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(c_threadsPerBlock >= SHIFTS,
  82                   "Threads per block in GPU bonded must be >= SHIFTS for the virial kernel "
  83                   "(calcVir=true)");
  84
  85     wcycle_ = wcycle;
  86
  87     allocateDeviceBuffer(&d_forceParams_, ffparams.numTypes(), deviceContext_);
  88     // This could be an async transfer (if the source is pinned), so
  89     // long as it uses the same stream as the kernels and we are happy
  90     // to consume additional pinned pages.
  91     copyToDeviceBuffer(&d_forceParams_, ffparams.iparams.data(), 0, ffparams.numTypes(),
  92                        deviceStream_, GpuApiCallBehavior::Sync, nullptr);
  93     vTot_.resize(F_NRE);
  94     allocateDeviceBuffer(&d_vTot_, F_NRE, deviceContext_);
  95     clearDeviceBufferAsync(&d_vTot_, 0, F_NRE, deviceStream_);
  96
  97     kernelParams_.electrostaticsScaleFactor = electrostaticsScaleFactor;
  98     kernelParams_.d_forceParams             = d_forceParams_;
  99     kernelParams_.d_xq                      = d_xq_;
 100     kernelParams_.d_f                       = d_f_;
 101     kernelParams_.d_fShift                  = d_fShift_;
 102     kernelParams_.d_vTot                    = d_vTot_;
 103     for (int i = 0; i < numFTypesOnGpu; i++)
 104     {
 105         kernelParams_.d_iatoms[i]        = nullptr;
 106         kernelParams_.fTypeRangeStart[i] = 0;
 107         kernelParams_.fTypeRangeEnd[i]   = -1;
 108     }
 109
 110     int fTypeRangeEnd = kernelParams_.fTypeRangeEnd[numFTypesOnGpu - 1];
 111
 112     kernelLaunchConfig_.blockSize[0] = c_threadsPerBlock;
 113     kernelLaunchConfig_.blockSize[1] = 1;
 114     kernelLaunchConfig_.blockSize[2] = 1;
 115     kernelLaunchConfig_.gridSize[0]  = (fTypeRangeEnd + c_threadsPerBlock) / c_threadsPerBlock;
 116     kernelLaunchConfig_.gridSize[1]  = 1;
 117     kernelLaunchConfig_.gridSize[2]  = 1;
 118     kernelLaunchConfig_.sharedMemorySize =
 119             SHIFTS * sizeof(float3) + (c_threadsPerBlock / warp_size) * 3 * sizeof(float);
 120 }
 121
 122 GpuBonded::Impl::~Impl()
 123 {
 124     for (int fType : fTypesOnGpu)
 125     {
 126         if (d_iLists_[fType].iatoms)
 127         {
 128             freeDeviceBuffer(&d_iLists_[fType].iatoms);
 129             d_iLists_[fType].iatoms = nullptr;
 130         }
 131     }
 132
 133     freeDeviceBuffer(&d_forceParams_);
 134     freeDeviceBuffer(&d_vTot_);
 135 }
 136
 137 //! Return whether function type \p fType in \p idef has perturbed interactions
 138 static bool fTypeHasPerturbedEntries(const InteractionDefinitions& idef, int fType)
 139 {
 140     GMX_ASSERT(idef.ilsort == ilsortNO_FE || idef.ilsort == ilsortFE_SORTED,
 141                "Perturbed interations should be sorted here");
 142
 143     const InteractionList& ilist = idef.il[fType];
 144
 145     return (idef.ilsort != ilsortNO_FE && idef.numNonperturbedInteractions[fType] != ilist.size());
 146 }
 147
 148 //! Converts \p src with atom indices in state order to \p dest in nbnxn order
 149 static void convertIlistToNbnxnOrder(const InteractionList& src,
 150                                      HostInteractionList*   dest,
 151                                      int                    numAtomsPerInteraction,
 152                                      ArrayRef<const int>    nbnxnAtomOrder)
 153 {
 154     GMX_ASSERT(src.size() == 0 || !nbnxnAtomOrder.empty(), "We need the nbnxn atom order");
 155
 156     dest->iatoms.resize(src.size());
 157
 158     // TODO use OpenMP to parallelise this loop
 159     for (int i = 0; i < src.size(); i += 1 + numAtomsPerInteraction)
 160     {
 161         dest->iatoms[i] = src.iatoms[i];
 162         for (int a = 0; a < numAtomsPerInteraction; a++)
 163         {
 164             dest->iatoms[i + 1 + a] = nbnxnAtomOrder[src.iatoms[i + 1 + a]];
 165         }
 166     }
 167 }
 168
 169 //! Returns \p input rounded up to the closest multiple of \p factor.
 170 static inline int roundUpToFactor(const int input, const int factor)
 171 {
 172     GMX_ASSERT(factor > 0, "The factor to round up to must be > 0.");
 173
 174     int remainder = input % factor;
 175
 176     if (remainder == 0)
 177     {
 178         return (input);
 179     }
 180     return (input + (factor - remainder));
 181 }
 182
 183 // TODO Consider whether this function should be a factory method that
 184 // makes an object that is the only one capable of the device
 185 // operations needed for the lifetime of an interaction list. This
 186 // would be harder to misuse than GpuBonded, and exchange the problem
 187 // of naming this method for the problem of what to name the
 188 // BondedDeviceInteractionListHandler type.
 189
 190 /*! Divides bonded interactions over threads and GPU.
 191  *  The bonded interactions are assigned by interaction type to GPU threads. The intereaction
 192  *  types are assigned in blocks sized as <warp_size>. The beginning and end (thread index) of each
 193  *  interaction type are stored in kernelParams_. Pointers to the relevant data structures on the
 194  *  GPU are also stored in kernelParams_.
 195  *
 196  * \todo Use DeviceBuffer for the d_xqPtr.
 197  */
 198 void GpuBonded::Impl::updateInteractionListsAndDeviceBuffers(ArrayRef<const int> nbnxnAtomOrder,
 199                                                              const InteractionDefinitions& idef,
 200                                                              void*                         d_xqPtr,
 201                                                              DeviceBuffer<RVec>            d_fPtr,
 202                                                              DeviceBuffer<RVec> d_fShiftPtr)
 203 {
 204     // TODO wallcycle sub start
 205     haveInteractions_ = false;
 206     int fTypesCounter = 0;
 207
 208     for (int fType : fTypesOnGpu)
 209     {
 210         auto& iList = iLists_[fType];
 211
 212         /* Perturbation is not implemented in the GPU bonded kernels.
 213          * But instead of doing all interactions on the CPU, we can
 214          * still easily handle the types that have no perturbed
 215          * interactions on the GPU. */
 216         if (!idef.il[fType].empty() && !fTypeHasPerturbedEntries(idef, fType))
 217         {
 218             haveInteractions_ = true;
 219
 220             convertIlistToNbnxnOrder(idef.il[fType], &iList, NRAL(fType), nbnxnAtomOrder);
 221         }
 222         else
 223         {
 224             iList.iatoms.clear();
 225         }
 226
 227         // Update the device if necessary. This can leave some
 228         // allocations on the device when the host size decreases to
 229         // zero, which is OK, since we deallocate everything at the
 230         // end.
 231         if (iList.size() > 0)
 232         {
 233             t_ilist& d_iList = d_iLists_[fType];
 234
 235             reallocateDeviceBuffer(&d_iList.iatoms, iList.size(), &d_iList.nr, &d_iList.nalloc,
 236                                    deviceContext_);
 237
 238             copyToDeviceBuffer(&d_iList.iatoms, iList.iatoms.data(), 0, iList.size(), deviceStream_,
 239                                GpuApiCallBehavior::Async, nullptr);
 240         }
 241         kernelParams_.fTypesOnGpu[fTypesCounter]    = fType;
 242         kernelParams_.numFTypeIAtoms[fTypesCounter] = iList.size();
 243         int numBonds = iList.size() / (interaction_function[fType].nratoms + 1);
 244         kernelParams_.numFTypeBonds[fTypesCounter] = numBonds;
 245         kernelParams_.d_iatoms[fTypesCounter]      = d_iLists_[fType].iatoms;
 246         if (fTypesCounter == 0)
 247         {
 248             kernelParams_.fTypeRangeStart[fTypesCounter] = 0;
 249         }
 250         else
 251         {
 252             kernelParams_.fTypeRangeStart[fTypesCounter] =
 253                     kernelParams_.fTypeRangeEnd[fTypesCounter - 1] + 1;
 254         }
 255         kernelParams_.fTypeRangeEnd[fTypesCounter] = kernelParams_.fTypeRangeStart[fTypesCounter]
 256                                                      + roundUpToFactor(numBonds, warp_size) - 1;
 257
 258         GMX_ASSERT(numBonds > 0
 259                            || kernelParams_.fTypeRangeEnd[fTypesCounter]
 260                                       <= kernelParams_.fTypeRangeStart[fTypesCounter],
 261                    "Invalid GPU listed forces setup. numBonds must be > 0 if there are threads "
 262                    "allocated to do work on that interaction function type.");
 263         GMX_ASSERT(kernelParams_.fTypeRangeStart[fTypesCounter] % warp_size == 0
 264                            && (kernelParams_.fTypeRangeEnd[fTypesCounter] + 1) % warp_size == 0,
 265                    "The bonded interactions must be assigned to the GPU in blocks of warp size.");
 266
 267         fTypesCounter++;
 268     }
 269
 270     int fTypeRangeEnd               = kernelParams_.fTypeRangeEnd[numFTypesOnGpu - 1];
 271     kernelLaunchConfig_.gridSize[0] = (fTypeRangeEnd + c_threadsPerBlock) / c_threadsPerBlock;
 272
 273     d_xq_     = static_cast<float4*>(d_xqPtr);
 274     d_f_      = asFloat3(d_fPtr);
 275     d_fShift_ = asFloat3(d_fShiftPtr);
 276
 277     kernelParams_.d_xq          = d_xq_;
 278     kernelParams_.d_f           = d_f_;
 279     kernelParams_.d_fShift      = d_fShift_;
 280     kernelParams_.d_forceParams = d_forceParams_;
 281     kernelParams_.d_vTot        = d_vTot_;
 282
 283     // TODO wallcycle sub stop
 284 }
 285
 286 void GpuBonded::Impl::setPbc(PbcType pbcType, const matrix box, bool canMoleculeSpanPbc)
 287 {
 288     PbcAiuc pbcAiuc;
 289     setPbcAiuc(canMoleculeSpanPbc ? numPbcDimensions(pbcType) : 0, box, &pbcAiuc);
 290     kernelParams_.pbcAiuc = pbcAiuc;
 291 }
 292
 293 bool GpuBonded::Impl::haveInteractions() const
 294 {
 295     return haveInteractions_;
 296 }
 297
 298 void GpuBonded::Impl::launchEnergyTransfer()
 299 {
 300     GMX_ASSERT(haveInteractions_,
 301                "No GPU bonded interactions, so no energies will be computed, so transfer should "
 302                "not be called");
 303     wallcycle_sub_start_nocount(wcycle_, ewcsLAUNCH_GPU_BONDED);
 304     // TODO add conditional on whether there has been any compute (and make sure host buffer doesn't contain garbage)
 305     float* h_vTot = vTot_.data();
 306     copyFromDeviceBuffer(h_vTot, &d_vTot_, 0, F_NRE, deviceStream_, GpuApiCallBehavior::Async, nullptr);
 307     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_GPU_BONDED);
 308 }
 309
 310 void GpuBonded::Impl::waitAccumulateEnergyTerms(gmx_enerdata_t* enerd)
 311 {
 312     GMX_ASSERT(haveInteractions_,
 313                "No GPU bonded interactions, so no energies will be computed or transferred, so "
 314                "accumulation should not occur");
 315
 316     wallcycle_start(wcycle_, ewcWAIT_GPU_BONDED);
 317     cudaError_t stat = cudaStreamSynchronize(deviceStream_.stream());
 318     CU_RET_ERR(stat, "D2H transfer of bonded energies failed");
 319     wallcycle_stop(wcycle_, ewcWAIT_GPU_BONDED);
 320
 321     for (int fType : fTypesOnGpu)
 322     {
 323         if (fType != F_LJ14 && fType != F_COUL14)
 324         {
 325             enerd->term[fType] += vTot_[fType];
 326         }
 327     }
 328
 329     // Note: We do not support energy groups here
 330     gmx_grppairener_t* grppener = &enerd->grpp;
 331     GMX_RELEASE_ASSERT(grppener->nener == 1, "No energy group support for bondeds on the GPU");
 332     grppener->ener[egLJ14][0] += vTot_[F_LJ14];
 333     grppener->ener[egCOUL14][0] += vTot_[F_COUL14];
 334 }
 335
 336 void GpuBonded::Impl::clearEnergies()
 337 {
 338     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 339     wallcycle_sub_start_nocount(wcycle_, ewcsLAUNCH_GPU_BONDED);
 340     clearDeviceBufferAsync(&d_vTot_, 0, F_NRE, deviceStream_);
 341     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_GPU_BONDED);
 342     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 343 }
 344
 345 // ---- GpuBonded
 346
 347 GpuBonded::GpuBonded(const gmx_ffparams_t& ffparams,
 348                      const float           electrostaticsScaleFactor,
 349                      const DeviceContext&  deviceContext,
 350                      const DeviceStream&   deviceStream,
 351                      gmx_wallcycle*        wcycle) :
 352     impl_(new Impl(ffparams, electrostaticsScaleFactor, deviceContext, deviceStream, wcycle))
 353 {
 354 }
 355
 356 GpuBonded::~GpuBonded() = default;
 357
 358 void GpuBonded::updateInteractionListsAndDeviceBuffers(ArrayRef<const int>           nbnxnAtomOrder,
 359                                                        const InteractionDefinitions& idef,
 360                                                        void*                         d_xq,
 361                                                        DeviceBuffer<RVec>            d_f,
 362                                                        DeviceBuffer<RVec>            d_fShift)
 363 {
 364     impl_->updateInteractionListsAndDeviceBuffers(nbnxnAtomOrder, idef, d_xq, d_f, d_fShift);
 365 }
 366
 367 void GpuBonded::setPbc(PbcType pbcType, const matrix box, bool canMoleculeSpanPbc)
 368 {
 369     impl_->setPbc(pbcType, box, canMoleculeSpanPbc);
 370 }
 371
 372 bool GpuBonded::haveInteractions() const
 373 {
 374     return impl_->haveInteractions();
 375 }
 376
 377 void GpuBonded::setPbcAndlaunchKernel(PbcType                  pbcType,
 378                                       const matrix             box,
 379                                       bool                     canMoleculeSpanPbc,
 380                                       const gmx::StepWorkload& stepWork)
 381 {
 382     setPbc(pbcType, box, canMoleculeSpanPbc);
 383     launchKernel(stepWork);
 384 }
 385
 386 void GpuBonded::launchEnergyTransfer()
 387 {
 388     impl_->launchEnergyTransfer();
 389 }
 390
 391 void GpuBonded::waitAccumulateEnergyTerms(gmx_enerdata_t* enerd)
 392 {
 393     impl_->waitAccumulateEnergyTerms(enerd);
 394 }
 395
 396 void GpuBonded::clearEnergies()
 397 {
 398     impl_->clearEnergies();
 399 }
 400
 401 } // namespace gmx