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/mdlib/dispersioncorrection.h"
  52 #include "gromacs/mdlib/force_flags.h"
  53 #include "gromacs/mdlib/forcerec.h"
  54 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  55 #include "gromacs/mdtypes/enerdata.h"
  56 #include "gromacs/mdtypes/forcerec.h"
  57 #include "gromacs/mdtypes/interaction_const.h"
  58 #include "gromacs/mdtypes/mdatom.h"
  59 #include "gromacs/mdtypes/simulation_workload.h"
  60 #include "gromacs/nbnxm/atomdata.h"
  61 #include "gromacs/nbnxm/gridset.h"
  62 #include "gromacs/nbnxm/nbnxm.h"
  63 #include "gromacs/nbnxm/nbnxm_simd.h"
  64 #include "gromacs/nbnxm/pairlistset.h"
  65 #include "gromacs/nbnxm/pairlistsets.h"
  66 #include "gromacs/nbnxm/pairsearch.h"
  67 #include "gromacs/pbcutil/ishift.h"
  68 #include "gromacs/pbcutil/pbc.h"
  69 #include "gromacs/simd/simd.h"
  70 #include "gromacs/timing/cyclecounter.h"
  71 #include "gromacs/utility/enumerationhelpers.h"
  72 #include "gromacs/utility/fatalerror.h"
  73 #include "gromacs/utility/logger.h"
  74
  75 #include "bench_system.h"
  76
  77 namespace Nbnxm
  78 {
  79
  80 /*! \brief Checks the kernel setup
  81  *
  82  * Returns an error string when the kernel is not available.
  83  */
  84 static std::optional<std::string> checkKernelSetup(const KernelBenchOptions& options)
  85 {
  86     GMX_RELEASE_ASSERT(options.nbnxmSimd < BenchMarkKernels::Count
  87                                && options.nbnxmSimd != BenchMarkKernels::SimdAuto,
  88                        "Need a valid kernel SIMD type");
  89
  90     // Check SIMD support
  91     if ((options.nbnxmSimd != BenchMarkKernels::SimdNo && !GMX_SIMD)
  92 #ifndef GMX_NBNXN_SIMD_4XN
  93         || options.nbnxmSimd == BenchMarkKernels::Simd4XM
  94 #endif
  95 #ifndef GMX_NBNXN_SIMD_2XNN
  96         || options.nbnxmSimd == BenchMarkKernels::Simd2XMM
  97 #endif
  98     )
  99     {
 100         return "the requested SIMD kernel was not set up at configuration time";
 101     }
 102
 103     if (options.reportTime && (0 > gmx_cycles_calibrate(1.0)))
 104     {
 105         return "the -time option is not supported on this system";
 106     }
 107
 108     return {};
 109 }
 110
 111 //! Helper to translate between the different enumeration values.
 112 static KernelType translateBenchmarkEnum(const BenchMarkKernels& kernel)
 113 {
 114     int kernelInt = static_cast<int>(kernel);
 115     return static_cast<KernelType>(kernelInt);
 116 }
 117
 118 /*! \brief Returns the kernel setup
 119  */
 120 static KernelSetup getKernelSetup(const KernelBenchOptions& options)
 121 {
 122     auto messageWhenInvalid = checkKernelSetup(options);
 123     GMX_RELEASE_ASSERT(!messageWhenInvalid, "Need valid options");
 124
 125     KernelSetup kernelSetup;
 126
 127     // The int enum options.nbnxnSimd is set up to match KernelType + 1
 128     kernelSetup.kernelType = translateBenchmarkEnum(options.nbnxmSimd);
 129     // The plain-C kernel does not support analytical ewald correction
 130     if (kernelSetup.kernelType == KernelType::Cpu4x4_PlainC)
 131     {
 132         kernelSetup.ewaldExclusionType = EwaldExclusionType::Table;
 133     }
 134     else
 135     {
 136         kernelSetup.ewaldExclusionType = options.useTabulatedEwaldCorr ? EwaldExclusionType::Table
 137                                                                        : EwaldExclusionType::Analytical;
 138     }
 139
 140     return kernelSetup;
 141 }
 142
 143 //! Return an interaction constants struct with members used in the benchmark set appropriately
 144 static interaction_const_t setupInteractionConst(const KernelBenchOptions& options)
 145
 146 {
 147     interaction_const_t ic;
 148
 149     ic.vdwtype      = VanDerWaalsType::Cut;
 150     ic.vdw_modifier = InteractionModifiers::PotShift;
 151     ic.rvdw         = options.pairlistCutoff;
 152
 153     ic.eeltype = (options.coulombType == BenchMarkCoulomb::Pme ? CoulombInteractionType::Pme
 154                                                                : CoulombInteractionType::RF);
 155     ic.coulomb_modifier = InteractionModifiers::PotShift;
 156     ic.rcoulomb         = options.pairlistCutoff;
 157
 158     // Reaction-field with reactionFieldPermitivity=inf
 159     // TODO: Replace by calc_rffac() after refactoring that
 160     ic.reactionFieldCoefficient = 0.5 * std::pow(ic.rcoulomb, -3);
 161     ic.reactionFieldShift = 1 / ic.rcoulomb + ic.reactionFieldCoefficient * ic.rcoulomb * ic.rcoulomb;
 162
 163     if (EEL_PME_EWALD(ic.eeltype))
 164     {
 165         // Ewald coefficients, we ignore the potential shift
 166         GMX_RELEASE_ASSERT(options.ewaldcoeff_q > 0, "Ewald coefficient should be > 0");
 167         ic.ewaldcoeff_q       = options.ewaldcoeff_q;
 168         ic.coulombEwaldTables = std::make_unique<EwaldCorrectionTables>();
 169         init_interaction_const_tables(nullptr, &ic, 0, 0);
 170     }
 171
 172     return ic;
 173 }
 174
 175 //! Sets up and returns a Nbnxm object for the given benchmark options and system
 176 static std::unique_ptr<nonbonded_verlet_t> setupNbnxmForBenchInstance(const KernelBenchOptions& options,
 177                                                                       const gmx::BenchmarkSystem& system)
 178 {
 179     const auto pinPolicy  = (options.useGpu ? gmx::PinningPolicy::PinnedIfSupported
 180                                            : gmx::PinningPolicy::CannotBePinned);
 181     const int  numThreads = options.numThreads;
 182     // Note: the options and Nbnxm combination rule enums values should match
 183     const int combinationRule = static_cast<int>(options.ljCombinationRule);
 184
 185     auto messageWhenInvalid = checkKernelSetup(options);
 186     if (messageWhenInvalid)
 187     {
 188         gmx_fatal(FARGS, "Requested kernel is unavailable because %s.", messageWhenInvalid->c_str());
 189     }
 190     Nbnxm::KernelSetup kernelSetup = getKernelSetup(options);
 191
 192     PairlistParams pairlistParams(kernelSetup.kernelType, false, options.pairlistCutoff, false);
 193
 194     GridSet gridSet(
 195             PbcType::Xyz, false, nullptr, nullptr, pairlistParams.pairlistType, false, numThreads, pinPolicy);
 196
 197     auto pairlistSets = std::make_unique<PairlistSets>(pairlistParams, false, 0);
 198
 199     auto pairSearch = std::make_unique<PairSearch>(
 200             PbcType::Xyz, false, nullptr, nullptr, pairlistParams.pairlistType, false, numThreads, pinPolicy);
 201
 202     auto atomData = std::make_unique<nbnxn_atomdata_t>(pinPolicy);
 203
 204     // Put everything together
 205     auto nbv = std::make_unique<nonbonded_verlet_t>(
 206             std::move(pairlistSets), std::move(pairSearch), std::move(atomData), kernelSetup, nullptr, nullptr);
 207
 208     nbnxn_atomdata_init(gmx::MDLogger(),
 209                         nbv->nbat.get(),
 210                         kernelSetup.kernelType,
 211                         combinationRule,
 212                         system.numAtomTypes,
 213                         system.nonbondedParameters,
 214                         1,
 215                         numThreads);
 216
 217     t_nrnb nrnb;
 218
 219     GMX_RELEASE_ASSERT(!TRICLINIC(system.box), "Only rectangular unit-cells are supported here");
 220     const rvec lowerCorner = { 0, 0, 0 };
 221     const rvec upperCorner = { system.box[XX][XX], system.box[YY][YY], system.box[ZZ][ZZ] };
 222
 223     gmx::ArrayRef<const int> atomInfo;
 224     if (options.useHalfLJOptimization)
 225     {
 226         atomInfo = system.atomInfoOxygenVdw;
 227     }
 228     else
 229     {
 230         atomInfo = system.atomInfoAllVdw;
 231     }
 232
 233     const real atomDensity = system.coordinates.size() / det(system.box);
 234
 235     nbnxn_put_on_grid(nbv.get(),
 236                       system.box,
 237                       0,
 238                       lowerCorner,
 239                       upperCorner,
 240                       nullptr,
 241                       { 0, int(system.coordinates.size()) },
 242                       atomDensity,
 243                       atomInfo,
 244                       system.coordinates,
 245                       0,
 246                       nullptr);
 247
 248     nbv->constructPairlist(gmx::InteractionLocality::Local, system.excls, 0, &nrnb);
 249
 250     nbv->setAtomProperties(system.atomTypes, system.charges, atomInfo);
 251
 252     return nbv;
 253 }
 254
 255 //! Add the options instance to the list for all requested kernel SIMD types
 256 static void expandSimdOptionAndPushBack(const KernelBenchOptions&        options,
 257                                         std::vector<KernelBenchOptions>* optionsList)
 258 {
 259     if (options.nbnxmSimd == BenchMarkKernels::SimdAuto)
 260     {
 261         bool addedInstance = false;
 262 #ifdef GMX_NBNXN_SIMD_4XN
 263         optionsList->push_back(options);
 264         optionsList->back().nbnxmSimd = BenchMarkKernels::Simd4XM;
 265         addedInstance                 = true;
 266 #endif
 267 #ifdef GMX_NBNXN_SIMD_2XNN
 268         optionsList->push_back(options);
 269         optionsList->back().nbnxmSimd = BenchMarkKernels::Simd2XMM;
 270         addedInstance                 = true;
 271 #endif
 272         if (!addedInstance)
 273         {
 274             optionsList->push_back(options);
 275             optionsList->back().nbnxmSimd = BenchMarkKernels::SimdNo;
 276         }
 277     }
 278     else
 279     {
 280         optionsList->push_back(options);
 281     }
 282 }
 283
 284 //! Sets up and runs the requested benchmark instance and prints the results
 285 //
 286 // When \p doWarmup is true runs the warmup iterations instead
 287 // of the normal ones and does not print any results
 288 static void setupAndRunInstance(const gmx::BenchmarkSystem& system,
 289                                 const KernelBenchOptions&   options,
 290                                 const bool                  doWarmup)
 291 {
 292     // Generate an, accurate, estimate of the number of non-zero pair interactions
 293     const real atomDensity = system.coordinates.size() / det(system.box);
 294     const real numPairsWithinCutoff =
 295             atomDensity * 4.0 / 3.0 * M_PI * std::pow(options.pairlistCutoff, 3);
 296     const real numUsefulPairs = system.coordinates.size() * 0.5 * (numPairsWithinCutoff + 1);
 297
 298     std::unique_ptr<nonbonded_verlet_t> nbv = setupNbnxmForBenchInstance(options, system);
 299
 300     // We set the interaction cut-off to the pairlist cut-off
 301     interaction_const_t ic = setupInteractionConst(options);
 302
 303     t_nrnb nrnb = { 0 };
 304
 305     gmx_enerdata_t enerd(1, 0);
 306
 307     gmx::StepWorkload stepWork;
 308     stepWork.computeForces = true;
 309     if (options.computeVirialAndEnergy)
 310     {
 311         stepWork.computeVirial = true;
 312         stepWork.computeEnergy = true;
 313     }
 314
 315     const gmx::EnumerationArray<BenchMarkKernels, std::string> kernelNames = {
 316         "auto", "no", "4xM", "2xMM"
 317     };
 318
 319     const gmx::EnumerationArray<BenchMarkCombRule, std::string> combruleNames = { "geom.",
 320                                                                                   "LB",
 321                                                                                   "none" };
 322
 323     if (!doWarmup)
 324     {
 325         fprintf(stdout,
 326                 "%-7s %-4s %-5s %-4s ",
 327                 options.coulombType == BenchMarkCoulomb::Pme ? "Ewald" : "RF",
 328                 options.useHalfLJOptimization ? "half" : "all",
 329                 combruleNames[options.ljCombinationRule].c_str(),
 330                 kernelNames[options.nbnxmSimd].c_str());
 331         if (!options.outputFile.empty())
 332         {
 333             fprintf(system.csv,
 334                     "\"%d\",\"%zu\",\"%g\",\"%d\",\"%d\",\"%s\",\"%s\",\"%s\",\"%s\",\"%s\",\"%"
 335                     "s\",",
 336 #if GMX_SIMD
 337                     (options.nbnxmSimd != BenchMarkKernels::SimdNo) ? GMX_SIMD_REAL_WIDTH : 0,
 338 #else
 339                     0,
 340 #endif
 341                     system.coordinates.size(),
 342                     options.pairlistCutoff,
 343                     options.numThreads,
 344                     options.numIterations,
 345                     options.computeVirialAndEnergy ? "yes" : "no",
 346                     (options.coulombType != BenchMarkCoulomb::ReactionField)
 347                             ? ((options.nbnxmSimd == BenchMarkKernels::SimdNo || options.useTabulatedEwaldCorr)
 348                                        ? "table"
 349                                        : "analytical")
 350                             : "",
 351                     options.coulombType == BenchMarkCoulomb::Pme ? "Ewald" : "RF",
 352                     options.useHalfLJOptimization ? "half" : "all",
 353                     combruleNames[options.ljCombinationRule].c_str(),
 354                     kernelNames[options.nbnxmSimd].c_str());
 355         }
 356     }
 357
 358     // Run pre-iteration to avoid cache misses
 359     for (int iter = 0; iter < options.numPreIterations; iter++)
 360     {
 361         nbv->dispatchNonbondedKernel(
 362                 gmx::InteractionLocality::Local, ic, stepWork, enbvClearFYes, system.forceRec, &enerd, &nrnb);
 363     }
 364
 365     const int numIterations = (doWarmup ? options.numWarmupIterations : options.numIterations);
 366     const PairlistSet& pairlistSet = nbv->pairlistSets().pairlistSet(gmx::InteractionLocality::Local);
 367     const gmx::index numPairs = pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_ + pairlistSet.natpair_q_;
 368     gmx_cycles_t cycles = gmx_cycles_read();
 369     for (int iter = 0; iter < numIterations; iter++)
 370     {
 371         // Run the kernel without force clearing
 372         nbv->dispatchNonbondedKernel(
 373                 gmx::InteractionLocality::Local, ic, stepWork, enbvClearFNo, system.forceRec, &enerd, &nrnb);
 374     }
 375     cycles = gmx_cycles_read() - cycles;
 376     if (!doWarmup)
 377     {
 378         if (options.reportTime)
 379         {
 380             const double uSec = static_cast<double>(cycles) * gmx_cycles_calibrate(1.0) * 1.e6;
 381             if (options.cyclesPerPair)
 382             {
 383                 fprintf(stdout,
 384                         "%13.2f %13.3f %10.3f %10.3f\n",
 385                         uSec,
 386                         uSec / options.numIterations,
 387                         uSec / (options.numIterations * numPairs),
 388                         uSec / (options.numIterations * numUsefulPairs));
 389                 if (!options.outputFile.empty())
 390                 {
 391                     fprintf(system.csv,
 392                             "\"%.3f\",\"%.4f\",\"%.4f\",\"%.4f\"\n",
 393                             uSec,
 394                             uSec / options.numIterations,
 395                             uSec / (options.numIterations * numPairs),
 396                             uSec / (options.numIterations * numUsefulPairs));
 397                 }
 398             }
 399             else
 400             {
 401                 fprintf(stdout,
 402                         "%13.2f %13.3f %10.3f %10.3f\n",
 403                         uSec,
 404                         uSec / options.numIterations,
 405                         options.numIterations * numPairs / uSec,
 406                         options.numIterations * numUsefulPairs / uSec);
 407                 if (!options.outputFile.empty())
 408                 {
 409                     fprintf(system.csv,
 410                             "\"%.3f\",\"%.4f\",\"%.4f\",\"%.4f\"\n",
 411                             uSec,
 412                             uSec / options.numIterations,
 413                             options.numIterations * numPairs / uSec,
 414                             options.numIterations * numUsefulPairs / uSec);
 415                 }
 416             }
 417         }
 418         else
 419         {
 420             const double dCycles = static_cast<double>(cycles);
 421             if (options.cyclesPerPair)
 422             {
 423                 fprintf(stdout,
 424                         "%10.3f %10.4f %8.4f %8.4f\n",
 425                         cycles * 1e-6,
 426                         dCycles / options.numIterations * 1e-6,
 427                         dCycles / (options.numIterations * numPairs),
 428                         dCycles / (options.numIterations * numUsefulPairs));
 429             }
 430             else
 431             {
 432                 fprintf(stdout,
 433                         "%10.3f %10.4f %8.4f %8.4f\n",
 434                         dCycles * 1e-6,
 435                         dCycles / options.numIterations * 1e-6,
 436                         options.numIterations * numPairs / dCycles,
 437                         options.numIterations * numUsefulPairs / dCycles);
 438             }
 439         }
 440     }
 441 }
 442
 443 void bench(const int sizeFactor, const KernelBenchOptions& options)
 444 {
 445     // We don't want to call gmx_omp_nthreads_init(), so we init what we need
 446     gmx_omp_nthreads_set(emntPairsearch, options.numThreads);
 447     gmx_omp_nthreads_set(emntNonbonded, options.numThreads);
 448
 449     const gmx::BenchmarkSystem system(sizeFactor, options.outputFile);
 450
 451     real minBoxSize = norm(system.box[XX]);
 452     for (int dim = YY; dim < DIM; dim++)
 453     {
 454         minBoxSize = std::min(minBoxSize, norm(system.box[dim]));
 455     }
 456     if (options.pairlistCutoff > 0.5 * minBoxSize)
 457     {
 458         gmx_fatal(FARGS, "The cut-off should be shorter than half the box size");
 459     }
 460
 461     std::vector<KernelBenchOptions> optionsList;
 462     if (options.doAll)
 463     {
 464         KernelBenchOptions                        opt = options;
 465         gmx::EnumerationWrapper<BenchMarkCoulomb> coulombIter;
 466         for (auto coulombType : coulombIter)
 467         {
 468             opt.coulombType = coulombType;
 469             for (int halfLJ = 0; halfLJ <= 1; halfLJ++)
 470             {
 471                 opt.useHalfLJOptimization = (halfLJ == 1);
 472
 473                 gmx::EnumerationWrapper<BenchMarkCombRule> combRuleIter;
 474                 for (auto combRule : combRuleIter)
 475                 {
 476                     opt.ljCombinationRule = combRule;
 477
 478                     expandSimdOptionAndPushBack(opt, &optionsList);
 479                 }
 480             }
 481         }
 482     }
 483     else
 484     {
 485         expandSimdOptionAndPushBack(options, &optionsList);
 486     }
 487     GMX_RELEASE_ASSERT(!optionsList.empty(), "Expect at least on benchmark setup");
 488
 489 #if GMX_SIMD
 490     if (options.nbnxmSimd != BenchMarkKernels::SimdNo)
 491     {
 492         fprintf(stdout, "SIMD width:           %d\n", GMX_SIMD_REAL_WIDTH);
 493     }
 494 #endif
 495     fprintf(stdout, "System size:          %zu atoms\n", system.coordinates.size());
 496     fprintf(stdout, "Cut-off radius:       %g nm\n", options.pairlistCutoff);
 497     fprintf(stdout, "Number of threads:    %d\n", options.numThreads);
 498     fprintf(stdout, "Number of iterations: %d\n", options.numIterations);
 499     fprintf(stdout, "Compute energies:     %s\n", options.computeVirialAndEnergy ? "yes" : "no");
 500     if (options.coulombType != BenchMarkCoulomb::ReactionField)
 501     {
 502         fprintf(stdout,
 503                 "Ewald excl. corr.:    %s\n",
 504                 options.nbnxmSimd == BenchMarkKernels::SimdNo || options.useTabulatedEwaldCorr
 505                         ? "table"
 506                         : "analytical");
 507     }
 508     printf("\n");
 509
 510     if (options.numWarmupIterations > 0)
 511     {
 512         setupAndRunInstance(system, optionsList[0], true);
 513     }
 514
 515     if (options.reportTime)
 516     {
 517         fprintf(stdout,
 518                 "Coulomb LJ   comb. SIMD       usec         usec/it.        %s\n",
 519                 options.cyclesPerPair ? "usec/pair" : "pairs/usec");
 520         if (!options.outputFile.empty())
 521         {
 522             fprintf(system.csv,
 523                     "\"width\",\"atoms\",\"cut-off radius\",\"threads\",\"iter\",\"compute "
 524                     "energy\",\"Ewald excl. "
 525                     "corr.\",\"Coulomb\",\"LJ\",\"comb\",\"SIMD\",\"usec\",\"usec/it\",\"total "
 526                     "pairs/usec\",\"useful pairs/usec\"\n");
 527         }
 528         fprintf(stdout,
 529                 "                                                        total      useful\n");
 530     }
 531     else
 532     {
 533         fprintf(stdout,
 534                 "Coulomb LJ   comb. SIMD    Mcycles  Mcycles/it.   %s\n",
 535                 options.cyclesPerPair ? "cycles/pair" : "pairs/cycle");
 536         if (!options.outputFile.empty())
 537         {
 538             fprintf(system.csv,
 539                     "\"width\",\"atoms\",\"cut-off radius\",\"threads\",\"iter\",\"compute "
 540                     "energy\",\"Ewald excl. "
 541                     "corr.\",\"Coulomb\",\"LJ\",\"comb\",\"SIMD\",\"Mcycles\",\"Mcycles/"
 542                     "it\",\"total "
 543                     "total cycles/pair\",\"total cycles per useful pair\"\n");
 544         }
 545         fprintf(stdout, "                                                total    useful\n");
 546     }
 547
 548     for (const auto& optionsInstance : optionsList)
 549     {
 550         setupAndRunInstance(system, optionsInstance, false);
 551     }
 552
 553     if (!options.outputFile.empty())
 554     {
 555         fclose(system.csv);
 556     }
 557 }
 558
 559 } // namespace Nbnxm