src/gromacs/applied_forces/awh/bias.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2015,2016,2017,2018,2019,2020,2021, 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  * Implements the Bias class.
  39  *
  40  * \author Viveca Lindahl
  41  * \author Berk Hess <hess@kth.se>
  42  * \ingroup module_awh
  43  */
  44
  45 #include "gmxpre.h"
  46
  47 #include "bias.h"
  48
  49 #include <cassert>
  50 #include <cinttypes>
  51 #include <cmath>
  52 #include <cstdio>
  53 #include <cstdlib>
  54 #include <cstring>
  55
  56 #include <algorithm>
  57 #include <memory>
  58
  59 #include "gromacs/fileio/gmxfio.h"
  60 #include "gromacs/gmxlib/network.h"
  61 #include "gromacs/math/functions.h"
  62 #include "gromacs/math/utilities.h"
  63 #include "gromacs/mdtypes/awh_history.h"
  64 #include "gromacs/mdtypes/awh_params.h"
  65 #include "gromacs/mdtypes/commrec.h"
  66 #include "gromacs/utility/exceptions.h"
  67 #include "gromacs/utility/gmxassert.h"
  68 #include "gromacs/utility/stringutil.h"
  69
  70 #include "biassharing.h"
  71 #include "correlationgrid.h"
  72 #include "correlationhistory.h"
  73 #include "pointstate.h"
  74
  75 namespace gmx
  76 {
  77
  78 void Bias::warnForHistogramAnomalies(double t, int64_t step, FILE* fplog)
  79 {
  80     const int maxNumWarningsInCheck = 1;  /* The maximum number of warnings to print per check */
  81     const int maxNumWarningsInRun   = 10; /* The maximum number of warnings to print in a run */
  82
  83     if (fplog == nullptr || numWarningsIssued_ >= maxNumWarningsInRun || state_.inInitialStage()
  84         || !params_.isCheckHistogramForAnomaliesStep(step))
  85     {
  86         return;
  87     }
  88
  89     numWarningsIssued_ +=
  90             state_.warnForHistogramAnomalies(grid_, biasIndex(), t, fplog, maxNumWarningsInCheck);
  91
  92     if (numWarningsIssued_ >= maxNumWarningsInRun)
  93     {
  94         fprintf(fplog, "\nawh%d: suppressing future AWH warnings.\n", biasIndex() + 1);
  95     }
  96 }
  97
  98 void Bias::doSkippedUpdatesForAllPoints()
  99 {
 100     if (params_.skipUpdates())
 101     {
 102         state_.doSkippedUpdatesForAllPoints(params_);
 103     }
 104 }
 105
 106 gmx::ArrayRef<const double> Bias::calcForceAndUpdateBias(const awh_dvec         coordValue,
 107                                                          ArrayRef<const double> neighborLambdaEnergies,
 108                                                          ArrayRef<const double> neighborLambdaDhdl,
 109                                                          double*                awhPotential,
 110                                                          double*                potentialJump,
 111                                                          double                 t,
 112                                                          int64_t                step,
 113                                                          int64_t                seed,
 114                                                          FILE*                  fplog)
 115 {
 116     if (step < 0)
 117     {
 118         GMX_THROW(InvalidInputError(
 119                 "The step number is negative which is not supported by the AWH code."));
 120     }
 121
 122     GMX_RELEASE_ASSERT(!(params_.convolveForce && grid_.hasLambdaAxis()),
 123                        "When using AWH to sample an FEP lambda dimension the AWH potential cannot "
 124                        "be convolved.");
 125
 126     state_.setCoordValue(grid_, coordValue);
 127
 128     std::vector<double, AlignedAllocator<double>>& probWeightNeighbor = alignedTempWorkSpace_;
 129
 130     /* If the convolved force is needed or this is a sampling step,
 131      * the bias in the current neighborhood needs to be up-to-date
 132      * and the probablity weights need to be calculated.
 133      */
 134     const bool        isSampleCoordStep = params_.isSampleCoordStep(step);
 135     const bool        moveUmbrella      = (isSampleCoordStep || step == 0);
 136     double            convolvedBias     = 0;
 137     const CoordState& coordState        = state_.coordState();
 138
 139     if (params_.convolveForce || moveUmbrella || isSampleCoordStep)
 140     {
 141         if (params_.skipUpdates())
 142         {
 143             state_.doSkippedUpdatesInNeighborhood(params_, grid_);
 144         }
 145         convolvedBias = state_.updateProbabilityWeightsAndConvolvedBias(
 146                 dimParams_, grid_, moveUmbrella ? neighborLambdaEnergies : ArrayRef<const double>{}, &probWeightNeighbor);
 147
 148         if (isSampleCoordStep)
 149         {
 150             updateForceCorrelationGrid(probWeightNeighbor, neighborLambdaDhdl, t);
 151
 152             state_.sampleCoordAndPmf(dimParams_, grid_, probWeightNeighbor, convolvedBias);
 153         }
 154     }
 155
 156     /* Set the bias force and get the potential contribution from this bias.
 157      * The potential jump occurs at different times depending on how
 158      * the force is applied (and how the potential is normalized).
 159      * For the convolved force it happens when the bias is updated,
 160      * for the umbrella when the umbrella is moved.
 161      */
 162     *potentialJump = 0;
 163     double potential;
 164     if (params_.convolveForce)
 165     {
 166         state_.calcConvolvedForce(dimParams_,
 167                                   grid_,
 168                                   probWeightNeighbor,
 169                                   moveUmbrella ? neighborLambdaDhdl : ArrayRef<const double>{},
 170                                   tempForce_,
 171                                   biasForce_);
 172
 173         potential = -convolvedBias * params_.invBeta;
 174     }
 175     else
 176     {
 177         /* Umbrella force */
 178         GMX_RELEASE_ASSERT(state_.points()[coordState.umbrellaGridpoint()].inTargetRegion(),
 179                            "AWH bias grid point for the umbrella reference value is outside of the "
 180                            "target region.");
 181         potential = state_.calcUmbrellaForceAndPotential(
 182                 dimParams_,
 183                 grid_,
 184                 coordState.umbrellaGridpoint(),
 185                 moveUmbrella ? neighborLambdaDhdl : ArrayRef<const double>{},
 186                 biasForce_);
 187
 188         /* Moving the umbrella results in a force correction and
 189          * a new potential. The umbrella center is sampled as often as
 190          * the coordinate so we know the probability weights needed
 191          * for moving the umbrella are up-to-date.
 192          */
 193         if (moveUmbrella)
 194         {
 195             const bool onlySampleUmbrellaGridpoint = false;
 196             double     newPotential                = state_.moveUmbrella(dimParams_,
 197                                                       grid_,
 198                                                       probWeightNeighbor,
 199                                                       neighborLambdaDhdl,
 200                                                       biasForce_,
 201                                                       step,
 202                                                       seed,
 203                                                       params_.biasIndex,
 204                                                       onlySampleUmbrellaGridpoint);
 205             *potentialJump                         = newPotential - potential;
 206         }
 207     }
 208
 209     /* Update the free energy estimates and bias and other history dependent method parameters */
 210     if (params_.isUpdateFreeEnergyStep(step))
 211     {
 212         state_.updateFreeEnergyAndAddSamplesToHistogram(
 213                 dimParams_, grid_, params_, t, step, fplog, &updateList_);
 214
 215         if (params_.convolveForce)
 216         {
 217             /* The update results in a potential jump, so we need the new convolved potential. */
 218             double newPotential = -calcConvolvedBias(coordState.coordValue()) * params_.invBeta;
 219             *potentialJump      = newPotential - potential;
 220         }
 221     }
 222     /* If there is a lambda axis it is still controlled using an umbrella even if the force
 223      * is convolved in the other dimensions. */
 224     if (moveUmbrella && params_.convolveForce && grid_.hasLambdaAxis())
 225     {
 226         const bool onlySampleUmbrellaGridpoint = true;
 227         state_.moveUmbrella(dimParams_,
 228                             grid_,
 229                             probWeightNeighbor,
 230                             neighborLambdaDhdl,
 231                             biasForce_,
 232                             step,
 233                             seed,
 234                             params_.biasIndex,
 235                             onlySampleUmbrellaGridpoint);
 236     }
 237
 238     /* Return the potential. */
 239     *awhPotential = potential;
 240
 241     /* Check the sampled histograms and potentially warn user if something is suspicious */
 242     warnForHistogramAnomalies(t, step, fplog);
 243
 244     return biasForce_;
 245 }
 246
 247 /*! \brief
 248  * Count the total number of samples / sample weight over all grid points.
 249  *
 250  * \param[in] pointState  The state of the points in a bias.
 251  * \returns the total sample count.
 252  */
 253 static int64_t countSamples(ArrayRef<const PointState> pointState)
 254 {
 255     double numSamples = 0;
 256     for (const PointState& point : pointState)
 257     {
 258         numSamples += point.weightSumTot();
 259     }
 260
 261     return gmx::roundToInt64(numSamples);
 262 }
 263
 264 /*! \brief
 265  * Check if the state (loaded from checkpoint) and the run are consistent.
 266  *
 267  * When the state and the run setup are inconsistent, an exception is thrown.
 268  *
 269  * \param[in] params  The parameters of the bias.
 270  * \param[in] state   The state of the bias.
 271  */
 272 static void ensureStateAndRunConsistency(const BiasParams& params, const BiasState& state)
 273 {
 274     int64_t numSamples = countSamples(state.points());
 275     int64_t numUpdatesFromSamples =
 276             numSamples / (params.numSamplesUpdateFreeEnergy_ * params.numSharedUpdate);
 277     int64_t numUpdatesExpected = state.histogramSize().numUpdates();
 278     if (numUpdatesFromSamples != numUpdatesExpected)
 279     {
 280         std::string mesg = gmx::formatString(
 281                 "The number of AWH updates in the checkpoint file (%" PRId64
 282                 ") does not match the total number of AWH samples divided by the number of samples "
 283                 "per update for %d sharing AWH bias(es) (%" PRId64 "/%d=%" PRId64 ")",
 284                 numUpdatesExpected,
 285                 params.numSharedUpdate,
 286                 numSamples,
 287                 params.numSamplesUpdateFreeEnergy_ * params.numSharedUpdate,
 288                 numUpdatesFromSamples);
 289         mesg += " Maybe you changed AWH parameters.";
 290         /* Unfortunately we currently do not store the number of simulations
 291          * sharing the bias or the state to checkpoint. But we can hint at
 292          * a mismatch in the number of sharing simulations.
 293          */
 294         if (numUpdatesFromSamples % state.histogramSize().numUpdates() == 0)
 295         {
 296             mesg += gmx::formatString(
 297                     " Or the run you continued from used %" PRId64
 298                     " sharing simulations, whereas you now specified %d sharing simulations.",
 299                     numUpdatesFromSamples / state.histogramSize().numUpdates(),
 300                     params.numSharedUpdate);
 301         }
 302         GMX_THROW(InvalidInputError(mesg));
 303     }
 304 }
 305
 306 void Bias::restoreStateFromHistory(const AwhBiasHistory* biasHistory, const t_commrec* cr)
 307 {
 308     GMX_RELEASE_ASSERT(thisRankDoesIO_ == MASTER(cr),
 309                        "The master rank should do I/O, the other ranks should not");
 310
 311     if (MASTER(cr))
 312     {
 313         GMX_RELEASE_ASSERT(biasHistory != nullptr,
 314                            "On the master rank we need a valid history object to restore from");
 315         state_.restoreFromHistory(*biasHistory, grid_);
 316
 317         /* Ensure that the state is consistent with our current run setup,
 318          * since the user can have changed AWH parameters or the number
 319          * of simulations sharing the bias.
 320          */
 321         ensureStateAndRunConsistency(params_, state_);
 322
 323         if (forceCorrelationGrid_ != nullptr)
 324         {
 325             forceCorrelationGrid_->restoreStateFromHistory(biasHistory->forceCorrelationGrid);
 326         }
 327     }
 328
 329     if (PAR(cr))
 330     {
 331         state_.broadcast(cr);
 332     }
 333 }
 334
 335 void Bias::initHistoryFromState(AwhBiasHistory* biasHistory) const
 336 {
 337     GMX_RELEASE_ASSERT(biasHistory != nullptr, "Need a valid biasHistory");
 338
 339     state_.initHistoryFromState(biasHistory);
 340
 341     if (forceCorrelationGrid_ != nullptr)
 342     {
 343         biasHistory->forceCorrelationGrid = initCorrelationGridHistoryFromState(forceCorrelationGrid());
 344     }
 345 }
 346
 347 void Bias::updateHistory(AwhBiasHistory* biasHistory) const
 348 {
 349     GMX_RELEASE_ASSERT(biasHistory != nullptr, "Need a valid biasHistory");
 350
 351     state_.updateHistory(biasHistory, grid_);
 352
 353     if (forceCorrelationGrid_ != nullptr)
 354     {
 355         updateCorrelationGridHistory(&biasHistory->forceCorrelationGrid, forceCorrelationGrid());
 356     }
 357 }
 358
 359 Bias::Bias(int                            biasIndexInCollection,
 360            const AwhParams&               awhParams,
 361            const AwhBiasParams&           awhBiasParams,
 362            ArrayRef<const DimParams>      dimParamsInit,
 363            double                         beta,
 364            double                         mdTimeStep,
 365            const BiasSharing*             biasSharing,
 366            const std::string&             biasInitFilename,
 367            ThisRankWillDoIO               thisRankWillDoIO,
 368            BiasParams::DisableUpdateSkips disableUpdateSkips) :
 369     dimParams_(dimParamsInit.begin(), dimParamsInit.end()),
 370     grid_(dimParamsInit, awhBiasParams.dimParams()),
 371     params_(awhParams,
 372             awhBiasParams,
 373             dimParams_,
 374             beta,
 375             mdTimeStep,
 376             disableUpdateSkips,
 377             biasSharing ? biasSharing->numSharingSimulations(biasIndexInCollection) : 1,
 378             grid_.axis(),
 379             biasIndexInCollection),
 380     state_(awhBiasParams, params_.initialHistogramSize, dimParams_, grid_, biasSharing),
 381     thisRankDoesIO_(thisRankWillDoIO == ThisRankWillDoIO::Yes),
 382     biasForce_(ndim()),
 383     tempForce_(ndim()),
 384     numWarningsIssued_(0)
 385 {
 386     /* For a global update updateList covers all points, so reserve that */
 387     updateList_.reserve(grid_.numPoints());
 388
 389     state_.initGridPointState(
 390             awhBiasParams, dimParams_, grid_, params_, biasInitFilename, awhParams.numBias());
 391
 392     if (thisRankDoesIO_)
 393     {
 394         /* Set up the force correlation object. */
 395
 396         /* We let the correlation init function set its parameters
 397          * to something useful for now.
 398          */
 399         double blockLength = 0;
 400         /* Construct the force correlation object. */
 401         forceCorrelationGrid_ =
 402                 std::make_unique<CorrelationGrid>(state_.points().size(),
 403                                                   ndim(),
 404                                                   blockLength,
 405                                                   CorrelationGrid::BlockLengthMeasure::Time,
 406                                                   awhParams.nstSampleCoord() * mdTimeStep);
 407
 408         writer_ = std::make_unique<BiasWriter>(*this);
 409     }
 410 }
 411
 412 void Bias::printInitializationToLog(FILE* fplog) const
 413 {
 414     if (fplog != nullptr && forceCorrelationGrid_ != nullptr)
 415     {
 416         std::string prefix = gmx::formatString("\nawh%d:", params_.biasIndex + 1);
 417
 418         fprintf(fplog,
 419                 "%s initial force correlation block length = %g %s"
 420                 "%s force correlation number of blocks = %d",
 421                 prefix.c_str(),
 422                 forceCorrelationGrid().getBlockLength(),
 423                 forceCorrelationGrid().blockLengthMeasure == CorrelationGrid::BlockLengthMeasure::Weight
 424                         ? ""
 425                         : "ps",
 426                 prefix.c_str(),
 427                 forceCorrelationGrid().getNumBlocks());
 428     }
 429 }
 430
 431 void Bias::updateForceCorrelationGrid(gmx::ArrayRef<const double> probWeightNeighbor,
 432                                       ArrayRef<const double>      neighborLambdaDhdl,
 433                                       double                      t)
 434 {
 435     if (forceCorrelationGrid_ == nullptr)
 436     {
 437         return;
 438     }
 439
 440     const std::vector<int>& neighbor = grid_.point(state_.coordState().gridpointIndex()).neighbor;
 441
 442     gmx::ArrayRef<double> forceFromNeighbor = tempForce_;
 443     for (size_t n = 0; n < neighbor.size(); n++)
 444     {
 445         double weightNeighbor = probWeightNeighbor[n];
 446         int    indexNeighbor  = neighbor[n];
 447
 448         /* Add the force data of this neighbor point. Note: the sum of these forces is the convolved force.
 449
 450            We actually add the force normalized by beta which has the units of 1/length. This means that the
 451            resulting correlation time integral is directly in units of friction time/length^2 which is really what
 452            we're interested in. */
 453         state_.calcUmbrellaForceAndPotential(
 454                 dimParams_, grid_, indexNeighbor, neighborLambdaDhdl, forceFromNeighbor);
 455
 456         /* Note: we might want to give a whole list of data to add instead and have this loop in the data adding function */
 457         forceCorrelationGrid_->addData(indexNeighbor, weightNeighbor, forceFromNeighbor, t);
 458     }
 459 }
 460
 461 /* Return the number of data blocks that have been prepared for writing. */
 462 int Bias::numEnergySubblocksToWrite() const
 463 {
 464     GMX_RELEASE_ASSERT(writer_ != nullptr, "Should only request data from an initialized writer");
 465
 466     return writer_->numBlocks();
 467 }
 468
 469 /* Write bias data blocks to energy subblocks. */
 470 int Bias::writeToEnergySubblocks(t_enxsubblock* subblock) const
 471 {
 472     GMX_RELEASE_ASSERT(writer_ != nullptr, "Should only request data from an initialized writer");
 473
 474     return writer_->writeToEnergySubblocks(*this, subblock);
 475 }
 476
 477 bool Bias::isSampleCoordStep(const int64_t step) const
 478 {
 479     return params_.isSampleCoordStep(step);
 480 }
 481
 482
 483 } // namespace gmx