src/gromacs/listed_forces/gpubonded_impl.cu

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 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 /*! \internal \file
  36  *
  37  * \brief Implements GPU bonded lists for CUDA
  38  *
  39  * \author Berk Hess <hess@kth.se>
  40  * \author Szilárd Páll <pall.szilard@gmail.com>
  41  * \author Mark Abraham <mark.j.abraham@gmail.com>
  42  * \author Magnus Lundborg <lundborg.magnus@gmail.com>
  43  *
  44  * \ingroup module_listed_forces
  45  */
  46
  47 #include "gmxpre.h"
  48
  49 #include "gpubonded_impl.h"
  50
  51 #include "gromacs/gpu_utils/cuda_arch_utils.cuh"
  52 #include "gromacs/gpu_utils/cudautils.cuh"
  53 #include "gromacs/gpu_utils/device_context.h"
  54 #include "gromacs/gpu_utils/device_stream.h"
  55 #include "gromacs/gpu_utils/devicebuffer.h"
  56 #include "gromacs/gpu_utils/typecasts.cuh"
  57 #include "gromacs/mdtypes/enerdata.h"
  58 #include "gromacs/timing/wallcycle.h"
  59 #include "gromacs/topology/forcefieldparameters.h"
  60
  61 struct t_forcerec;
  62
  63 namespace gmx
  64 {
  65 // Number of CUDA threads in a block
  66 constexpr static int c_threadsPerBlock = 256;
  67
  68 // ---- GpuBonded::Impl
  69
  70 GpuBonded::Impl::Impl(const gmx_ffparams_t& ffparams,
  71                       const float           electrostaticsScaleFactor,
  72                       const DeviceContext&  deviceContext,
  73                       const DeviceStream&   deviceStream,
  74                       gmx_wallcycle*        wcycle) :
  75     deviceContext_(deviceContext),
  76     deviceStream_(deviceStream)
  77 {
  78     GMX_RELEASE_ASSERT(deviceStream.isValid(),
  79                        "Can't run GPU version of bonded forces in stream that is not valid.");
  80
  81     static_assert(
  82             c_threadsPerBlock >= c_numShiftVectors,
  83             "Threads per block in GPU bonded must be >= c_numShiftVectors for the virial kernel "
  84             "(calcVir=true)");
  85
  86     wcycle_ = wcycle;
  87
  88     allocateDeviceBuffer(&d_forceParams_, ffparams.numTypes(), deviceContext_);
  89     // This could be an async transfer (if the source is pinned), so
  90     // long as it uses the same stream as the kernels and we are happy
  91     // to consume additional pinned pages.
  92     copyToDeviceBuffer(&d_forceParams_,
  93                        ffparams.iparams.data(),
  94                        0,
  95                        ffparams.numTypes(),
  96                        deviceStream_,
  97                        GpuApiCallBehavior::Sync,
  98                        nullptr);
  99     vTot_.resize(F_NRE);
 100     allocateDeviceBuffer(&d_vTot_, F_NRE, deviceContext_);
 101     clearDeviceBufferAsync(&d_vTot_, 0, F_NRE, deviceStream_);
 102
 103     kernelParams_.electrostaticsScaleFactor = electrostaticsScaleFactor;
 104     kernelParams_.d_forceParams             = d_forceParams_;
 105     kernelParams_.d_xq                      = d_xq_;
 106     kernelParams_.d_f                       = d_f_;
 107     kernelParams_.d_fShift                  = d_fShift_;
 108     kernelParams_.d_vTot                    = d_vTot_;
 109     for (int i = 0; i < numFTypesOnGpu; i++)
 110     {
 111         kernelParams_.d_iatoms[i]        = nullptr;
 112         kernelParams_.fTypeRangeStart[i] = 0;
 113         kernelParams_.fTypeRangeEnd[i]   = -1;
 114     }
 115
 116     int fTypeRangeEnd = kernelParams_.fTypeRangeEnd[numFTypesOnGpu - 1];
 117
 118     kernelLaunchConfig_.blockSize[0] = c_threadsPerBlock;
 119     kernelLaunchConfig_.blockSize[1] = 1;
 120     kernelLaunchConfig_.blockSize[2] = 1;
 121     kernelLaunchConfig_.gridSize[0]  = (fTypeRangeEnd + c_threadsPerBlock) / c_threadsPerBlock;
 122     kernelLaunchConfig_.gridSize[1]  = 1;
 123     kernelLaunchConfig_.gridSize[2]  = 1;
 124     kernelLaunchConfig_.sharedMemorySize =
 125             c_numShiftVectors * sizeof(float3) + (c_threadsPerBlock / warp_size) * 3 * sizeof(float);
 126 }
 127
 128 GpuBonded::Impl::~Impl()
 129 {
 130     for (int fType : fTypesOnGpu)
 131     {
 132         if (d_iLists_[fType].iatoms)
 133         {
 134             freeDeviceBuffer(&d_iLists_[fType].iatoms);
 135             d_iLists_[fType].iatoms = nullptr;
 136         }
 137     }
 138
 139     freeDeviceBuffer(&d_forceParams_);
 140     freeDeviceBuffer(&d_vTot_);
 141 }
 142
 143 //! Return whether function type \p fType in \p idef has perturbed interactions
 144 static bool fTypeHasPerturbedEntries(const InteractionDefinitions& idef, int fType)
 145 {
 146     GMX_ASSERT(idef.ilsort == ilsortNO_FE || idef.ilsort == ilsortFE_SORTED,
 147                "Perturbed interations should be sorted here");
 148
 149     const InteractionList& ilist = idef.il[fType];
 150
 151     return (idef.ilsort != ilsortNO_FE && idef.numNonperturbedInteractions[fType] != ilist.size());
 152 }
 153
 154 //! Converts \p src with atom indices in state order to \p dest in nbnxn order
 155 static void convertIlistToNbnxnOrder(const InteractionList& src,
 156                                      HostInteractionList*   dest,
 157                                      int                    numAtomsPerInteraction,
 158                                      ArrayRef<const int>    nbnxnAtomOrder)
 159 {
 160     GMX_ASSERT(src.size() == 0 || !nbnxnAtomOrder.empty(), "We need the nbnxn atom order");
 161
 162     dest->iatoms.resize(src.size());
 163
 164     // TODO use OpenMP to parallelise this loop
 165     for (int i = 0; i < src.size(); i += 1 + numAtomsPerInteraction)
 166     {
 167         dest->iatoms[i] = src.iatoms[i];
 168         for (int a = 0; a < numAtomsPerInteraction; a++)
 169         {
 170             dest->iatoms[i + 1 + a] = nbnxnAtomOrder[src.iatoms[i + 1 + a]];
 171         }
 172     }
 173 }
 174
 175 //! Returns \p input rounded up to the closest multiple of \p factor.
 176 static inline int roundUpToFactor(const int input, const int factor)
 177 {
 178     GMX_ASSERT(factor > 0, "The factor to round up to must be > 0.");
 179
 180     int remainder = input % factor;
 181
 182     if (remainder == 0)
 183     {
 184         return (input);
 185     }
 186     return (input + (factor - remainder));
 187 }
 188
 189 // TODO Consider whether this function should be a factory method that
 190 // makes an object that is the only one capable of the device
 191 // operations needed for the lifetime of an interaction list. This
 192 // would be harder to misuse than GpuBonded, and exchange the problem
 193 // of naming this method for the problem of what to name the
 194 // BondedDeviceInteractionListHandler type.
 195
 196 /*! Divides bonded interactions over threads and GPU.
 197  *  The bonded interactions are assigned by interaction type to GPU threads. The intereaction
 198  *  types are assigned in blocks sized as <warp_size>. The beginning and end (thread index) of each
 199  *  interaction type are stored in kernelParams_. Pointers to the relevant data structures on the
 200  *  GPU are also stored in kernelParams_.
 201  *
 202  * \todo Use DeviceBuffer for the d_xqPtr.
 203  */
 204 void GpuBonded::Impl::updateInteractionListsAndDeviceBuffers(ArrayRef<const int> nbnxnAtomOrder,
 205                                                              const InteractionDefinitions& idef,
 206                                                              void*                         d_xqPtr,
 207                                                              DeviceBuffer<RVec>            d_fPtr,
 208                                                              DeviceBuffer<RVec> d_fShiftPtr)
 209 {
 210     // TODO wallcycle sub start
 211     haveInteractions_ = false;
 212     int fTypesCounter = 0;
 213
 214     for (int fType : fTypesOnGpu)
 215     {
 216         auto& iList = iLists_[fType];
 217
 218         /* Perturbation is not implemented in the GPU bonded kernels.
 219          * But instead of doing all interactions on the CPU, we can
 220          * still easily handle the types that have no perturbed
 221          * interactions on the GPU. */
 222         if (!idef.il[fType].empty() && !fTypeHasPerturbedEntries(idef, fType))
 223         {
 224             haveInteractions_ = true;
 225
 226             convertIlistToNbnxnOrder(idef.il[fType], &iList, NRAL(fType), nbnxnAtomOrder);
 227         }
 228         else
 229         {
 230             iList.iatoms.clear();
 231         }
 232
 233         // Update the device if necessary. This can leave some
 234         // allocations on the device when the host size decreases to
 235         // zero, which is OK, since we deallocate everything at the
 236         // end.
 237         if (iList.size() > 0)
 238         {
 239             t_ilist& d_iList = d_iLists_[fType];
 240
 241             reallocateDeviceBuffer(
 242                     &d_iList.iatoms, iList.size(), &d_iList.nr, &d_iList.nalloc, deviceContext_);
 243
 244             copyToDeviceBuffer(&d_iList.iatoms,
 245                                iList.iatoms.data(),
 246                                0,
 247                                iList.size(),
 248                                deviceStream_,
 249                                GpuApiCallBehavior::Async,
 250                                nullptr);
 251         }
 252         kernelParams_.fTypesOnGpu[fTypesCounter]    = fType;
 253         kernelParams_.numFTypeIAtoms[fTypesCounter] = iList.size();
 254         int numBonds = iList.size() / (interaction_function[fType].nratoms + 1);
 255         kernelParams_.numFTypeBonds[fTypesCounter] = numBonds;
 256         kernelParams_.d_iatoms[fTypesCounter]      = d_iLists_[fType].iatoms;
 257         if (fTypesCounter == 0)
 258         {
 259             kernelParams_.fTypeRangeStart[fTypesCounter] = 0;
 260         }
 261         else
 262         {
 263             kernelParams_.fTypeRangeStart[fTypesCounter] =
 264                     kernelParams_.fTypeRangeEnd[fTypesCounter - 1] + 1;
 265         }
 266         kernelParams_.fTypeRangeEnd[fTypesCounter] = kernelParams_.fTypeRangeStart[fTypesCounter]
 267                                                      + roundUpToFactor(numBonds, warp_size) - 1;
 268
 269         GMX_ASSERT(numBonds > 0
 270                            || kernelParams_.fTypeRangeEnd[fTypesCounter]
 271                                       <= kernelParams_.fTypeRangeStart[fTypesCounter],
 272                    "Invalid GPU listed forces setup. numBonds must be > 0 if there are threads "
 273                    "allocated to do work on that interaction function type.");
 274         GMX_ASSERT(kernelParams_.fTypeRangeStart[fTypesCounter] % warp_size == 0
 275                            && (kernelParams_.fTypeRangeEnd[fTypesCounter] + 1) % warp_size == 0,
 276                    "The bonded interactions must be assigned to the GPU in blocks of warp size.");
 277
 278         fTypesCounter++;
 279     }
 280
 281     int fTypeRangeEnd               = kernelParams_.fTypeRangeEnd[numFTypesOnGpu - 1];
 282     kernelLaunchConfig_.gridSize[0] = (fTypeRangeEnd + c_threadsPerBlock) / c_threadsPerBlock;
 283
 284     d_xq_     = static_cast<float4*>(d_xqPtr);
 285     d_f_      = asFloat3(d_fPtr);
 286     d_fShift_ = asFloat3(d_fShiftPtr);
 287
 288     kernelParams_.d_xq          = d_xq_;
 289     kernelParams_.d_f           = d_f_;
 290     kernelParams_.d_fShift      = d_fShift_;
 291     kernelParams_.d_forceParams = d_forceParams_;
 292     kernelParams_.d_vTot        = d_vTot_;
 293
 294     // TODO wallcycle sub stop
 295 }
 296
 297 void GpuBonded::Impl::setPbc(PbcType pbcType, const matrix box, bool canMoleculeSpanPbc)
 298 {
 299     PbcAiuc pbcAiuc;
 300     setPbcAiuc(canMoleculeSpanPbc ? numPbcDimensions(pbcType) : 0, box, &pbcAiuc);
 301     kernelParams_.pbcAiuc = pbcAiuc;
 302 }
 303
 304 bool GpuBonded::Impl::haveInteractions() const
 305 {
 306     return haveInteractions_;
 307 }
 308
 309 void GpuBonded::Impl::launchEnergyTransfer()
 310 {
 311     GMX_ASSERT(haveInteractions_,
 312                "No GPU bonded interactions, so no energies will be computed, so transfer should "
 313                "not be called");
 314     wallcycle_sub_start_nocount(wcycle_, ewcsLAUNCH_GPU_BONDED);
 315     // TODO add conditional on whether there has been any compute (and make sure host buffer doesn't contain garbage)
 316     float* h_vTot = vTot_.data();
 317     copyFromDeviceBuffer(h_vTot, &d_vTot_, 0, F_NRE, deviceStream_, GpuApiCallBehavior::Async, nullptr);
 318     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_GPU_BONDED);
 319 }
 320
 321 void GpuBonded::Impl::waitAccumulateEnergyTerms(gmx_enerdata_t* enerd)
 322 {
 323     GMX_ASSERT(haveInteractions_,
 324                "No GPU bonded interactions, so no energies will be computed or transferred, so "
 325                "accumulation should not occur");
 326
 327     wallcycle_start(wcycle_, ewcWAIT_GPU_BONDED);
 328     cudaError_t stat = cudaStreamSynchronize(deviceStream_.stream());
 329     CU_RET_ERR(stat, "D2H transfer of bonded energies failed");
 330     wallcycle_stop(wcycle_, ewcWAIT_GPU_BONDED);
 331
 332     for (int fType : fTypesOnGpu)
 333     {
 334         if (fType != F_LJ14 && fType != F_COUL14)
 335         {
 336             enerd->term[fType] += vTot_[fType];
 337         }
 338     }
 339
 340     // Note: We do not support energy groups here
 341     gmx_grppairener_t* grppener = &enerd->grpp;
 342     GMX_RELEASE_ASSERT(grppener->nener == 1, "No energy group support for bondeds on the GPU");
 343     grppener->energyGroupPairTerms[NonBondedEnergyTerms::LJ14][0] += vTot_[F_LJ14];
 344     grppener->energyGroupPairTerms[NonBondedEnergyTerms::Coulomb14][0] += vTot_[F_COUL14];
 345 }
 346
 347 void GpuBonded::Impl::clearEnergies()
 348 {
 349     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 350     wallcycle_sub_start_nocount(wcycle_, ewcsLAUNCH_GPU_BONDED);
 351     clearDeviceBufferAsync(&d_vTot_, 0, F_NRE, deviceStream_);
 352     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_GPU_BONDED);
 353     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 354 }
 355
 356 // ---- GpuBonded
 357
 358 GpuBonded::GpuBonded(const gmx_ffparams_t& ffparams,
 359                      const float           electrostaticsScaleFactor,
 360                      const DeviceContext&  deviceContext,
 361                      const DeviceStream&   deviceStream,
 362                      gmx_wallcycle*        wcycle) :
 363     impl_(new Impl(ffparams, electrostaticsScaleFactor, deviceContext, deviceStream, wcycle))
 364 {
 365 }
 366
 367 GpuBonded::~GpuBonded() = default;
 368
 369 void GpuBonded::updateInteractionListsAndDeviceBuffers(ArrayRef<const int>           nbnxnAtomOrder,
 370                                                        const InteractionDefinitions& idef,
 371                                                        void*                         d_xq,
 372                                                        DeviceBuffer<RVec>            d_f,
 373                                                        DeviceBuffer<RVec>            d_fShift)
 374 {
 375     impl_->updateInteractionListsAndDeviceBuffers(nbnxnAtomOrder, idef, d_xq, d_f, d_fShift);
 376 }
 377
 378 void GpuBonded::setPbc(PbcType pbcType, const matrix box, bool canMoleculeSpanPbc)
 379 {
 380     impl_->setPbc(pbcType, box, canMoleculeSpanPbc);
 381 }
 382
 383 bool GpuBonded::haveInteractions() const
 384 {
 385     return impl_->haveInteractions();
 386 }
 387
 388 void GpuBonded::setPbcAndlaunchKernel(PbcType                  pbcType,
 389                                       const matrix             box,
 390                                       bool                     canMoleculeSpanPbc,
 391                                       const gmx::StepWorkload& stepWork)
 392 {
 393     setPbc(pbcType, box, canMoleculeSpanPbc);
 394     launchKernel(stepWork);
 395 }
 396
 397 void GpuBonded::launchEnergyTransfer()
 398 {
 399     impl_->launchEnergyTransfer();
 400 }
 401
 402 void GpuBonded::waitAccumulateEnergyTerms(gmx_enerdata_t* enerd)
 403 {
 404     impl_->waitAccumulateEnergyTerms(enerd);
 405 }
 406
 407 void GpuBonded::clearEnergies()
 408 {
 409     impl_->clearEnergies();
 410 }
 411
 412 } // namespace gmx