src/gromacs/nbnxm/benchmark/bench_setup.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   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.
   8  *
   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.
  13  *
  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.
  18  *
  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.
  23  *
  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.
  31  *
  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.
  34  */
  35
  36 /*! \internal \file
  37  * \brief
  38  * This file defines functions for setting up kernel benchmarks
  39  *
  40  * \author Berk Hess <hess@kth.se>
  41  * \ingroup module_nbnxm
  42  */
  43
  44 #include "gmxpre.h"
  45
  46 #include "bench_setup.h"
  47
  48 #include <optional>
  49
  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"
  75
  76 #include "bench_system.h"
  77
  78 namespace Nbnxm
  79 {
  80
  81 /*! \brief Checks the kernel setup
  82  *
  83  * Returns an error string when the kernel is not available.
  84  */
  85 static std::optional<std::string> checkKernelSetup(const KernelBenchOptions& options)
  86 {
  87     GMX_RELEASE_ASSERT(options.nbnxmSimd < BenchMarkKernels::Count
  88                                && options.nbnxmSimd != BenchMarkKernels::SimdAuto,
  89                        "Need a valid kernel SIMD type");
  90
  91     // Check SIMD support
  92     if ((options.nbnxmSimd != BenchMarkKernels::SimdNo && !GMX_SIMD)
  93 #ifndef GMX_NBNXN_SIMD_4XN
  94         || options.nbnxmSimd == BenchMarkKernels::Simd4XM
  95 #endif
  96 #ifndef GMX_NBNXN_SIMD_2XNN
  97         || options.nbnxmSimd == BenchMarkKernels::Simd2XMM
  98 #endif
  99     )
 100     {
 101         return "the requested SIMD kernel was not set up at configuration time";
 102     }
 103
 104     if (options.reportTime && (0 > gmx_cycles_calibrate(1.0)))
 105     {
 106         return "the -time option is not supported on this system";
 107     }
 108
 109     return {};
 110 }
 111
 112 //! Helper to translate between the different enumeration values.
 113 static KernelType translateBenchmarkEnum(const BenchMarkKernels& kernel)
 114 {
 115     int kernelInt = static_cast<int>(kernel);
 116     return static_cast<KernelType>(kernelInt);
 117 }
 118
 119 /*! \brief Returns the kernel setup
 120  */
 121 static KernelSetup getKernelSetup(const KernelBenchOptions& options)
 122 {
 123     auto messageWhenInvalid = checkKernelSetup(options);
 124     GMX_RELEASE_ASSERT(!messageWhenInvalid, "Need valid options");
 125
 126     KernelSetup kernelSetup;
 127
 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)
 132     {
 133         kernelSetup.ewaldExclusionType = EwaldExclusionType::Table;
 134     }
 135     else
 136     {
 137         kernelSetup.ewaldExclusionType = options.useTabulatedEwaldCorr ? EwaldExclusionType::Table
 138                                                                        : EwaldExclusionType::Analytical;
 139     }
 140
 141     return kernelSetup;
 142 }
 143
 144 //! Return an interaction constants struct with members used in the benchmark set appropriately
 145 static interaction_const_t setupInteractionConst(const KernelBenchOptions& options)
 146
 147 {
 148     interaction_const_t ic;
 149
 150     ic.vdwtype      = VanDerWaalsType::Cut;
 151     ic.vdw_modifier = InteractionModifiers::PotShift;
 152     ic.rvdw         = options.pairlistCutoff;
 153
 154     ic.eeltype = (options.coulombType == BenchMarkCoulomb::Pme ? CoulombInteractionType::Pme
 155                                                                : CoulombInteractionType::RF);
 156     ic.coulomb_modifier = InteractionModifiers::PotShift;
 157     ic.rcoulomb         = options.pairlistCutoff;
 158
 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;
 163
 164     if (EEL_PME_EWALD(ic.eeltype))
 165     {
 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);
 171     }
 172
 173     return ic;
 174 }
 175
 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)
 179 {
 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);
 185
 186     auto messageWhenInvalid = checkKernelSetup(options);
 187     if (messageWhenInvalid)
 188     {
 189         gmx_fatal(FARGS, "Requested kernel is unavailable because %s.", messageWhenInvalid->c_str());
 190     }
 191     Nbnxm::KernelSetup kernelSetup = getKernelSetup(options);
 192
 193     PairlistParams pairlistParams(kernelSetup.kernelType, false, options.pairlistCutoff, false);
 194
 195     GridSet gridSet(
 196             PbcType::Xyz, false, nullptr, nullptr, pairlistParams.pairlistType, false, numThreads, pinPolicy);
 197
 198     auto pairlistSets = std::make_unique<PairlistSets>(pairlistParams, false, 0);
 199
 200     auto pairSearch = std::make_unique<PairSearch>(
 201             PbcType::Xyz, false, nullptr, nullptr, pairlistParams.pairlistType, false, numThreads, pinPolicy);
 202
 203     auto atomData = std::make_unique<nbnxn_atomdata_t>(pinPolicy);
 204
 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);
 208
 209     nbnxn_atomdata_init(gmx::MDLogger(),
 210                         nbv->nbat.get(),
 211                         kernelSetup.kernelType,
 212                         combinationRule,
 213                         system.numAtomTypes,
 214                         system.nonbondedParameters,
 215                         1,
 216                         numThreads);
 217
 218     t_nrnb nrnb;
 219
 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] };
 223
 224     gmx::ArrayRef<const int> atomInfo;
 225     if (options.useHalfLJOptimization)
 226     {
 227         atomInfo = system.atomInfoOxygenVdw;
 228     }
 229     else
 230     {
 231         atomInfo = system.atomInfoAllVdw;
 232     }
 233
 234     const real atomDensity = system.coordinates.size() / det(system.box);
 235
 236     nbnxn_put_on_grid(nbv.get(),
 237                       system.box,
 238                       0,
 239                       lowerCorner,
 240                       upperCorner,
 241                       nullptr,
 242                       { 0, int(system.coordinates.size()) },
 243                       atomDensity,
 244                       atomInfo,
 245                       system.coordinates,
 246                       0,
 247                       nullptr);
 248
 249     nbv->constructPairlist(gmx::InteractionLocality::Local, system.excls, 0, &nrnb);
 250
 251     nbv->setAtomProperties(system.atomTypes, system.charges, atomInfo);
 252
 253     return nbv;
 254 }
 255
 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)
 259 {
 260     if (options.nbnxmSimd == BenchMarkKernels::SimdAuto)
 261     {
 262         bool addedInstance = false;
 263 #ifdef GMX_NBNXN_SIMD_4XN
 264         optionsList->push_back(options);
 265         optionsList->back().nbnxmSimd = BenchMarkKernels::Simd4XM;
 266         addedInstance                 = true;
 267 #endif
 268 #ifdef GMX_NBNXN_SIMD_2XNN
 269         optionsList->push_back(options);
 270         optionsList->back().nbnxmSimd = BenchMarkKernels::Simd2XMM;
 271         addedInstance                 = true;
 272 #endif
 273         if (!addedInstance)
 274         {
 275             optionsList->push_back(options);
 276             optionsList->back().nbnxmSimd = BenchMarkKernels::SimdNo;
 277         }
 278     }
 279     else
 280     {
 281         optionsList->push_back(options);
 282     }
 283 }
 284
 285 //! Sets up and runs the requested benchmark instance and prints the results
 286 //
 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,
 291                                 const bool                  doWarmup)
 292 {
 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);
 298
 299     std::unique_ptr<nonbonded_verlet_t> nbv = setupNbnxmForBenchInstance(options, system);
 300
 301     // We set the interaction cut-off to the pairlist cut-off
 302     interaction_const_t ic = setupInteractionConst(options);
 303
 304     t_nrnb nrnb = { 0 };
 305
 306     gmx_enerdata_t enerd(1, 0);
 307
 308     gmx::StepWorkload stepWork;
 309     stepWork.computeForces = true;
 310     if (options.computeVirialAndEnergy)
 311     {
 312         stepWork.computeVirial = true;
 313         stepWork.computeEnergy = true;
 314     }
 315
 316     const gmx::EnumerationArray<BenchMarkKernels, std::string> kernelNames = {
 317         "auto", "no", "4xM", "2xMM"
 318     };
 319
 320     const gmx::EnumerationArray<BenchMarkCombRule, std::string> combruleNames = { "geom.",
 321                                                                                   "LB",
 322                                                                                   "none" };
 323
 324     if (!doWarmup)
 325     {
 326         fprintf(stdout,
 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())
 333         {
 334             fprintf(system.csv,
 335                     "\"%d\",\"%zu\",\"%g\",\"%d\",\"%d\",\"%s\",\"%s\",\"%s\",\"%s\",\"%s\",\"%"
 336                     "s\",",
 337 #if GMX_SIMD
 338                     (options.nbnxmSimd != BenchMarkKernels::SimdNo) ? GMX_SIMD_REAL_WIDTH : 0,
 339 #else
 340                     0,
 341 #endif
 342                     system.coordinates.size(),
 343                     options.pairlistCutoff,
 344                     options.numThreads,
 345                     options.numIterations,
 346                     options.computeVirialAndEnergy ? "yes" : "no",
 347                     (options.coulombType != BenchMarkCoulomb::ReactionField)
 348                             ? ((options.nbnxmSimd == BenchMarkKernels::SimdNo || options.useTabulatedEwaldCorr)
 349                                        ? "table"
 350                                        : "analytical")
 351                             : "",
 352                     options.coulombType == BenchMarkCoulomb::Pme ? "Ewald" : "RF",
 353                     options.useHalfLJOptimization ? "half" : "all",
 354                     combruleNames[options.ljCombinationRule].c_str(),
 355                     kernelNames[options.nbnxmSimd].c_str());
 356         }
 357     }
 358
 359     // Run pre-iteration to avoid cache misses
 360     for (int iter = 0; iter < options.numPreIterations; iter++)
 361     {
 362         nbv->dispatchNonbondedKernel(
 363                 gmx::InteractionLocality::Local, ic, stepWork, enbvClearFYes, system.forceRec, &enerd, &nrnb);
 364     }
 365
 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++)
 371     {
 372         // Run the kernel without force clearing
 373         nbv->dispatchNonbondedKernel(
 374                 gmx::InteractionLocality::Local, ic, stepWork, enbvClearFNo, system.forceRec, &enerd, &nrnb);
 375     }
 376     cycles = gmx_cycles_read() - cycles;
 377     if (!doWarmup)
 378     {
 379         if (options.reportTime)
 380         {
 381             const double uSec = static_cast<double>(cycles) * gmx_cycles_calibrate(1.0) * 1.e6;
 382             if (options.cyclesPerPair)
 383             {
 384                 fprintf(stdout,
 385                         "%13.2f %13.3f %10.3f %10.3f\n",
 386                         uSec,
 387                         uSec / options.numIterations,
 388                         uSec / (options.numIterations * numPairs),
 389                         uSec / (options.numIterations * numUsefulPairs));
 390                 if (!options.outputFile.empty())
 391                 {
 392                     fprintf(system.csv,
 393                             "\"%.3f\",\"%.4f\",\"%.4f\",\"%.4f\"\n",
 394                             uSec,
 395                             uSec / options.numIterations,
 396                             uSec / (options.numIterations * numPairs),
 397                             uSec / (options.numIterations * numUsefulPairs));
 398                 }
 399             }
 400             else
 401             {
 402                 fprintf(stdout,
 403                         "%13.2f %13.3f %10.3f %10.3f\n",
 404                         uSec,
 405                         uSec / options.numIterations,
 406                         options.numIterations * numPairs / uSec,
 407                         options.numIterations * numUsefulPairs / uSec);
 408                 if (!options.outputFile.empty())
 409                 {
 410                     fprintf(system.csv,
 411                             "\"%.3f\",\"%.4f\",\"%.4f\",\"%.4f\"\n",
 412                             uSec,
 413                             uSec / options.numIterations,
 414                             options.numIterations * numPairs / uSec,
 415                             options.numIterations * numUsefulPairs / uSec);
 416                 }
 417             }
 418         }
 419         else
 420         {
 421             const double dCycles = static_cast<double>(cycles);
 422             if (options.cyclesPerPair)
 423             {
 424                 fprintf(stdout,
 425                         "%10.3f %10.4f %8.4f %8.4f\n",
 426                         cycles * 1e-6,
 427                         dCycles / options.numIterations * 1e-6,
 428                         dCycles / (options.numIterations * numPairs),
 429                         dCycles / (options.numIterations * numUsefulPairs));
 430             }
 431             else
 432             {
 433                 fprintf(stdout,
 434                         "%10.3f %10.4f %8.4f %8.4f\n",
 435                         dCycles * 1e-6,
 436                         dCycles / options.numIterations * 1e-6,
 437                         options.numIterations * numPairs / dCycles,
 438                         options.numIterations * numUsefulPairs / dCycles);
 439             }
 440         }
 441     }
 442 }
 443
 444 void bench(const int sizeFactor, const KernelBenchOptions& options)
 445 {
 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);
 449
 450     const gmx::BenchmarkSystem system(sizeFactor, options.outputFile);
 451
 452     real minBoxSize = norm(system.box[XX]);
 453     for (int dim = YY; dim < DIM; dim++)
 454     {
 455         minBoxSize = std::min(minBoxSize, norm(system.box[dim]));
 456     }
 457     if (options.pairlistCutoff > 0.5 * minBoxSize)
 458     {
 459         gmx_fatal(FARGS, "The cut-off should be shorter than half the box size");
 460     }
 461
 462     std::vector<KernelBenchOptions> optionsList;
 463     if (options.doAll)
 464     {
 465         KernelBenchOptions                        opt = options;
 466         gmx::EnumerationWrapper<BenchMarkCoulomb> coulombIter;
 467         for (auto coulombType : coulombIter)
 468         {
 469             opt.coulombType = coulombType;
 470             for (int halfLJ = 0; halfLJ <= 1; halfLJ++)
 471             {
 472                 opt.useHalfLJOptimization = (halfLJ == 1);
 473
 474                 gmx::EnumerationWrapper<BenchMarkCombRule> combRuleIter;
 475                 for (auto combRule : combRuleIter)
 476                 {
 477                     opt.ljCombinationRule = combRule;
 478
 479                     expandSimdOptionAndPushBack(opt, &optionsList);
 480                 }
 481             }
 482         }
 483     }
 484     else
 485     {
 486         expandSimdOptionAndPushBack(options, &optionsList);
 487     }
 488     GMX_RELEASE_ASSERT(!optionsList.empty(), "Expect at least on benchmark setup");
 489
 490 #if GMX_SIMD
 491     if (options.nbnxmSimd != BenchMarkKernels::SimdNo)
 492     {
 493         fprintf(stdout, "SIMD width:           %d\n", GMX_SIMD_REAL_WIDTH);
 494     }
 495 #endif
 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)
 502     {
 503         fprintf(stdout,
 504                 "Ewald excl. corr.:    %s\n",
 505                 options.nbnxmSimd == BenchMarkKernels::SimdNo || options.useTabulatedEwaldCorr
 506                         ? "table"
 507                         : "analytical");
 508     }
 509     printf("\n");
 510
 511     if (options.numWarmupIterations > 0)
 512     {
 513         setupAndRunInstance(system, optionsList[0], true);
 514     }
 515
 516     if (options.reportTime)
 517     {
 518         fprintf(stdout,
 519                 "Coulomb LJ   comb. SIMD       usec         usec/it.        %s\n",
 520                 options.cyclesPerPair ? "usec/pair" : "pairs/usec");
 521         if (!options.outputFile.empty())
 522         {
 523             fprintf(system.csv,
 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");
 528         }
 529         fprintf(stdout,
 530                 "                                                        total      useful\n");
 531     }
 532     else
 533     {
 534         fprintf(stdout,
 535                 "Coulomb LJ   comb. SIMD    Mcycles  Mcycles/it.   %s\n",
 536                 options.cyclesPerPair ? "cycles/pair" : "pairs/cycle");
 537         if (!options.outputFile.empty())
 538         {
 539             fprintf(system.csv,
 540                     "\"width\",\"atoms\",\"cut-off radius\",\"threads\",\"iter\",\"compute "
 541                     "energy\",\"Ewald excl. "
 542                     "corr.\",\"Coulomb\",\"LJ\",\"comb\",\"SIMD\",\"Mcycles\",\"Mcycles/"
 543                     "it\",\"total "
 544                     "total cycles/pair\",\"total cycles per useful pair\"\n");
 545         }
 546         fprintf(stdout, "                                                total    useful\n");
 547     }
 548
 549     for (const auto& optionsInstance : optionsList)
 550     {
 551         setupAndRunInstance(system, optionsInstance, false);
 552     }
 553
 554     if (!options.outputFile.empty())
 555     {
 556         fclose(system.csv);
 557     }
 558 }
 559
 560 } // namespace Nbnxm