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