2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2017,2018,2019, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
38 * \brief Implements functions for tuning adjustable parameters for the nbnxn non-bonded search and interaction kernels
40 * \author Berk Hess <hess@kth.se>
41 * \ingroup __module_nb_verlet
46 #include "pairlist_tuning.h"
55 #include "gromacs/domdec/domdec.h"
56 #include "gromacs/hardware/cpuinfo.h"
57 #include "gromacs/math/vec.h"
58 #include "gromacs/mdlib/calc_verletbuf.h"
59 #include "gromacs/mdtypes/commrec.h"
60 #include "gromacs/mdtypes/inputrec.h"
61 #include "gromacs/mdtypes/interaction_const.h"
62 #include "gromacs/mdtypes/state.h"
63 #include "gromacs/nbnxm/nbnxm.h"
64 #include "gromacs/pbcutil/pbc.h"
65 #include "gromacs/topology/topology.h"
66 #include "gromacs/utility/cstringutil.h"
67 #include "gromacs/utility/fatalerror.h"
68 #include "gromacs/utility/gmxassert.h"
69 #include "gromacs/utility/logger.h"
70 #include "gromacs/utility/strconvert.h"
71 #include "gromacs/utility/stringutil.h"
73 /*! \brief Returns if we can (heuristically) change nstlist and rlist
75 * \param [in] ir The input parameter record
77 static bool supportsDynamicPairlistGenerationInterval(const t_inputrec &ir)
80 ir.cutoff_scheme == ecutsVERLET &&
82 !(EI_MD(ir.eI) && ir.etc == etcNO) &&
86 /*! \brief Cost of non-bonded kernels
88 * We determine the extra cost of the non-bonded kernels compared to
89 * a reference nstlist value of 10 (which is the default in grompp).
91 static const int nbnxnReferenceNstlist = 10;
92 //! The values to try when switching
93 const int nstlist_try[] = { 20, 25, 40, 50, 80, 100 };
94 //! Number of elements in the neighborsearch list trials.
95 #define NNSTL (sizeof(nstlist_try)/sizeof(nstlist_try[0]))
96 /* Increase nstlist until the size of the pair-list increased by
97 * \p c_nbnxnListSizeFactor??? or more, but never more than
98 * \p c_nbnxnListSizeFactor??? + \p c_nbnxnListSizeFactorMargin.
99 * Since we have dynamic pair list pruning, the force kernel cost depends
100 * only very weakly on nstlist. It depends strongly on nstlistPrune.
101 * Increasing nstlist mainly affects the cost of the pair search (down due
102 * to lower frequency, up due to larger list) and the list pruning kernel.
103 * We increase nstlist conservatively with regard to kernel performance.
104 * In serial the search cost is not high and thus we don't gain much by
105 * increasing nstlist a lot. In parallel the MPI and CPU-GPU communication
106 * volume as well as the communication buffer preparation and reduction time
107 * increase quickly with rlist and thus nslist. Therefore we should avoid
108 * large nstlist, even if that also reduces the domain decomposition cost.
109 * With GPUs we perform the dynamic pruning in a rolling fashion and this
110 * overlaps with the update on the CPU, which allows even larger nstlist.
112 // CPU: pair-search is a factor ~1.5 slower than the non-bonded kernel.
113 //! Target pair-list size increase ratio for CPU
114 static const float c_nbnxnListSizeFactorCpu = 1.25;
115 // Intel KNL: pair-search is a factor ~2-3 slower than the non-bonded kernel.
116 //! Target pair-list size increase ratio for Intel KNL
117 static const float c_nbnxnListSizeFactorIntelXeonPhi = 1.4;
118 // GPU: pair-search is a factor 1.5-3 slower than the non-bonded kernel.
119 //! Target pair-list size increase ratio for GPU
120 static const float c_nbnxnListSizeFactorGPU = 1.4;
121 //! Never increase the size of the pair-list more than the factor above plus this margin
122 static const float c_nbnxnListSizeFactorMargin = 0.1;
124 void increaseNstlist(FILE *fp, t_commrec *cr,
125 t_inputrec *ir, int nstlist_cmdline,
126 const gmx_mtop_t *mtop,
128 bool useOrEmulateGpuForNonbondeds,
129 const gmx::CpuInfo &cpuinfo)
131 if (!EI_DYNAMICS(ir->eI))
133 /* Can only increase nstlist with dynamics */
137 float listfac_ok, listfac_max;
138 int nstlist_orig, nstlist_prev;
139 real rlistWithReferenceNstlist, rlist_inc, rlist_ok, rlist_max;
140 real rlist_new, rlist_prev;
141 size_t nstlist_ind = 0;
142 gmx_bool bBox, bDD, bCont;
143 const char *nstl_gpu = "\nFor optimal performance with a GPU nstlist (now %d) should be larger.\nThe optimum depends on your CPU and GPU resources.\nYou might want to try several nstlist values.\n";
144 const char *nve_err = "Can not increase nstlist because an NVE ensemble is used";
145 const char *vbd_err = "Can not increase nstlist because verlet-buffer-tolerance is not set or used";
146 const char *box_err = "Can not increase nstlist because the box is too small";
147 const char *dd_err = "Can not increase nstlist because of domain decomposition limitations";
150 if (nstlist_cmdline <= 0)
152 if (ir->nstlist == 1)
154 /* The user probably set nstlist=1 for a reason,
155 * don't mess with the settings.
160 /* With a GPU and fixed nstlist suggest tuning nstlist */
162 useOrEmulateGpuForNonbondeds &&
163 ir->nstlist < nstlist_try[0] &&
164 !supportsDynamicPairlistGenerationInterval(*ir))
166 fprintf(fp, nstl_gpu, ir->nstlist);
170 while (nstlist_ind < NNSTL && ir->nstlist >= nstlist_try[nstlist_ind])
174 if (nstlist_ind == NNSTL)
176 /* There are no larger nstlist value to try */
181 if (EI_MD(ir->eI) && ir->etc == etcNO)
185 fprintf(stderr, "%s\n", nve_err);
189 fprintf(fp, "%s\n", nve_err);
195 if (ir->verletbuf_tol == 0 && useOrEmulateGpuForNonbondeds)
197 gmx_fatal(FARGS, "You are using an old tpr file with a GPU, please generate a new tpr file with an up to date version of grompp");
200 if (ir->verletbuf_tol < 0)
204 fprintf(stderr, "%s\n", vbd_err);
208 fprintf(fp, "%s\n", vbd_err);
214 GMX_RELEASE_ASSERT(supportsDynamicPairlistGenerationInterval(*ir), "In all cases that do not support dynamic nstlist, we should have returned with an appropriate message above");
216 if (useOrEmulateGpuForNonbondeds)
218 listfac_ok = c_nbnxnListSizeFactorGPU;
220 else if (cpuinfo.brandString().find("Xeon Phi") != std::string::npos)
222 listfac_ok = c_nbnxnListSizeFactorIntelXeonPhi;
226 listfac_ok = c_nbnxnListSizeFactorCpu;
228 listfac_max = listfac_ok + c_nbnxnListSizeFactorMargin;
230 nstlist_orig = ir->nstlist;
231 if (nstlist_cmdline > 0)
235 sprintf(buf, "Getting nstlist=%d from command line option",
238 ir->nstlist = nstlist_cmdline;
241 ListSetupType listType = (useOrEmulateGpuForNonbondeds ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
242 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
244 /* Allow rlist to make the list a given factor larger than the list
245 * would be with the reference value for nstlist (10).
247 nstlist_prev = ir->nstlist;
248 ir->nstlist = nbnxnReferenceNstlist;
249 calc_verlet_buffer_size(mtop, det(box), ir, ir->nstlist, ir->nstlist - 1,
251 nullptr, &rlistWithReferenceNstlist);
252 ir->nstlist = nstlist_prev;
254 /* Determine the pair list size increase due to zero interactions */
255 rlist_inc = nbnxn_get_rlist_effective_inc(listSetup.cluster_size_j,
256 mtop->natoms/det(box));
257 rlist_ok = (rlistWithReferenceNstlist + rlist_inc)*std::cbrt(listfac_ok) - rlist_inc;
258 rlist_max = (rlistWithReferenceNstlist + rlist_inc)*std::cbrt(listfac_max) - rlist_inc;
261 fprintf(debug, "nstlist tuning: rlist_inc %.3f rlist_ok %.3f rlist_max %.3f\n",
262 rlist_inc, rlist_ok, rlist_max);
265 nstlist_prev = nstlist_orig;
266 rlist_prev = ir->rlist;
269 if (nstlist_cmdline <= 0)
271 ir->nstlist = nstlist_try[nstlist_ind];
274 /* Set the pair-list buffer size in ir */
275 calc_verlet_buffer_size(mtop, det(box), ir, ir->nstlist, ir->nstlist - 1, -1, &listSetup, nullptr, &rlist_new);
277 /* Does rlist fit in the box? */
278 bBox = (gmx::square(rlist_new) < max_cutoff2(ir->ePBC, box));
280 if (bBox && DOMAINDECOMP(cr))
282 /* Check if rlist fits in the domain decomposition */
283 if (inputrec2nboundeddim(ir) < DIM)
285 gmx_incons("Changing nstlist with domain decomposition and unbounded dimensions is not implemented yet");
288 copy_mat(box, state_tmp.box);
289 bDD = change_dd_cutoff(cr, state_tmp, rlist_new);
294 fprintf(debug, "nstlist %d rlist %.3f bBox %s bDD %s\n",
295 ir->nstlist, rlist_new, gmx::boolToString(bBox), gmx::boolToString(bDD));
300 if (nstlist_cmdline <= 0)
302 if (bBox && bDD && rlist_new <= rlist_max)
304 /* Increase nstlist */
305 nstlist_prev = ir->nstlist;
306 rlist_prev = rlist_new;
307 bCont = (nstlist_ind+1 < NNSTL && rlist_new < rlist_ok);
311 /* Stick with the previous nstlist */
312 ir->nstlist = nstlist_prev;
313 rlist_new = rlist_prev;
325 gmx_warning("%s", !bBox ? box_err : dd_err);
328 fprintf(fp, "\n%s\n", !bBox ? box_err : dd_err);
330 ir->nstlist = nstlist_orig;
332 else if (ir->nstlist != nstlist_orig || rlist_new != ir->rlist)
334 sprintf(buf, "Changing nstlist from %d to %d, rlist from %g to %g",
335 nstlist_orig, ir->nstlist,
336 ir->rlist, rlist_new);
339 fprintf(stderr, "%s\n\n", buf);
343 fprintf(fp, "%s\n\n", buf);
345 ir->rlist = rlist_new;
349 /*! \brief The interval in steps at which we perform dynamic, rolling pruning on a GPU.
351 * Ideally we should auto-tune this value.
352 * Not considering overheads, 1 would be the ideal value. But 2 seems
353 * a reasonable compromise that reduces GPU kernel launch overheads and
354 * also avoids inefficiency on large GPUs when pruning small lists.
355 * Because with domain decomposition we alternate local/non-local pruning
356 * at even/odd steps, which gives a period of 2, this value currenly needs
357 * to be 2, which is indirectly asserted when the GPU pruning is dispatched
358 * during the force evaluation.
360 static const int c_nbnxnGpuRollingListPruningInterval = 2;
362 /*! \brief The minimum nstlist for dynamic pair list pruning.
364 * In most cases going lower than 4 will lead to a too high pruning cost.
365 * This value should be a multiple of \p c_nbnxnGpuRollingListPruningInterval
367 static const int c_nbnxnDynamicListPruningMinLifetime = 4;
369 /*! \brief Set the dynamic pairlist pruning parameters in \p ic
371 * \param[in] ir The input parameter record
372 * \param[in] mtop The global topology
373 * \param[in] box The unit cell
374 * \param[in] useGpu Tells if we are using a GPU for non-bondeds
375 * \param[in] listSetup The nbnxn pair list setup
376 * \param[in] userSetNstlistPrune The user set ic->nstlistPrune (using an env.var.)
377 * \param[in] ic The nonbonded interactions constants
378 * \param[in,out] listParams The list setup parameters
381 setDynamicPairlistPruningParameters(const t_inputrec *ir,
382 const gmx_mtop_t *mtop,
385 const VerletbufListSetup &listSetup,
386 bool userSetNstlistPrune,
387 const interaction_const_t *ic,
388 NbnxnListParameters *listParams)
390 /* When nstlistPrune was set by the user, we need to execute one loop
391 * iteration to determine rlistInner.
392 * Otherwise we compute rlistInner and increase nstlist as long as
393 * we have a pairlist buffer of length 0 (i.e. rlistInner == cutoff).
395 const real interactionCutoff = std::max(ic->rcoulomb, ic->rvdw);
396 int tunedNstlistPrune = listParams->nstlistPrune;
399 /* Dynamic pruning on the GPU is performed on the list for
400 * the next step on the coordinates of the current step,
401 * so the list lifetime is nstlistPrune (not the usual nstlist-1).
403 int listLifetime = tunedNstlistPrune - (useGpu ? 0 : 1);
404 listParams->nstlistPrune = tunedNstlistPrune;
405 calc_verlet_buffer_size(mtop, det(box), ir,
406 tunedNstlistPrune, listLifetime,
407 -1, &listSetup, nullptr,
408 &listParams->rlistInner);
410 /* On the GPU we apply the dynamic pruning in a rolling fashion
411 * every c_nbnxnGpuRollingListPruningInterval steps,
412 * so keep nstlistPrune a multiple of the interval.
414 tunedNstlistPrune += useGpu ? c_nbnxnGpuRollingListPruningInterval : 1;
416 while (!userSetNstlistPrune &&
417 tunedNstlistPrune < ir->nstlist &&
418 listParams->rlistInner == interactionCutoff);
420 if (userSetNstlistPrune)
422 listParams->useDynamicPruning = true;
426 /* Determine the pair list size increase due to zero interactions */
427 real rlistInc = nbnxn_get_rlist_effective_inc(listSetup.cluster_size_j,
428 mtop->natoms/det(box));
430 /* Dynamic pruning is only useful when the inner list is smaller than
431 * the outer. The factor 0.99 ensures at least 3% list size reduction.
433 * With dynamic pruning on the CPU we prune after updating,
434 * so nstlistPrune=nstlist-1 would add useless extra work.
435 * With the GPU there will probably be more overhead than gain
436 * with nstlistPrune=nstlist-1, so we disable dynamic pruning.
437 * Note that in such cases the first sub-condition is likely also false.
439 listParams->useDynamicPruning =
440 (listParams->rlistInner + rlistInc < 0.99*(listParams->rlistOuter + rlistInc) &&
441 listParams->nstlistPrune < ir->nstlist - 1);
444 if (!listParams->useDynamicPruning)
446 /* These parameters should not be used, but set them to useful values */
447 listParams->nstlistPrune = -1;
448 listParams->rlistInner = listParams->rlistOuter;
452 /*! \brief Returns a string describing the setup of a single pair-list
454 * \param[in] listName Short name of the list, can be ""
455 * \param[in] nstList The list update interval in steps
456 * \param[in] nstListForSpacing Update interval for setting the number characters for printing \p nstList
457 * \param[in] rList List cut-off radius
458 * \param[in] interactionCutoff The interaction cut-off, use for printing the list buffer size
460 static std::string formatListSetup(const std::string &listName,
462 int nstListForSpacing,
464 real interactionCutoff)
466 std::string listSetup = " ";
467 if (!listName.empty())
469 listSetup += listName + " list: ";
471 listSetup += "updated every ";
472 // Make the shortest int format string that fits nstListForSpacing
473 std::string nstListFormat = "%" + gmx::formatString("%zu", gmx::formatString("%d", nstListForSpacing).size()) + "d";
474 listSetup += gmx::formatString(nstListFormat.c_str(), nstList);
475 listSetup += gmx::formatString(" steps, buffer %.3f nm, rlist %.3f nm\n",
476 rList - interactionCutoff, rList);
481 void setupDynamicPairlistPruning(const gmx::MDLogger &mdlog,
482 const t_inputrec *ir,
483 const gmx_mtop_t *mtop,
486 const interaction_const_t *ic,
487 NbnxnListParameters *listParams)
489 GMX_RELEASE_ASSERT(listParams->rlistOuter > 0, "With the nbnxn setup rlist should be > 0");
491 /* Initialize the parameters to no dynamic list pruning */
492 listParams->useDynamicPruning = false;
494 const VerletbufListSetup ls = verletbufGetListSetup(nbnxnKernelType);
496 /* Currently emulation mode does not support dual pair-lists */
497 const bool useGpu = (nbnxnKernelType == nbnxnk8x8x8_GPU);
499 if (supportsDynamicPairlistGenerationInterval(*ir) &&
500 getenv("GMX_DISABLE_DYNAMICPRUNING") == nullptr)
502 /* Note that nstlistPrune can have any value independently of nstlist.
503 * Actually applying rolling pruning is only useful when
504 * nstlistPrune < nstlist -1
506 char *env = getenv("GMX_NSTLIST_DYNAMICPRUNING");
507 bool userSetNstlistPrune = (env != nullptr);
509 if (userSetNstlistPrune)
512 listParams->nstlistPrune = strtol(env, &end, 10);
513 if (!end || (*end != 0) ||
514 !(listParams->nstlistPrune > 0 && listParams->nstlistPrune < ir->nstlist))
516 gmx_fatal(FARGS, "Invalid value passed in GMX_NSTLIST_DYNAMICPRUNING=%s, should be > 0 and < nstlist", env);
521 static_assert(c_nbnxnDynamicListPruningMinLifetime % c_nbnxnGpuRollingListPruningInterval == 0,
522 "c_nbnxnDynamicListPruningMinLifetime sets the starting value for nstlistPrune, which should be divisible by the rolling pruning interval for efficiency reasons.");
524 // TODO: Use auto-tuning to determine nstlistPrune
525 listParams->nstlistPrune = c_nbnxnDynamicListPruningMinLifetime;
528 setDynamicPairlistPruningParameters(ir, mtop, box, useGpu, ls,
529 userSetNstlistPrune, ic,
532 if (listParams->useDynamicPruning && useGpu)
534 /* Note that we can round down here. This makes the effective
535 * rolling pruning interval slightly shorter than nstlistTune,
536 * thus giving correct results, but a slightly lower efficiency.
538 GMX_RELEASE_ASSERT(listParams->nstlistPrune >= c_nbnxnGpuRollingListPruningInterval,
539 ( "With dynamic list pruning on GPUs pruning frequency must be at least as large as the rolling pruning interval (" +
540 std::to_string(c_nbnxnGpuRollingListPruningInterval) +
542 listParams->numRollingParts = listParams->nstlistPrune/c_nbnxnGpuRollingListPruningInterval;
546 listParams->numRollingParts = 1;
552 const real interactionCutoff = std::max(ic->rcoulomb, ic->rvdw);
553 if (listParams->useDynamicPruning)
555 mesg += gmx::formatString("Using a dual %dx%d pair-list setup updated with dynamic%s pruning:\n",
556 ls.cluster_size_i, ls.cluster_size_j,
557 listParams->numRollingParts > 1 ? ", rolling" : "");
558 mesg += formatListSetup("outer", ir->nstlist, ir->nstlist, listParams->rlistOuter, interactionCutoff);
559 mesg += formatListSetup("inner", listParams->nstlistPrune, ir->nstlist, listParams->rlistInner, interactionCutoff);
563 mesg += gmx::formatString("Using a %dx%d pair-list setup:\n",
564 ls.cluster_size_i, ls.cluster_size_j);
565 mesg += formatListSetup("", ir->nstlist, ir->nstlist, listParams->rlistOuter, interactionCutoff);
567 if (supportsDynamicPairlistGenerationInterval(*ir))
569 VerletbufListSetup listSetup1x1 = { 1, 1 };
572 calc_verlet_buffer_size(mtop, det(box), ir, ir->nstlist, ir->nstlist - 1,
573 -1, &listSetup1x1, nullptr, &rlistOuter);
574 if (listParams->useDynamicPruning)
576 int listLifeTime = listParams->nstlistPrune - (useGpu ? 0 : 1);
577 calc_verlet_buffer_size(mtop, det(box), ir, listParams->nstlistPrune, listLifeTime,
578 -1, &listSetup1x1, nullptr, &rlistInner);
581 mesg += gmx::formatString("At tolerance %g kJ/mol/ps per atom, equivalent classical 1x1 list would be:\n",
583 if (listParams->useDynamicPruning)
585 mesg += formatListSetup("outer", ir->nstlist, ir->nstlist, rlistOuter, interactionCutoff);
586 mesg += formatListSetup("inner", listParams->nstlistPrune, ir->nstlist, rlistInner, interactionCutoff);
590 mesg += formatListSetup("", ir->nstlist, ir->nstlist, rlistOuter, interactionCutoff);
594 GMX_LOG(mdlog.info).asParagraph().appendText(mesg);