2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2019,2020,2021, 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 * This file defines functions for setting up kernel benchmarks
40 * \author Berk Hess <hess@kth.se>
41 * \ingroup module_nbnxm
46 #include "bench_setup.h"
50 #include "gromacs/gmxlib/nrnb.h"
51 #include "gromacs/math/units.h"
52 #include "gromacs/mdlib/dispersioncorrection.h"
53 #include "gromacs/mdlib/force_flags.h"
54 #include "gromacs/mdlib/forcerec.h"
55 #include "gromacs/mdlib/gmx_omp_nthreads.h"
56 #include "gromacs/mdtypes/enerdata.h"
57 #include "gromacs/mdtypes/forcerec.h"
58 #include "gromacs/mdtypes/interaction_const.h"
59 #include "gromacs/mdtypes/mdatom.h"
60 #include "gromacs/mdtypes/simulation_workload.h"
61 #include "gromacs/nbnxm/atomdata.h"
62 #include "gromacs/nbnxm/gridset.h"
63 #include "gromacs/nbnxm/nbnxm.h"
64 #include "gromacs/nbnxm/nbnxm_simd.h"
65 #include "gromacs/nbnxm/pairlistset.h"
66 #include "gromacs/nbnxm/pairlistsets.h"
67 #include "gromacs/nbnxm/pairsearch.h"
68 #include "gromacs/pbcutil/ishift.h"
69 #include "gromacs/pbcutil/pbc.h"
70 #include "gromacs/simd/simd.h"
71 #include "gromacs/timing/cyclecounter.h"
72 #include "gromacs/utility/enumerationhelpers.h"
73 #include "gromacs/utility/fatalerror.h"
74 #include "gromacs/utility/logger.h"
76 #include "bench_system.h"
81 /*! \brief Checks the kernel setup
83 * Returns an error string when the kernel is not available.
85 static std::optional<std::string> checkKernelSetup(const KernelBenchOptions& options)
87 GMX_RELEASE_ASSERT(options.nbnxmSimd < BenchMarkKernels::Count
88 && options.nbnxmSimd != BenchMarkKernels::SimdAuto,
89 "Need a valid kernel SIMD type");
92 if ((options.nbnxmSimd != BenchMarkKernels::SimdNo && !GMX_SIMD)
93 #ifndef GMX_NBNXN_SIMD_4XN
94 || options.nbnxmSimd == BenchMarkKernels::Simd4XM
96 #ifndef GMX_NBNXN_SIMD_2XNN
97 || options.nbnxmSimd == BenchMarkKernels::Simd2XMM
101 return "the requested SIMD kernel was not set up at configuration time";
104 if (options.reportTime && (0 > gmx_cycles_calibrate(1.0)))
106 return "the -time option is not supported on this system";
112 //! Helper to translate between the different enumeration values.
113 static KernelType translateBenchmarkEnum(const BenchMarkKernels& kernel)
115 int kernelInt = static_cast<int>(kernel);
116 return static_cast<KernelType>(kernelInt);
119 /*! \brief Returns the kernel setup
121 static KernelSetup getKernelSetup(const KernelBenchOptions& options)
123 auto messageWhenInvalid = checkKernelSetup(options);
124 GMX_RELEASE_ASSERT(!messageWhenInvalid, "Need valid options");
126 KernelSetup kernelSetup;
128 // The int enum options.nbnxnSimd is set up to match KernelType + 1
129 kernelSetup.kernelType = translateBenchmarkEnum(options.nbnxmSimd);
130 // The plain-C kernel does not support analytical ewald correction
131 if (kernelSetup.kernelType == KernelType::Cpu4x4_PlainC)
133 kernelSetup.ewaldExclusionType = EwaldExclusionType::Table;
137 kernelSetup.ewaldExclusionType = options.useTabulatedEwaldCorr ? EwaldExclusionType::Table
138 : EwaldExclusionType::Analytical;
144 //! Return an interaction constants struct with members used in the benchmark set appropriately
145 static interaction_const_t setupInteractionConst(const KernelBenchOptions& options)
148 interaction_const_t ic;
150 ic.vdwtype = VanDerWaalsType::Cut;
151 ic.vdw_modifier = InteractionModifiers::PotShift;
152 ic.rvdw = options.pairlistCutoff;
154 ic.eeltype = (options.coulombType == BenchMarkCoulomb::Pme ? CoulombInteractionType::Pme
155 : CoulombInteractionType::RF);
156 ic.coulomb_modifier = InteractionModifiers::PotShift;
157 ic.rcoulomb = options.pairlistCutoff;
159 // Reaction-field with reactionFieldPermitivity=inf
160 // TODO: Replace by calc_rffac() after refactoring that
161 ic.reactionFieldCoefficient = 0.5 * std::pow(ic.rcoulomb, -3);
162 ic.reactionFieldShift = 1 / ic.rcoulomb + ic.reactionFieldCoefficient * ic.rcoulomb * ic.rcoulomb;
164 if (EEL_PME_EWALD(ic.eeltype))
166 // Ewald coefficients, we ignore the potential shift
167 GMX_RELEASE_ASSERT(options.ewaldcoeff_q > 0, "Ewald coefficient should be > 0");
168 ic.ewaldcoeff_q = options.ewaldcoeff_q;
169 ic.coulombEwaldTables = std::make_unique<EwaldCorrectionTables>();
170 init_interaction_const_tables(nullptr, &ic, 0, 0);
176 //! Sets up and returns a Nbnxm object for the given benchmark options and system
177 static std::unique_ptr<nonbonded_verlet_t> setupNbnxmForBenchInstance(const KernelBenchOptions& options,
178 const gmx::BenchmarkSystem& system)
180 const auto pinPolicy = (options.useGpu ? gmx::PinningPolicy::PinnedIfSupported
181 : gmx::PinningPolicy::CannotBePinned);
182 const int numThreads = options.numThreads;
183 // Note: the options and Nbnxm combination rule enums values should match
184 const int combinationRule = static_cast<int>(options.ljCombinationRule);
186 auto messageWhenInvalid = checkKernelSetup(options);
187 if (messageWhenInvalid)
189 gmx_fatal(FARGS, "Requested kernel is unavailable because %s.", messageWhenInvalid->c_str());
191 Nbnxm::KernelSetup kernelSetup = getKernelSetup(options);
193 PairlistParams pairlistParams(kernelSetup.kernelType, false, options.pairlistCutoff, false);
196 PbcType::Xyz, false, nullptr, nullptr, pairlistParams.pairlistType, false, numThreads, pinPolicy);
198 auto pairlistSets = std::make_unique<PairlistSets>(pairlistParams, false, 0);
200 auto pairSearch = std::make_unique<PairSearch>(
201 PbcType::Xyz, false, nullptr, nullptr, pairlistParams.pairlistType, false, numThreads, pinPolicy);
203 auto atomData = std::make_unique<nbnxn_atomdata_t>(pinPolicy);
205 // Put everything together
206 auto nbv = std::make_unique<nonbonded_verlet_t>(
207 std::move(pairlistSets), std::move(pairSearch), std::move(atomData), kernelSetup, nullptr, nullptr);
209 nbnxn_atomdata_init(gmx::MDLogger(),
211 kernelSetup.kernelType,
214 system.nonbondedParameters,
220 GMX_RELEASE_ASSERT(!TRICLINIC(system.box), "Only rectangular unit-cells are supported here");
221 const rvec lowerCorner = { 0, 0, 0 };
222 const rvec upperCorner = { system.box[XX][XX], system.box[YY][YY], system.box[ZZ][ZZ] };
224 gmx::ArrayRef<const int> atomInfo;
225 if (options.useHalfLJOptimization)
227 atomInfo = system.atomInfoOxygenVdw;
231 atomInfo = system.atomInfoAllVdw;
234 const real atomDensity = system.coordinates.size() / det(system.box);
236 nbnxn_put_on_grid(nbv.get(),
242 { 0, int(system.coordinates.size()) },
249 nbv->constructPairlist(gmx::InteractionLocality::Local, system.excls, 0, &nrnb);
251 nbv->setAtomProperties(system.atomTypes, system.charges, atomInfo);
256 //! Add the options instance to the list for all requested kernel SIMD types
257 static void expandSimdOptionAndPushBack(const KernelBenchOptions& options,
258 std::vector<KernelBenchOptions>* optionsList)
260 if (options.nbnxmSimd == BenchMarkKernels::SimdAuto)
262 bool addedInstance = false;
263 #ifdef GMX_NBNXN_SIMD_4XN
264 optionsList->push_back(options);
265 optionsList->back().nbnxmSimd = BenchMarkKernels::Simd4XM;
266 addedInstance = true;
268 #ifdef GMX_NBNXN_SIMD_2XNN
269 optionsList->push_back(options);
270 optionsList->back().nbnxmSimd = BenchMarkKernels::Simd2XMM;
271 addedInstance = true;
275 optionsList->push_back(options);
276 optionsList->back().nbnxmSimd = BenchMarkKernels::SimdNo;
281 optionsList->push_back(options);
285 //! Sets up and runs the requested benchmark instance and prints the results
287 // When \p doWarmup is true runs the warmup iterations instead
288 // of the normal ones and does not print any results
289 static void setupAndRunInstance(const gmx::BenchmarkSystem& system,
290 const KernelBenchOptions& options,
293 // Generate an, accurate, estimate of the number of non-zero pair interactions
294 const real atomDensity = system.coordinates.size() / det(system.box);
295 const real numPairsWithinCutoff =
296 atomDensity * 4.0 / 3.0 * M_PI * std::pow(options.pairlistCutoff, 3);
297 const real numUsefulPairs = system.coordinates.size() * 0.5 * (numPairsWithinCutoff + 1);
299 std::unique_ptr<nonbonded_verlet_t> nbv = setupNbnxmForBenchInstance(options, system);
301 // We set the interaction cut-off to the pairlist cut-off
302 interaction_const_t ic = setupInteractionConst(options);
306 gmx_enerdata_t enerd(1, 0);
308 gmx::StepWorkload stepWork;
309 stepWork.computeForces = true;
310 if (options.computeVirialAndEnergy)
312 stepWork.computeVirial = true;
313 stepWork.computeEnergy = true;
316 const gmx::EnumerationArray<BenchMarkKernels, std::string> kernelNames = {
317 "auto", "no", "4xM", "2xMM"
320 const gmx::EnumerationArray<BenchMarkCombRule, std::string> combruleNames = { "geom.",
327 "%-7s %-4s %-5s %-4s ",
328 options.coulombType == BenchMarkCoulomb::Pme ? "Ewald" : "RF",
329 options.useHalfLJOptimization ? "half" : "all",
330 combruleNames[options.ljCombinationRule].c_str(),
331 kernelNames[options.nbnxmSimd].c_str());
332 if (!options.outputFile.empty())
335 "\"%d\",\"%zu\",\"%g\",\"%d\",\"%d\",\"%s\",\"%s\",\"%s\",\"%s\",\"%s\",\"%"
338 (options.nbnxmSimd != BenchMarkKernels::SimdNo) ? GMX_SIMD_REAL_WIDTH : 0,
342 system.coordinates.size(),
343 options.pairlistCutoff,
345 options.numIterations,
346 options.computeVirialAndEnergy ? "yes" : "no",
347 (options.coulombType != BenchMarkCoulomb::ReactionField)
348 ? ((options.nbnxmSimd == BenchMarkKernels::SimdNo || options.useTabulatedEwaldCorr)
352 options.coulombType == BenchMarkCoulomb::Pme ? "Ewald" : "RF",
353 options.useHalfLJOptimization ? "half" : "all",
354 combruleNames[options.ljCombinationRule].c_str(),
355 kernelNames[options.nbnxmSimd].c_str());
359 // Run pre-iteration to avoid cache misses
360 for (int iter = 0; iter < options.numPreIterations; iter++)
362 nbv->dispatchNonbondedKernel(
363 gmx::InteractionLocality::Local, ic, stepWork, enbvClearFYes, system.forceRec, &enerd, &nrnb);
366 const int numIterations = (doWarmup ? options.numWarmupIterations : options.numIterations);
367 const PairlistSet& pairlistSet = nbv->pairlistSets().pairlistSet(gmx::InteractionLocality::Local);
368 const gmx::index numPairs = pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_ + pairlistSet.natpair_q_;
369 gmx_cycles_t cycles = gmx_cycles_read();
370 for (int iter = 0; iter < numIterations; iter++)
372 // Run the kernel without force clearing
373 nbv->dispatchNonbondedKernel(
374 gmx::InteractionLocality::Local, ic, stepWork, enbvClearFNo, system.forceRec, &enerd, &nrnb);
376 cycles = gmx_cycles_read() - cycles;
379 if (options.reportTime)
381 const double uSec = static_cast<double>(cycles) * gmx_cycles_calibrate(1.0) * 1.e6;
382 if (options.cyclesPerPair)
385 "%13.2f %13.3f %10.3f %10.3f\n",
387 uSec / options.numIterations,
388 uSec / (options.numIterations * numPairs),
389 uSec / (options.numIterations * numUsefulPairs));
390 if (!options.outputFile.empty())
393 "\"%.3f\",\"%.4f\",\"%.4f\",\"%.4f\"\n",
395 uSec / options.numIterations,
396 uSec / (options.numIterations * numPairs),
397 uSec / (options.numIterations * numUsefulPairs));
403 "%13.2f %13.3f %10.3f %10.3f\n",
405 uSec / options.numIterations,
406 options.numIterations * numPairs / uSec,
407 options.numIterations * numUsefulPairs / uSec);
408 if (!options.outputFile.empty())
411 "\"%.3f\",\"%.4f\",\"%.4f\",\"%.4f\"\n",
413 uSec / options.numIterations,
414 options.numIterations * numPairs / uSec,
415 options.numIterations * numUsefulPairs / uSec);
421 const double dCycles = static_cast<double>(cycles);
422 if (options.cyclesPerPair)
425 "%10.3f %10.4f %8.4f %8.4f\n",
427 dCycles / options.numIterations * 1e-6,
428 dCycles / (options.numIterations * numPairs),
429 dCycles / (options.numIterations * numUsefulPairs));
434 "%10.3f %10.4f %8.4f %8.4f\n",
436 dCycles / options.numIterations * 1e-6,
437 options.numIterations * numPairs / dCycles,
438 options.numIterations * numUsefulPairs / dCycles);
444 void bench(const int sizeFactor, const KernelBenchOptions& options)
446 // We don't want to call gmx_omp_nthreads_init(), so we init what we need
447 gmx_omp_nthreads_set(emntPairsearch, options.numThreads);
448 gmx_omp_nthreads_set(emntNonbonded, options.numThreads);
450 const gmx::BenchmarkSystem system(sizeFactor, options.outputFile);
452 real minBoxSize = norm(system.box[XX]);
453 for (int dim = YY; dim < DIM; dim++)
455 minBoxSize = std::min(minBoxSize, norm(system.box[dim]));
457 if (options.pairlistCutoff > 0.5 * minBoxSize)
459 gmx_fatal(FARGS, "The cut-off should be shorter than half the box size");
462 std::vector<KernelBenchOptions> optionsList;
465 KernelBenchOptions opt = options;
466 gmx::EnumerationWrapper<BenchMarkCoulomb> coulombIter;
467 for (auto coulombType : coulombIter)
469 opt.coulombType = coulombType;
470 for (int halfLJ = 0; halfLJ <= 1; halfLJ++)
472 opt.useHalfLJOptimization = (halfLJ == 1);
474 gmx::EnumerationWrapper<BenchMarkCombRule> combRuleIter;
475 for (auto combRule : combRuleIter)
477 opt.ljCombinationRule = combRule;
479 expandSimdOptionAndPushBack(opt, &optionsList);
486 expandSimdOptionAndPushBack(options, &optionsList);
488 GMX_RELEASE_ASSERT(!optionsList.empty(), "Expect at least on benchmark setup");
491 if (options.nbnxmSimd != BenchMarkKernels::SimdNo)
493 fprintf(stdout, "SIMD width: %d\n", GMX_SIMD_REAL_WIDTH);
496 fprintf(stdout, "System size: %zu atoms\n", system.coordinates.size());
497 fprintf(stdout, "Cut-off radius: %g nm\n", options.pairlistCutoff);
498 fprintf(stdout, "Number of threads: %d\n", options.numThreads);
499 fprintf(stdout, "Number of iterations: %d\n", options.numIterations);
500 fprintf(stdout, "Compute energies: %s\n", options.computeVirialAndEnergy ? "yes" : "no");
501 if (options.coulombType != BenchMarkCoulomb::ReactionField)
504 "Ewald excl. corr.: %s\n",
505 options.nbnxmSimd == BenchMarkKernels::SimdNo || options.useTabulatedEwaldCorr
511 if (options.numWarmupIterations > 0)
513 setupAndRunInstance(system, optionsList[0], true);
516 if (options.reportTime)
519 "Coulomb LJ comb. SIMD usec usec/it. %s\n",
520 options.cyclesPerPair ? "usec/pair" : "pairs/usec");
521 if (!options.outputFile.empty())
524 "\"width\",\"atoms\",\"cut-off radius\",\"threads\",\"iter\",\"compute "
525 "energy\",\"Ewald excl. "
526 "corr.\",\"Coulomb\",\"LJ\",\"comb\",\"SIMD\",\"usec\",\"usec/it\",\"total "
527 "pairs/usec\",\"useful pairs/usec\"\n");
535 "Coulomb LJ comb. SIMD Mcycles Mcycles/it. %s\n",
536 options.cyclesPerPair ? "cycles/pair" : "pairs/cycle");
537 if (!options.outputFile.empty())
540 "\"width\",\"atoms\",\"cut-off radius\",\"threads\",\"iter\",\"compute "
541 "energy\",\"Ewald excl. "
542 "corr.\",\"Coulomb\",\"LJ\",\"comb\",\"SIMD\",\"Mcycles\",\"Mcycles/"
544 "total cycles/pair\",\"total cycles per useful pair\"\n");
546 fprintf(stdout, " total useful\n");
549 for (const auto& optionsInstance : optionsList)
551 setupAndRunInstance(system, optionsInstance, false);
554 if (!options.outputFile.empty())