src/gromacs/mdtypes/threaded_force_buffer.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 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 /*! \internal \file
  36  *
  37  * \brief This file defines the implementation of ThreadForceBuffer and ThreadedForceBuffer.
  38  *
  39  * \author Berk Hess <hess@kth.se>
  40  *
  41  * \ingroup module_mdtypes
  42  */
  43 #include "gmxpre.h"
  44
  45 #include "threaded_force_buffer.h"
  46
  47 #include "gromacs/math/vec.h"
  48 #include "gromacs/mdtypes/forceoutput.h"
  49 #include "gromacs/pbcutil/ishift.h"
  50 #include "gromacs/utility/alignedallocator.h"
  51 #include "gromacs/utility/exceptions.h"
  52 #include "gromacs/utility/fatalerror.h"
  53
  54 namespace gmx
  55 {
  56
  57 /*! \brief The max thread number is arbitrary, we used a fixed number
  58  * to avoid memory management.  Using more than 16 threads is probably
  59  * never useful performance wise. */
  60 static constexpr int s_maxNumThreadsForReduction = 256;
  61
  62 template<typename ForceBufferElementType>
  63 ThreadForceBuffer<ForceBufferElementType>::ThreadForceBuffer(const int  threadIndex,
  64                                                              const bool useEnergyTerms,
  65                                                              const int  numEnergyGroups) :
  66     threadIndex_(threadIndex), shiftForces_(c_numShiftVectors), groupPairEnergies_(numEnergyGroups)
  67 {
  68     if (useEnergyTerms)
  69     {
  70         energyTerms_.resize(F_NRE);
  71     }
  72 }
  73
  74 template<typename ForceBufferElementType>
  75 void ThreadForceBuffer<ForceBufferElementType>::clearForcesAndEnergies()
  76 {
  77     constexpr int c_numComponents = sizeof(ForceBufferElementType) / sizeof(real);
  78
  79     for (int atomIndex : usedBlockIndices_)
  80     {
  81         const int bufferIndexBegin = atomIndex * s_reductionBlockSize * c_numComponents;
  82         const int bufferIndexEnd   = bufferIndexBegin + s_reductionBlockSize * c_numComponents;
  83         std::fill(forceBuffer_.begin() + bufferIndexBegin, forceBuffer_.begin() + bufferIndexEnd, 0.0_real);
  84     }
  85
  86     const RVec zeroVec = { 0.0_real, 0.0_real, 0.0_real };
  87     std::fill(shiftForces_.begin(), shiftForces_.end(), zeroVec);
  88
  89     std::fill(energyTerms_.begin(), energyTerms_.end(), 0.0_real);
  90
  91     for (int i = 0; i < static_cast<int>(NonBondedEnergyTerms::Count); i++)
  92     {
  93         for (int j = 0; j < groupPairEnergies_.nener; j++)
  94         {
  95             groupPairEnergies_.energyGroupPairTerms[i][j] = 0.0_real;
  96         }
  97     }
  98     for (auto i : keysOf(dvdl_))
  99     {
 100         dvdl_[i] = 0;
 101     }
 102 }
 103
 104 template<typename ForceBufferElementType>
 105 void ThreadForceBuffer<ForceBufferElementType>::resizeBufferAndClearMask(const int numAtoms)
 106 {
 107     numAtoms_ = numAtoms;
 108
 109     const int numBlocks = (numAtoms + s_reductionBlockSize - 1) >> s_numReductionBlockBits;
 110
 111     reductionMask_.resize(numBlocks);
 112     forceBuffer_.resize(numBlocks * s_reductionBlockSize * sizeof(ForceBufferElementType) / sizeof(real));
 113
 114     for (gmx_bitmask_t& mask : reductionMask_)
 115     {
 116         bitmask_clear(&mask);
 117     }
 118 }
 119
 120 template<typename ForceBufferElementType>
 121 void ThreadForceBuffer<ForceBufferElementType>::processMask()
 122 {
 123     // Now we are done setting the masks, generate the new list of used blocks
 124     usedBlockIndices_.clear();
 125     for (int b = 0; b < ssize(reductionMask_); b++)
 126     {
 127         if (bitmask_is_set(reductionMask_[b], threadIndex_))
 128         {
 129             usedBlockIndices_.push_back(b);
 130         }
 131     }
 132 }
 133
 134 namespace
 135 {
 136
 137 //! \brief Reduce thread-local force buffers into \p force (does not reduce shift forces)
 138 template<typename ForceBufferElementType>
 139 void reduceThreadForceBuffers(ArrayRef<gmx::RVec> force,
 140                               ArrayRef<std::unique_ptr<ThreadForceBuffer<ForceBufferElementType>>> threadForceBuffers,
 141                               ArrayRef<const gmx_bitmask_t> masks,
 142                               ArrayRef<const int>           usedBlockIndices)
 143 {
 144     const int numBuffers = threadForceBuffers.size();
 145     GMX_ASSERT(numBuffers <= s_maxNumThreadsForReduction,
 146                "There is a limit on the number of buffers we can use for reduction");
 147
 148     const int numAtoms = threadForceBuffers[0]->size();
 149
 150     rvec* gmx_restrict f = as_rvec_array(force.data());
 151
 152     /* This reduction can run on any number of threads, independently of the number of buffers.
 153      * But if the number of threads matches the number of buffers, which it currently does,
 154      * the uniform distribution of the touched blocks over nthreads will
 155      * match the distribution of bonded over threads well in most cases,
 156      * which means that threads mostly reduce their own data which increases
 157      * the number of cache hits.
 158      * Additionally, we should always use the same number of threads in parallel
 159      * regions in OpenMP, otherwise the performance will degrade significantly.
 160      */
 161     const int gmx_unused numThreadsForReduction = threadForceBuffers.size();
 162 #pragma omp parallel for num_threads(numThreadsForReduction) schedule(static)
 163     for (int b = 0; b < usedBlockIndices.ssize(); b++)
 164     {
 165         try
 166         {
 167             // Reduce the buffers that contribute to this block
 168             ForceBufferElementType* fp[s_maxNumThreadsForReduction];
 169
 170             const int blockIndex = usedBlockIndices[b];
 171
 172             // Make a list of threads that have this block index set in the mask
 173             int numContributingBuffers = 0;
 174             for (int ft = 0; ft < numBuffers; ft++)
 175             {
 176                 if (bitmask_is_set(masks[blockIndex], ft))
 177                 {
 178                     fp[numContributingBuffers++] = threadForceBuffers[ft]->forceBuffer();
 179                 }
 180             }
 181             if (numContributingBuffers > 0)
 182             {
 183                 // Reduce the selected buffers
 184                 int a0 = blockIndex * ThreadForceBuffer<ForceBufferElementType>::s_reductionBlockSize;
 185                 int a1 = (blockIndex + 1) * ThreadForceBuffer<ForceBufferElementType>::s_reductionBlockSize;
 186                 // Note: It would be nice if we could pad f to avoid this min()
 187                 a1 = std::min(a1, numAtoms);
 188                 if (numContributingBuffers == 1)
 189                 {
 190                     // Avoid double loop for the case of a single buffer
 191                     for (int a = a0; a < a1; a++)
 192                     {
 193                         rvec_inc(f[a], fp[0][a]);
 194                     }
 195                 }
 196                 else
 197                 {
 198                     for (int a = a0; a < a1; a++)
 199                     {
 200                         for (int fb = 0; fb < numContributingBuffers; fb++)
 201                         {
 202                             rvec_inc(f[a], fp[fb][a]);
 203                         }
 204                     }
 205                 }
 206             }
 207         }
 208         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 209     }
 210 }
 211
 212 } // namespace
 213
 214 template<typename ForceBufferElementType>
 215 ThreadedForceBuffer<ForceBufferElementType>::ThreadedForceBuffer(const int  numThreads,
 216                                                                  const bool useEnergyTerms,
 217                                                                  const int  numEnergyGroups) :
 218     useEnergyTerms_(useEnergyTerms)
 219 {
 220     threadForceBuffers_.resize(numThreads);
 221 #pragma omp parallel for num_threads(numThreads) schedule(static)
 222     for (int t = 0; t < numThreads; t++)
 223     {
 224         try
 225         {
 226             /* Note that thread 0 uses the global fshift and energy arrays,
 227              * but to keep the code simple, we initialize all data here.
 228              */
 229             threadForceBuffers_[t] = std::make_unique<ThreadForceBuffer<ForceBufferElementType>>(
 230                     t, useEnergyTerms_, numEnergyGroups);
 231         }
 232         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 233     }
 234 }
 235
 236 template<typename ForceBufferElementType>
 237 void ThreadedForceBuffer<ForceBufferElementType>::setupReduction()
 238 {
 239     const int numBuffers = threadForceBuffers_.size();
 240
 241     const int numAtoms = threadForceBuffers_[0]->size();
 242
 243     const int totalNumBlocks =
 244             (numAtoms + ThreadForceBuffer<ForceBufferElementType>::s_reductionBlockSize - 1)
 245             >> ThreadForceBuffer<ForceBufferElementType>::s_numReductionBlockBits;
 246
 247     // Check that all thread buffers have matching sizes
 248     for (const auto& threadForceBuffer : threadForceBuffers_)
 249     {
 250         GMX_RELEASE_ASSERT(threadForceBuffer->size() == numAtoms,
 251                            "All buffers should have the same size");
 252         GMX_RELEASE_ASSERT(threadForceBuffer->reductionMask().ssize() == totalNumBlocks,
 253                            "The block count should match");
 254     }
 255
 256     /* Reduce the masks over the threads and determine which blocks
 257      * we need to reduce over.
 258      */
 259     reductionMask_.resize(totalNumBlocks);
 260
 261     usedBlockIndices_.clear();
 262     int numBlocksUsed = 0;
 263     for (int b = 0; b < totalNumBlocks; b++)
 264     {
 265         gmx_bitmask_t& mask = reductionMask_[b];
 266
 267         /* Generate the union over the threads of the bitmask */
 268         bitmask_clear(&mask);
 269         for (int t = 0; t < numBuffers; t++)
 270         {
 271             bitmask_union(&mask, threadForceBuffers_[t]->reductionMask()[b]);
 272         }
 273         if (!bitmask_is_zero(mask))
 274         {
 275             usedBlockIndices_.push_back(b);
 276         }
 277
 278         if (debug)
 279         {
 280             int c = 0;
 281             for (int t = 0; t < numBuffers; t++)
 282             {
 283                 if (bitmask_is_set(mask, t))
 284                 {
 285                     c++;
 286                 }
 287             }
 288             numBlocksUsed += c;
 289
 290             if (gmx_debug_at)
 291             {
 292                 fprintf(debug, "block %d flags %s count %d\n", b, to_hex_string(mask).c_str(), c);
 293             }
 294         }
 295     }
 296     if (debug)
 297     {
 298         fprintf(debug,
 299                 "Number of %d atom blocks to reduce: %d\n",
 300                 ThreadForceBuffer<ForceBufferElementType>::s_reductionBlockSize,
 301                 int(ssize(usedBlockIndices_)));
 302         fprintf(debug,
 303                 "Reduction density %.2f for touched blocks only %.2f\n",
 304                 numBlocksUsed * ThreadForceBuffer<ForceBufferElementType>::s_reductionBlockSize
 305                         / static_cast<double>(numAtoms),
 306                 numBlocksUsed / static_cast<double>(ssize(usedBlockIndices_)));
 307     }
 308 }
 309
 310 template<typename ForceBufferElementType>
 311 void ThreadedForceBuffer<ForceBufferElementType>::reduce(gmx::ForceWithShiftForces* forceWithShiftForces,
 312                                                          real*                      ener,
 313                                                          gmx_grppairener_t*         grpp,
 314                                                          gmx::ArrayRef<real>        dvdl,
 315                                                          const gmx::StepWorkload& stepWork,
 316                                                          const int reductionBeginIndex)
 317 {
 318     if (stepWork.computeForces && !usedBlockIndices_.empty())
 319     {
 320         /* Reduce the force buffer */
 321         GMX_ASSERT(forceWithShiftForces, "Need a valid force buffer for reduction");
 322
 323         reduceThreadForceBuffers<ForceBufferElementType>(
 324                 forceWithShiftForces->force(), threadForceBuffers_, reductionMask_, usedBlockIndices_);
 325     }
 326
 327     const int numBuffers = numThreadBuffers();
 328
 329     /* When necessary, reduce energy and virial using one thread only */
 330     if ((stepWork.computeEnergy || stepWork.computeVirial || stepWork.computeDhdl)
 331         && numBuffers > reductionBeginIndex)
 332     {
 333         gmx::ArrayRef<const std::unique_ptr<ThreadForceBuffer<ForceBufferElementType>>> f_t =
 334                 threadForceBuffers_;
 335
 336         if (stepWork.computeVirial)
 337         {
 338             GMX_ASSERT(forceWithShiftForces, "Need a valid force buffer for reduction");
 339
 340             rvec* gmx_restrict fshift = as_rvec_array(forceWithShiftForces->shiftForces().data());
 341
 342             for (int i = 0; i < gmx::c_numShiftVectors; i++)
 343             {
 344                 for (int t = reductionBeginIndex; t < numBuffers; t++)
 345                 {
 346                     rvec_inc(fshift[i], f_t[t]->shiftForces()[i]);
 347                 }
 348             }
 349         }
 350         if (stepWork.computeEnergy && useEnergyTerms_)
 351         {
 352             GMX_ASSERT(ener, "Need a valid energy buffer for reduction");
 353
 354             for (int i = 0; i < F_NRE; i++)
 355             {
 356                 for (int t = reductionBeginIndex; t < numBuffers; t++)
 357                 {
 358                     ener[i] += f_t[t]->energyTerms()[i];
 359                 }
 360             }
 361         }
 362
 363         if (stepWork.computeEnergy)
 364         {
 365             GMX_ASSERT(grpp, "Need a valid group pair energy buffer for reduction");
 366
 367             for (int i = 0; i < static_cast<int>(NonBondedEnergyTerms::Count); i++)
 368             {
 369                 for (int j = 0; j < f_t[0]->groupPairEnergies().nener; j++)
 370                 {
 371                     for (int t = reductionBeginIndex; t < numBuffers; t++)
 372                     {
 373                         grpp->energyGroupPairTerms[i][j] +=
 374                                 f_t[t]->groupPairEnergies().energyGroupPairTerms[i][j];
 375                     }
 376                 }
 377             }
 378         }
 379         if (stepWork.computeDhdl)
 380         {
 381             GMX_ASSERT(!dvdl.empty(), "Need a valid dV/dl buffer for reduction");
 382
 383             for (auto i : keysOf(f_t[0]->dvdl()))
 384             {
 385
 386                 for (int t = reductionBeginIndex; t < numBuffers; t++)
 387                 {
 388                     dvdl[static_cast<int>(i)] += f_t[t]->dvdl()[i];
 389                 }
 390             }
 391         }
 392     }
 393 }
 394
 395 template class ThreadForceBuffer<RVec>;
 396 template class ThreadedForceBuffer<RVec>;
 397
 398 template class ThreadForceBuffer<rvec4>;
 399 template class ThreadedForceBuffer<rvec4>;
 400
 401 } // namespace gmx