */
484 void nbnxn_atomdata_init(const gmx::MDLogger &mdlog,
485 nbnxn_atomdata_t *nbat,
487 int enbnxninitcombrule,
488 int ntype, const real *nbfp,
491 nbnxn_alloc_t *alloc,
497 gmx_bool simple, bCombGeom, bCombLB, bSIMD;
499 if (alloc == nullptr)
501 nbat->alloc = nbnxn_alloc_aligned;
509 nbat->free = nbnxn_free_aligned;
518 fprintf(debug, "There are %d atom types in the system, adding one for nbnxn_atomdata_t\n", ntype);
520 nbat->ntype = ntype + 1;
521 nbat->alloc((void **)&nbat->nbfp,
522 nbat->ntype*nbat->ntype*2*sizeof(*nbat->nbfp));
523 nbat->alloc((void **)&nbat->nbfp_comb, nbat->ntype*2*sizeof(*nbat->nbfp_comb));
525 /* A tolerance of 1e-5 seems reasonable for (possibly hand-typed)
526 * force-field floating point parameters.
529 ptr = getenv("GMX_LJCOMB_TOL");
534 sscanf(ptr, "%lf", &dbl);
540 /* Temporarily fill nbat->nbfp_comb with sigma and epsilon
541 * to check for the LB rule.
543 for (int i = 0; i < ntype; i++)
545 c6 = nbfp[(i*ntype+i)*2 ]/6.0;
546 c12 = nbfp[(i*ntype+i)*2+1]/12.0;
547 if (c6 > 0 && c12 > 0)
549 nbat->nbfp_comb[i*2 ] = gmx::sixthroot(c12/c6);
550 nbat->nbfp_comb[i*2+1] = 0.25*c6*c6/c12;
552 else if (c6 == 0 && c12 == 0)
554 nbat->nbfp_comb[i*2 ] = 0;
555 nbat->nbfp_comb[i*2+1] = 0;
559 /* Can not use LB rule with only dispersion or repulsion */
564 for (int i = 0; i < nbat->ntype; i++)
566 for (int j = 0; j < nbat->ntype; j++)
568 if (i < ntype && j < ntype)
570 /* fr->nbfp has been updated, so that array too now stores c6/c12 including
571 * the 6.0/12.0 prefactors to save 2 flops in the most common case (force-only).
573 c6 = nbfp[(i*ntype+j)*2 ];
574 c12 = nbfp[(i*ntype+j)*2+1];
575 nbat->nbfp[(i*nbat->ntype+j)*2 ] = c6;
576 nbat->nbfp[(i*nbat->ntype+j)*2+1] = c12;
578 /* Compare 6*C6 and 12*C12 for geometric cobination rule */
579 bCombGeom = bCombGeom &&
580 gmx_within_tol(c6*c6, nbfp[(i*ntype+i)*2 ]*nbfp[(j*ntype+j)*2 ], tol) &&
581 gmx_within_tol(c12*c12, nbfp[(i*ntype+i)*2+1]*nbfp[(j*ntype+j)*2+1], tol);
583 /* Compare C6 and C12 for Lorentz-Berthelot combination rule */
587 ((c6 == 0 && c12 == 0 &&
588 (nbat->nbfp_comb[i*2+1] == 0 || nbat->nbfp_comb[j*2+1] == 0)) ||
589 (c6 > 0 && c12 > 0 &&
590 gmx_within_tol(gmx::sixthroot(c12/c6),
591 0.5*(nbat->nbfp_comb[i*2]+nbat->nbfp_comb[j*2]), tol) &&
592 gmx_within_tol(0.25*c6*c6/c12, std::sqrt(nbat->nbfp_comb[i*2+1]*nbat->nbfp_comb[j*2+1]), tol)));
596 /* Add zero parameters for the additional dummy atom type */
597 nbat->nbfp[(i*nbat->ntype+j)*2 ] = 0;
598 nbat->nbfp[(i*nbat->ntype+j)*2+1] = 0;
604 fprintf(debug, "Combination rules: geometric %d Lorentz-Berthelot %d\n",
608 simple = nbnxn_kernel_pairlist_simple(nb_kernel_type);
610 switch (enbnxninitcombrule)
612 case enbnxninitcombruleDETECT:
613 /* We prefer the geometic combination rule,
614 * as that gives a slightly faster kernel than the LB rule.
618 nbat->comb_rule = ljcrGEOM;
622 nbat->comb_rule = ljcrLB;
626 nbat->comb_rule = ljcrNONE;
628 nbat->free(nbat->nbfp_comb);
633 if (nbat->comb_rule == ljcrNONE)
635 mesg = "Using full Lennard-Jones parameter combination matrix";
639 mesg = gmx::formatString("Using %s Lennard-Jones combination rule",
640 nbat->comb_rule == ljcrGEOM ? "geometric" : "Lorentz-Berthelot");
642 GMX_LOG(mdlog.info).asParagraph().appendText(mesg);
645 case enbnxninitcombruleGEOM:
646 nbat->comb_rule = ljcrGEOM;
648 case enbnxninitcombruleLB:
649 nbat->comb_rule = ljcrLB;
651 case enbnxninitcombruleNONE:
652 nbat->comb_rule = ljcrNONE;
654 nbat->free(nbat->nbfp_comb);
657 gmx_incons("Unknown enbnxninitcombrule");
660 bSIMD = (nb_kernel_type == nbnxnk4xN_SIMD_4xN ||
661 nb_kernel_type == nbnxnk4xN_SIMD_2xNN);
663 set_lj_parameter_data(nbat, bSIMD);
666 nbat->type = nullptr;
667 nbat->lj_comb = nullptr;
674 pack_x = std::max(NBNXN_CPU_CLUSTER_I_SIZE,
675 nbnxn_kernel_to_cluster_j_size(nb_kernel_type));
679 nbat->XFormat = nbatX4;
682 nbat->XFormat = nbatX8;
685 gmx_incons("Unsupported packing width");
690 nbat->XFormat = nbatXYZ;
693 nbat->FFormat = nbat->XFormat;
697 nbat->XFormat = nbatXYZQ;
698 nbat->FFormat = nbatXYZ;
701 nbat->nenergrp = n_energygroups;
704 /* Energy groups not supported yet for super-sub lists */
705 if (n_energygroups > 1)
707 GMX_LOG(mdlog.warning).asParagraph().appendText("NOTE: With GPUs, reporting energy group contributions is not supported");
711 /* Temporary storage goes as #grp^3*simd_width^2/2, so limit to 64 */
712 if (nbat->nenergrp > 64)
714 gmx_fatal(FARGS, "With NxN kernels not more than 64 energy groups are supported\n");
717 while (nbat->nenergrp > (1<<nbat->neg_2log))
721 nbat->energrp = nullptr;
722 nbat->alloc((void **)&nbat->shift_vec, SHIFTS*sizeof(*nbat->shift_vec));
723 nbat->xstride = (nbat->XFormat == nbatXYZQ ? STRIDE_XYZQ : DIM);
724 nbat->fstride = (nbat->FFormat == nbatXYZQ ? STRIDE_XYZQ : DIM);
730 nbnxn_atomdata_init_simple_exclusion_masks(nbat);
734 /* Initialize the output data structures */
736 snew(nbat->out, nbat->nout);
738 for (int i = 0; i < nbat->nout; i++)
740 nbnxn_atomdata_output_init(&nbat->out[i],
742 nbat->nenergrp, 1<<nbat->neg_2log,
745 nbat->buffer_flags.flag = nullptr;
746 nbat->buffer_flags.flag_nalloc = 0;
748 nth = gmx_omp_nthreads_get(emntNonbonded);
750 ptr = getenv("GMX_USE_TREEREDUCE");
753 nbat->bUseTreeReduce = strtol(ptr, nullptr, 10);
756 else if (nth > 8) /*on the CPU we currently don't benefit even at 32*/
758 nbat->bUseTreeReduce = 1;
763 nbat->bUseTreeReduce = 0;
765 if (nbat->bUseTreeReduce)
767 GMX_LOG(mdlog.info).asParagraph().appendText("Using tree force reduction");
769 snew(nbat->syncStep, nth);
773 template<int packSize>
774 static void copy_lj_to_nbat_lj_comb(const real *ljparam_type,
775 const int *type, int na,
778 /* The LJ params follow the combination rule:
779 * copy the params for the type array to the atom array.
781 for (int is = 0; is < na; is += packSize)
783 for (int k = 0; k < packSize; k++)
786 ljparam_at[is*2 + k] = ljparam_type[type[i]*2 ];
787 ljparam_at[is*2 + packSize + k] = ljparam_type[type[i]*2 + 1];
792 /* Sets the atom type in nbnxn_atomdata_t */
793 static void nbnxn_atomdata_set_atomtypes(nbnxn_atomdata_t *nbat,
794 const nbnxn_search_t nbs,
797 for (int g = 0; g < nbs->ngrid; g++)
799 const nbnxn_grid_t * grid = &nbs->grid[g];
801 /* Loop over all columns and copy and fill */
802 for (int i = 0; i < grid->ncx*grid->ncy; i++)
804 int ncz = grid->cxy_ind[i+1] - grid->cxy_ind[i];
805 int ash = (grid->cell0 + grid->cxy_ind[i])*grid->na_sc;
807 copy_int_to_nbat_int(nbs->a+ash, grid->cxy_na[i], ncz*grid->na_sc,
808 type, nbat->ntype-1, nbat->type+ash);
813 /* Sets the LJ combination rule parameters in nbnxn_atomdata_t */
814 static void nbnxn_atomdata_set_ljcombparams(nbnxn_atomdata_t *nbat,
815 const nbnxn_search_t nbs)
817 if (nbat->comb_rule != ljcrNONE)
819 for (int g = 0; g < nbs->ngrid; g++)
821 const nbnxn_grid_t * grid = &nbs->grid[g];
823 /* Loop over all columns and copy and fill */
824 for (int i = 0; i < grid->ncx*grid->ncy; i++)
826 int ncz = grid->cxy_ind[i+1] - grid->cxy_ind[i];
827 int ash = (grid->cell0 + grid->cxy_ind[i])*grid->na_sc;
829 if (nbat->XFormat == nbatX4)
831 copy_lj_to_nbat_lj_comb<c_packX4>(nbat->nbfp_comb,
834 nbat->lj_comb + ash*2);
836 else if (nbat->XFormat == nbatX8)
838 copy_lj_to_nbat_lj_comb<c_packX8>(nbat->nbfp_comb,
841 nbat->lj_comb + ash*2);
843 else if (nbat->XFormat == nbatXYZQ)
845 copy_lj_to_nbat_lj_comb<1>(nbat->nbfp_comb,
848 nbat->lj_comb + ash*2);
855 /* Sets the charges in nbnxn_atomdata_t *nbat */
856 static void nbnxn_atomdata_set_charges(nbnxn_atomdata_t *nbat,
857 const nbnxn_search_t nbs,
863 for (int g = 0; g < nbs->ngrid; g++)
865 const nbnxn_grid_t * grid = &nbs->grid[g];
867 /* Loop over all columns and copy and fill */
868 for (int cxy = 0; cxy < grid->ncx*grid->ncy; cxy++)
870 int ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
871 int na = grid->cxy_na[cxy];
872 int na_round = (grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy])*grid->na_sc;
874 if (nbat->XFormat == nbatXYZQ)
876 q = nbat->x + ash*STRIDE_XYZQ + ZZ + 1;
877 for (i = 0; i < na; i++)
879 *q = charge[nbs->a[ash+i]];
882 /* Complete the partially filled last cell with zeros */
883 for (; i < na_round; i++)
892 for (i = 0; i < na; i++)
894 *q = charge[nbs->a[ash+i]];
897 /* Complete the partially filled last cell with zeros */
898 for (; i < na_round; i++)
908 /* Set the charges of perturbed atoms in nbnxn_atomdata_t to 0.
909 * This is to automatically remove the RF/PME self term in the nbnxn kernels.
910 * Part of the zero interactions are still calculated in the normal kernels.
911 * All perturbed interactions are calculated in the free energy kernel,
912 * using the original charge and LJ data, not nbnxn_atomdata_t.
914 static void nbnxn_atomdata_mask_fep(nbnxn_atomdata_t *nbat,
915 const nbnxn_search_t nbs)
920 if (nbat->XFormat == nbatXYZQ)
922 q = nbat->x + ZZ + 1;
923 stride_q = STRIDE_XYZQ;
931 for (int g = 0; g < nbs->ngrid; g++)
933 const nbnxn_grid_t * grid = &nbs->grid[g];
940 nsubc = c_gpuNumClusterPerCell;
943 int c_offset = grid->cell0*grid->na_sc;
945 /* Loop over all columns and copy and fill */
946 for (int c = 0; c < grid->nc*nsubc; c++)
948 /* Does this cluster contain perturbed particles? */
949 if (grid->fep[c] != 0)
951 for (int i = 0; i < grid->na_c; i++)
953 /* Is this a perturbed particle? */
954 if (grid->fep[c] & (1 << i))
956 int ind = c_offset + c*grid->na_c + i;
957 /* Set atom type and charge to non-interacting */
958 nbat->type[ind] = nbat->ntype - 1;
967 /* Copies the energy group indices to a reordered and packed array */
968 static void copy_egp_to_nbat_egps(const int *a, int na, int na_round,
969 int na_c, int bit_shift,
970 const int *in, int *innb)
976 for (i = 0; i < na; i += na_c)
978 /* Store na_c energy group numbers into one int */
980 for (int sa = 0; sa < na_c; sa++)
985 comb |= (GET_CGINFO_GID(in[at]) << (sa*bit_shift));
990 /* Complete the partially filled last cell with fill */
991 for (; i < na_round; i += na_c)
997 /* Set the energy group indices for atoms in nbnxn_atomdata_t */
998 static void nbnxn_atomdata_set_energygroups(nbnxn_atomdata_t *nbat,
999 const nbnxn_search_t nbs,
1002 if (nbat->nenergrp == 1)
1007 for (int g = 0; g < nbs->ngrid; g++)
1009 const nbnxn_grid_t * grid = &nbs->grid[g];
1011 /* Loop over all columns and copy and fill */
1012 for (int i = 0; i < grid->ncx*grid->ncy; i++)
1014 int ncz = grid->cxy_ind[i+1] - grid->cxy_ind[i];
1015 int ash = (grid->cell0 + grid->cxy_ind[i])*grid->na_sc;
1017 copy_egp_to_nbat_egps(nbs->a+ash, grid->cxy_na[i], ncz*grid->na_sc,
1018 nbat->na_c, nbat->neg_2log,
1019 atinfo, nbat->energrp+(ash>>grid->na_c_2log));
1024 /* Sets all required atom parameter data in nbnxn_atomdata_t */
1025 void nbnxn_atomdata_set(nbnxn_atomdata_t *nbat,
1026 const nbnxn_search_t nbs,
1027 const t_mdatoms *mdatoms,
1030 nbnxn_atomdata_set_atomtypes(nbat, nbs, mdatoms->typeA);
1032 nbnxn_atomdata_set_charges(nbat, nbs, mdatoms->chargeA);
1036 nbnxn_atomdata_mask_fep(nbat, nbs);
1039 /* This must be done after masking types for FEP */
1040 nbnxn_atomdata_set_ljcombparams(nbat, nbs);
1042 nbnxn_atomdata_set_energygroups(nbat, nbs, atinfo);
1045 /* Copies the shift vector array to nbnxn_atomdata_t */
1046 void nbnxn_atomdata_copy_shiftvec(gmx_bool bDynamicBox,
1048 nbnxn_atomdata_t *nbat)
1052 nbat->bDynamicBox = bDynamicBox;
1053 for (i = 0; i < SHIFTS; i++)
1055 copy_rvec(shift_vec[i], nbat->shift_vec[i]);
1059 /* Copies (and reorders) the coordinates to nbnxn_atomdata_t */
1060 void nbnxn_atomdata_copy_x_to_nbat_x(const nbnxn_search_t nbs,
1064 nbnxn_atomdata_t *nbat)
1087 nbat->natoms_local = nbs->grid[0].nc*nbs->grid[0].na_sc;
1090 nth = gmx_omp_nthreads_get(emntPairsearch);
1092 #pragma omp parallel for num_threads(nth) schedule(static)
1093 for (th = 0; th < nth; th++)
1097 for (int g = g0; g < g1; g++)
1099 const nbnxn_grid_t *grid;
1102 grid = &nbs->grid[g];
1104 cxy0 = (grid->ncx*grid->ncy* th +nth-1)/nth;
1105 cxy1 = (grid->ncx*grid->ncy*(th+1)+nth-1)/nth;
1107 for (int cxy = cxy0; cxy < cxy1; cxy++)
1109 int na, ash, na_fill;
1111 na = grid->cxy_na[cxy];
1112 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1114 if (g == 0 && FillLocal)
1117 (grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy])*grid->na_sc;
1121 /* We fill only the real particle locations.
1122 * We assume the filling entries at the end have been
1123 * properly set before during pair-list generation.
1127 copy_rvec_to_nbat_real(nbs->a+ash, na, na_fill, x,
1128 nbat->XFormat, nbat->x, ash);
1132 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1137 nbnxn_atomdata_clear_reals(real * gmx_restrict dest,
1140 for (int i = i0; i < i1; i++)
1146 gmx_unused static void
1147 nbnxn_atomdata_reduce_reals(real * gmx_restrict dest,
1149 real ** gmx_restrict src,
1155 /* The destination buffer contains data, add to it */
1156 for (int i = i0; i < i1; i++)
1158 for (int s = 0; s < nsrc; s++)
1160 dest[i] += src[s][i];
1166 /* The destination buffer is unitialized, set it first */
1167 for (int i = i0; i < i1; i++)
1169 dest[i] = src[0][i];
1170 for (int s = 1; s < nsrc; s++)
1172 dest[i] += src[s][i];
1178 gmx_unused static void
1179 nbnxn_atomdata_reduce_reals_simd(real gmx_unused * gmx_restrict dest,
1180 gmx_bool gmx_unused bDestSet,
1181 real gmx_unused ** gmx_restrict src,
1182 int gmx_unused nsrc,
1183 int gmx_unused i0, int gmx_unused i1)
1186 /* The SIMD width here is actually independent of that in the kernels,
1187 * but we use the same width for simplicity (usually optimal anyhow).
1189 SimdReal dest_SSE, src_SSE;
1193 for (int i = i0; i < i1; i += GMX_SIMD_REAL_WIDTH)
1195 dest_SSE = load<SimdReal>(dest+i);
1196 for (int s = 0; s < nsrc; s++)
1198 src_SSE = load<SimdReal>(src[s]+i);
1199 dest_SSE = dest_SSE + src_SSE;
1201 store(dest+i, dest_SSE);
1206 for (int i = i0; i < i1; i += GMX_SIMD_REAL_WIDTH)
1208 dest_SSE = load<SimdReal>(src[0]+i);
1209 for (int s = 1; s < nsrc; s++)
1211 src_SSE = load<SimdReal>(src[s]+i);
1212 dest_SSE = dest_SSE + src_SSE;
1214 store(dest+i, dest_SSE);
1220 /* Add part of the force array(s) from nbnxn_atomdata_t to f */
1222 nbnxn_atomdata_add_nbat_f_to_f_part(const nbnxn_search_t nbs,
1223 const nbnxn_atomdata_t *nbat,
1224 nbnxn_atomdata_output_t *out,
1234 /* Loop over all columns and copy and fill */
1235 switch (nbat->FFormat)
1243 for (int a = a0; a < a1; a++)
1245 int i = cell[a]*nbat->fstride;
1248 f[a][YY] += fnb[i+1];
1249 f[a][ZZ] += fnb[i+2];
1254 for (int a = a0; a < a1; a++)
1256 int i = cell[a]*nbat->fstride;
1258 for (int fa = 0; fa < nfa; fa++)
1260 f[a][XX] += out[fa].f[i];
1261 f[a][YY] += out[fa].f[i+1];
1262 f[a][ZZ] += out[fa].f[i+2];
1272 for (int a = a0; a < a1; a++)
1274 int i = atom_to_x_index<c_packX4>(cell[a]);
1276 f[a][XX] += fnb[i+XX*c_packX4];
1277 f[a][YY] += fnb[i+YY*c_packX4];
1278 f[a][ZZ] += fnb[i+ZZ*c_packX4];
1283 for (int a = a0; a < a1; a++)
1285 int i = atom_to_x_index<c_packX4>(cell[a]);
1287 for (int fa = 0; fa < nfa; fa++)
1289 f[a][XX] += out[fa].f[i+XX*c_packX4];
1290 f[a][YY] += out[fa].f[i+YY*c_packX4];
1291 f[a][ZZ] += out[fa].f[i+ZZ*c_packX4];
1301 for (int a = a0; a < a1; a++)
1303 int i = atom_to_x_index<c_packX8>(cell[a]);
1305 f[a][XX] += fnb[i+XX*c_packX8];
1306 f[a][YY] += fnb[i+YY*c_packX8];
1307 f[a][ZZ] += fnb[i+ZZ*c_packX8];
1312 for (int a = a0; a < a1; a++)
1314 int i = atom_to_x_index<c_packX8>(cell[a]);
1316 for (int fa = 0; fa < nfa; fa++)
1318 f[a][XX] += out[fa].f[i+XX*c_packX8];
1319 f[a][YY] += out[fa].f[i+YY*c_packX8];
1320 f[a][ZZ] += out[fa].f[i+ZZ*c_packX8];
1326 gmx_incons("Unsupported nbnxn_atomdata_t format");
1330 static gmx_inline unsigned char reverse_bits(unsigned char b)
1332 /* http://graphics.stanford.edu/~seander/bithacks.html#ReverseByteWith64BitsDiv */
1333 return (b * 0x0202020202ULL & 0x010884422010ULL) % 1023;
1336 static void nbnxn_atomdata_add_nbat_f_to_f_treereduce(const nbnxn_atomdata_t *nbat,
1339 const nbnxn_buffer_flags_t *flags = &nbat->buffer_flags;
1341 int next_pow2 = 1<<(gmx::log2I(nth-1)+1);
1343 assert(nbat->nout == nth); /* tree-reduce currently only works for nout==nth */
1345 memset(nbat->syncStep, 0, sizeof(*(nbat->syncStep))*nth);
1347 #pragma omp parallel num_threads(nth)
1355 th = gmx_omp_get_thread_num();
1357 for (group_size = 2; group_size < 2*next_pow2; group_size *= 2)
1359 int index[2], group_pos, partner_pos, wu;
1360 int partner_th = th ^ (group_size/2);
1365 /* wait on partner thread - replaces full barrier */
1366 int sync_th, sync_group_size;
1368 tMPI_Atomic_memory_barrier(); /* gurantee data is saved before marking work as done */
1369 tMPI_Atomic_set(&(nbat->syncStep[th]), group_size/2); /* mark previous step as completed */
1371 /* find thread to sync with. Equal to partner_th unless nth is not a power of two. */
1372 for (sync_th = partner_th, sync_group_size = group_size; sync_th >= nth && sync_group_size > 2; sync_group_size /= 2)
1374 sync_th &= ~(sync_group_size/4);
1376 if (sync_th < nth) /* otherwise nothing to sync index[1] will be >=nout */
1378 /* wait on the thread which computed input data in previous step */
1379 while (tMPI_Atomic_get((volatile tMPI_Atomic_t*)&(nbat->syncStep[sync_th])) < group_size/2)
1383 /* guarantee that no later load happens before wait loop is finisehd */
1384 tMPI_Atomic_memory_barrier();
1386 #else /* TMPI_ATOMICS */
1391 /* Calculate buffers to sum (result goes into first buffer) */
1392 group_pos = th % group_size;
1393 index[0] = th - group_pos;
1394 index[1] = index[0] + group_size/2;
1396 /* If no second buffer, nothing to do */
1397 if (index[1] >= nbat->nout && group_size > 2)
1402 #if NBNXN_BUFFERFLAG_MAX_THREADS > 256
1403 #error reverse_bits assumes max 256 threads
1405 /* Position is permuted so that one of the 2 vectors being added was computed on the same thread in the previous step.
1406 This improves locality and enables to sync with just a single thread between steps (=the levels in the btree).
1407 The permutation which allows this corresponds to reversing the bits of the group position.
1409 group_pos = reverse_bits(group_pos)/(256/group_size);
1411 partner_pos = group_pos ^ 1;
1413 /* loop over two work-units (own and partner) */
1414 for (wu = 0; wu < 2; wu++)
1418 if (partner_th < nth)
1420 break; /* partner exists we don't have to do his work */
1424 group_pos = partner_pos;
1428 /* Calculate the cell-block range for our thread */
1429 b0 = (flags->nflag* group_pos )/group_size;
1430 b1 = (flags->nflag*(group_pos+1))/group_size;
1432 for (b = b0; b < b1; b++)
1434 i0 = b *NBNXN_BUFFERFLAG_SIZE*nbat->fstride;
1435 i1 = (b+1)*NBNXN_BUFFERFLAG_SIZE*nbat->fstride;
1437 if (bitmask_is_set(flags->flag[b], index[1]) || group_size > 2)
1440 nbnxn_atomdata_reduce_reals_simd
1442 nbnxn_atomdata_reduce_reals
1444 (nbat->out[index[0]].f,
1445 bitmask_is_set(flags->flag[b], index[0]) || group_size > 2,
1446 &(nbat->out[index[1]].f), 1, i0, i1);
1449 else if (!bitmask_is_set(flags->flag[b], index[0]))
1451 nbnxn_atomdata_clear_reals(nbat->out[index[0]].f,
1458 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1463 static void nbnxn_atomdata_add_nbat_f_to_f_stdreduce(const nbnxn_atomdata_t *nbat,
1466 #pragma omp parallel for num_threads(nth) schedule(static)
1467 for (int th = 0; th < nth; th++)
1471 const nbnxn_buffer_flags_t *flags;
1473 real *fptr[NBNXN_BUFFERFLAG_MAX_THREADS];
1475 flags = &nbat->buffer_flags;
1477 /* Calculate the cell-block range for our thread */
1478 int b0 = (flags->nflag* th )/nth;
1479 int b1 = (flags->nflag*(th+1))/nth;
1481 for (int b = b0; b < b1; b++)
1483 int i0 = b *NBNXN_BUFFERFLAG_SIZE*nbat->fstride;
1484 int i1 = (b+1)*NBNXN_BUFFERFLAG_SIZE*nbat->fstride;
1487 for (int out = 1; out < nbat->nout; out++)
1489 if (bitmask_is_set(flags->flag[b], out))
1491 fptr[nfptr++] = nbat->out[out].f;
1497 nbnxn_atomdata_reduce_reals_simd
1499 nbnxn_atomdata_reduce_reals
1502 bitmask_is_set(flags->flag[b], 0),
1506 else if (!bitmask_is_set(flags->flag[b], 0))
1508 nbnxn_atomdata_clear_reals(nbat->out[0].f,
1513 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1517 /* Add the force array(s) from nbnxn_atomdata_t to f */
1518 void nbnxn_atomdata_add_nbat_f_to_f(const nbnxn_search_t nbs,
1520 const nbnxn_atomdata_t *nbat,
1525 nbs_cycle_start(&nbs->cc[enbsCCreducef]);
1531 na = nbs->natoms_nonlocal;
1535 na = nbs->natoms_local;
1538 a0 = nbs->natoms_local;
1539 na = nbs->natoms_nonlocal - nbs->natoms_local;
1543 int nth = gmx_omp_nthreads_get(emntNonbonded);
1547 if (locality != eatAll)
1549 gmx_incons("add_f_to_f called with nout>1 and locality!=eatAll");
1552 /* Reduce the force thread output buffers into buffer 0, before adding
1553 * them to the, differently ordered, "real" force buffer.
1555 if (nbat->bUseTreeReduce)
1557 nbnxn_atomdata_add_nbat_f_to_f_treereduce(nbat, nth);
1561 nbnxn_atomdata_add_nbat_f_to_f_stdreduce(nbat, nth);
1564 #pragma omp parallel for num_threads(nth) schedule(static)
1565 for (int th = 0; th < nth; th++)
1569 nbnxn_atomdata_add_nbat_f_to_f_part(nbs, nbat,
1576 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1579 nbs_cycle_stop(&nbs->cc[enbsCCreducef]);
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);