src/gromacs/nbnxm/kerneldispatch.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2016 by the GROMACS development team.
   5  * Copyright (c) 2017,2018,2019,2020,2021, by the GROMACS development team, led by
   6  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   7  * and including many others, as listed in the AUTHORS file in the
   8  * top-level source directory and at http://www.gromacs.org.
   9  *
  10  * GROMACS is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public License
  12  * as published by the Free Software Foundation; either version 2.1
  13  * of the License, or (at your option) any later version.
  14  *
  15  * GROMACS is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with GROMACS; if not, see
  22  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  23  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  24  *
  25  * If you want to redistribute modifications to GROMACS, please
  26  * consider that scientific software is very special. Version
  27  * control is crucial - bugs must be traceable. We will be happy to
  28  * consider code for inclusion in the official distribution, but
  29  * derived work must not be called official GROMACS. Details are found
  30  * in the README & COPYING files - if they are missing, get the
  31  * official version at http://www.gromacs.org.
  32  *
  33  * To help us fund GROMACS development, we humbly ask that you cite
  34  * the research papers on the package. Check out http://www.gromacs.org.
  35  */
  36
  37 #include "gmxpre.h"
  38
  39 #include "gromacs/gmxlib/nrnb.h"
  40 #include "gromacs/math/vectypes.h"
  41 #include "gromacs/mdlib/enerdata_utils.h"
  42 #include "gromacs/mdlib/force.h"
  43 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  44 #include "gromacs/mdtypes/enerdata.h"
  45 #include "gromacs/mdtypes/forceoutput.h"
  46 #include "gromacs/mdtypes/forcerec.h"
  47 #include "gromacs/mdtypes/inputrec.h"
  48 #include "gromacs/mdtypes/interaction_const.h"
  49 #include "gromacs/mdtypes/md_enums.h"
  50 #include "gromacs/mdtypes/mdatom.h"
  51 #include "gromacs/mdtypes/nblist.h"
  52 #include "gromacs/mdtypes/simulation_workload.h"
  53 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  54 #include "gromacs/nbnxm/nbnxm.h"
  55 #include "gromacs/simd/simd.h"
  56 #include "gromacs/timing/wallcycle.h"
  57 #include "gromacs/utility/enumerationhelpers.h"
  58 #include "gromacs/utility/fatalerror.h"
  59 #include "gromacs/utility/gmxassert.h"
  60 #include "gromacs/utility/real.h"
  61
  62 #include "kernel_common.h"
  63 #include "nbnxm_gpu.h"
  64 #include "nbnxm_simd.h"
  65 #include "pairlistset.h"
  66 #include "pairlistsets.h"
  67 #include "kernels_reference/kernel_gpu_ref.h"
  68 #define INCLUDE_KERNELFUNCTION_TABLES
  69 #include "kernels_reference/kernel_ref.h"
  70 #ifdef GMX_NBNXN_SIMD_2XNN
  71 #    include "kernels_simd_2xmm/kernels.h"
  72 #endif
  73 #ifdef GMX_NBNXN_SIMD_4XN
  74 #    include "kernels_simd_4xm/kernels.h"
  75 #endif
  76 #undef INCLUDE_FUNCTION_TABLES
  77
  78 /*! \brief Clears the energy group output buffers
  79  *
  80  * \param[in,out] out  nbnxn kernel output struct
  81  */
  82 static void clearGroupEnergies(nbnxn_atomdata_output_t* out)
  83 {
  84     std::fill(out->Vvdw.begin(), out->Vvdw.end(), 0.0_real);
  85     std::fill(out->Vc.begin(), out->Vc.end(), 0.0_real);
  86     std::fill(out->VSvdw.begin(), out->VSvdw.end(), 0.0_real);
  87     std::fill(out->VSc.begin(), out->VSc.end(), 0.0_real);
  88 }
  89
  90 /*! \brief Reduce the group-pair energy buffers produced by a SIMD kernel
  91  * to single terms in the output buffers.
  92  *
  93  * The SIMD kernels produce a large number of energy buffer in SIMD registers
  94  * to avoid scattered reads and writes.
  95  *
  96  * \tparam        unrollj         The unroll size for j-particles in the SIMD kernel
  97  * \param[in]     numGroups       The number of energy groups
  98  * \param[in]     numGroups_2log  Log2 of numGroups, rounded up
  99  * \param[in,out] out             Struct with energy buffers
 100  */
 101 template<int unrollj>
 102 static void reduceGroupEnergySimdBuffers(int numGroups, int numGroups_2log, nbnxn_atomdata_output_t* out)
 103 {
 104     const int unrollj_half = unrollj / 2;
 105     /* Energies are stored in SIMD registers with size 2^numGroups_2log */
 106     const int numGroupsStorage = (1 << numGroups_2log);
 107
 108     const real* gmx_restrict vVdwSimd     = out->VSvdw.data();
 109     const real* gmx_restrict vCoulombSimd = out->VSc.data();
 110     real* gmx_restrict       vVdw         = out->Vvdw.data();
 111     real* gmx_restrict       vCoulomb     = out->Vc.data();
 112
 113     /* The size of the SIMD energy group buffer array is:
 114      * numGroups*numGroups*numGroupsStorage*unrollj_half*simd_width
 115      */
 116     for (int i = 0; i < numGroups; i++)
 117     {
 118         for (int j1 = 0; j1 < numGroups; j1++)
 119         {
 120             for (int j0 = 0; j0 < numGroups; j0++)
 121             {
 122                 int c = ((i * numGroups + j1) * numGroupsStorage + j0) * unrollj_half * unrollj;
 123                 for (int s = 0; s < unrollj_half; s++)
 124                 {
 125                     vVdw[i * numGroups + j0] += vVdwSimd[c + 0];
 126                     vVdw[i * numGroups + j1] += vVdwSimd[c + 1];
 127                     vCoulomb[i * numGroups + j0] += vCoulombSimd[c + 0];
 128                     vCoulomb[i * numGroups + j1] += vCoulombSimd[c + 1];
 129                     c += unrollj + 2;
 130                 }
 131             }
 132         }
 133     }
 134 }
 135
 136 CoulombKernelType getCoulombKernelType(const Nbnxm::EwaldExclusionType ewaldExclusionType,
 137                                        const CoulombInteractionType    coulombInteractionType,
 138                                        const bool                      haveEqualCoulombVwdRadii)
 139 {
 140
 141     if (EEL_RF(coulombInteractionType) || coulombInteractionType == CoulombInteractionType::Cut)
 142     {
 143         return CoulombKernelType::ReactionField;
 144     }
 145     else
 146     {
 147         if (ewaldExclusionType == Nbnxm::EwaldExclusionType::Table)
 148         {
 149             if (haveEqualCoulombVwdRadii)
 150             {
 151                 return CoulombKernelType::Table;
 152             }
 153             else
 154             {
 155                 return CoulombKernelType::TableTwin;
 156             }
 157         }
 158         else
 159         {
 160             if (haveEqualCoulombVwdRadii)
 161             {
 162                 return CoulombKernelType::Ewald;
 163             }
 164             else
 165             {
 166                 return CoulombKernelType::EwaldTwin;
 167             }
 168         }
 169     }
 170 }
 171
 172 int getVdwKernelType(const Nbnxm::KernelType    kernelType,
 173                      const LJCombinationRule    ljCombinationRule,
 174                      const VanDerWaalsType      vanDerWaalsType,
 175                      const InteractionModifiers interactionModifiers,
 176                      const LongRangeVdW         longRangeVdW)
 177 {
 178     if (vanDerWaalsType == VanDerWaalsType::Cut)
 179     {
 180         switch (interactionModifiers)
 181         {
 182             case InteractionModifiers::None:
 183             case InteractionModifiers::PotShift:
 184                 switch (ljCombinationRule)
 185                 {
 186                     case LJCombinationRule::Geometric: return vdwktLJCUT_COMBGEOM;
 187                     case LJCombinationRule::LorentzBerthelot: return vdwktLJCUT_COMBLB;
 188                     case LJCombinationRule::None: return vdwktLJCUT_COMBNONE;
 189                     default: GMX_THROW(gmx::InvalidInputError("Unknown combination rule"));
 190                 }
 191             case InteractionModifiers::ForceSwitch: return vdwktLJFORCESWITCH;
 192             case InteractionModifiers::PotSwitch: return vdwktLJPOTSWITCH;
 193             default:
 194                 std::string errorMsg =
 195                         gmx::formatString("Unsupported VdW interaction modifier %s (%d)",
 196                                           enumValueToString(interactionModifiers),
 197                                           static_cast<int>(interactionModifiers));
 198                 GMX_THROW(gmx::InvalidInputError(errorMsg));
 199         }
 200     }
 201     else if (vanDerWaalsType == VanDerWaalsType::Pme)
 202     {
 203         if (longRangeVdW == LongRangeVdW::Geom)
 204         {
 205             return vdwktLJEWALDCOMBGEOM;
 206         }
 207         else
 208         {
 209             /* At setup we (should have) selected the C reference kernel */
 210             GMX_RELEASE_ASSERT(kernelType == Nbnxm::KernelType::Cpu4x4_PlainC,
 211                                "Only the C reference nbnxn SIMD kernel supports LJ-PME with LB "
 212                                "combination rules");
 213             return vdwktLJEWALDCOMBLB;
 214         }
 215     }
 216     else
 217     {
 218         std::string errorMsg = gmx::formatString("Unsupported VdW interaction type %s (%d)",
 219                                                  enumValueToString(vanDerWaalsType),
 220                                                  static_cast<int>(vanDerWaalsType));
 221         GMX_THROW(gmx::InvalidInputError(errorMsg));
 222     }
 223 }
 224
 225 /*! \brief Dispatches the non-bonded N versus M atom cluster CPU kernels.
 226  *
 227  * OpenMP parallelization is performed within this function.
 228  * Energy reduction, but not force and shift force reduction, is performed
 229  * within this function.
 230  *
 231  * \param[in]     pairlistSet   Pairlists with local or non-local interactions to compute
 232  * \param[in]     kernelSetup   The non-bonded kernel setup
 233  * \param[in,out] nbat          The atomdata for the interactions
 234  * \param[in]     ic            Non-bonded interaction constants
 235  * \param[in]     shiftVectors  The PBC shift vectors
 236  * \param[in]     stepWork      Flags that tell what to compute
 237  * \param[in]     clearF        Enum that tells if to clear the force output buffer
 238  * \param[out]    vCoulomb      Output buffer for Coulomb energies
 239  * \param[out]    vVdw          Output buffer for Van der Waals energies
 240  * \param[in]     wcycle        Pointer to cycle counting data structure.
 241  */
 242 static void nbnxn_kernel_cpu(const PairlistSet&             pairlistSet,
 243                              const Nbnxm::KernelSetup&      kernelSetup,
 244                              nbnxn_atomdata_t*              nbat,
 245                              const interaction_const_t&     ic,
 246                              gmx::ArrayRef<const gmx::RVec> shiftVectors,
 247                              const gmx::StepWorkload&       stepWork,
 248                              int                            clearF,
 249                              real*                          vCoulomb,
 250                              real*                          vVdw,
 251                              gmx_wallcycle*                 wcycle)
 252 {
 253
 254     const nbnxn_atomdata_t::Params& nbatParams = nbat->params();
 255
 256     const int coulkt = static_cast<int>(getCoulombKernelType(
 257             kernelSetup.ewaldExclusionType, ic.eeltype, (ic.rcoulomb == ic.rvdw)));
 258     const int vdwkt  = getVdwKernelType(
 259             kernelSetup.kernelType, nbatParams.ljCombinationRule, ic.vdwtype, ic.vdw_modifier, ic.ljpme_comb_rule);
 260
 261     gmx::ArrayRef<const NbnxnPairlistCpu> pairlists = pairlistSet.cpuLists();
 262
 263     const auto* shiftVecPointer = as_rvec_array(shiftVectors.data());
 264
 265     int gmx_unused nthreads = gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded);
 266     wallcycle_sub_start(wcycle, WallCycleSubCounter::NonbondedClear);
 267 #pragma omp parallel for schedule(static) num_threads(nthreads)
 268     for (gmx::index nb = 0; nb < pairlists.ssize(); nb++)
 269     {
 270         // Presently, the kernels do not call C++ code that can throw,
 271         // so no need for a try/catch pair in this OpenMP region.
 272         nbnxn_atomdata_output_t* out = &nbat->out[nb];
 273
 274         if (clearF == enbvClearFYes)
 275         {
 276             clearForceBuffer(nbat, nb);
 277
 278             clear_fshift(out->fshift.data());
 279         }
 280
 281         if (nb == 0)
 282         {
 283             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NonbondedClear);
 284             wallcycle_sub_start(wcycle, WallCycleSubCounter::NonbondedKernel);
 285         }
 286
 287         // TODO: Change to reference
 288         const NbnxnPairlistCpu* pairlist = &pairlists[nb];
 289
 290         if (!stepWork.computeEnergy)
 291         {
 292             /* Don't calculate energies */
 293             switch (kernelSetup.kernelType)
 294             {
 295                 case Nbnxm::KernelType::Cpu4x4_PlainC:
 296                     nbnxn_kernel_noener_ref[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 297                     break;
 298 #ifdef GMX_NBNXN_SIMD_2XNN
 299                 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 300                     nbnxm_kernel_noener_simd_2xmm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 301                     break;
 302 #endif
 303 #ifdef GMX_NBNXN_SIMD_4XN
 304                 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 305                     nbnxm_kernel_noener_simd_4xm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 306                     break;
 307 #endif
 308                 default: GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 309             }
 310         }
 311         else if (out->Vvdw.size() == 1)
 312         {
 313             /* A single energy group (pair) */
 314             out->Vvdw[0] = 0;
 315             out->Vc[0]   = 0;
 316
 317             switch (kernelSetup.kernelType)
 318             {
 319                 case Nbnxm::KernelType::Cpu4x4_PlainC:
 320                     nbnxn_kernel_ener_ref[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 321                     break;
 322 #ifdef GMX_NBNXN_SIMD_2XNN
 323                 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 324                     nbnxm_kernel_ener_simd_2xmm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 325                     break;
 326 #endif
 327 #ifdef GMX_NBNXN_SIMD_4XN
 328                 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 329                     nbnxm_kernel_ener_simd_4xm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 330                     break;
 331 #endif
 332                 default: GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 333             }
 334         }
 335         else
 336         {
 337             /* Calculate energy group contributions */
 338             clearGroupEnergies(out);
 339
 340             int unrollj = 0;
 341
 342             switch (kernelSetup.kernelType)
 343             {
 344                 case Nbnxm::KernelType::Cpu4x4_PlainC:
 345                     unrollj = c_nbnxnCpuIClusterSize;
 346                     nbnxn_kernel_energrp_ref[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 347                     break;
 348 #ifdef GMX_NBNXN_SIMD_2XNN
 349                 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 350                     unrollj = GMX_SIMD_REAL_WIDTH / 2;
 351                     nbnxm_kernel_energrp_simd_2xmm[coulkt][vdwkt](
 352                             pairlist, nbat, &ic, shiftVecPointer, out);
 353                     break;
 354 #endif
 355 #ifdef GMX_NBNXN_SIMD_4XN
 356                 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 357                     unrollj = GMX_SIMD_REAL_WIDTH;
 358                     nbnxm_kernel_energrp_simd_4xm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 359                     break;
 360 #endif
 361                 default: GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 362             }
 363
 364             if (kernelSetup.kernelType != Nbnxm::KernelType::Cpu4x4_PlainC)
 365             {
 366                 switch (unrollj)
 367                 {
 368                     case 2:
 369                         reduceGroupEnergySimdBuffers<2>(nbatParams.nenergrp, nbatParams.neg_2log, out);
 370                         break;
 371                     case 4:
 372                         reduceGroupEnergySimdBuffers<4>(nbatParams.nenergrp, nbatParams.neg_2log, out);
 373                         break;
 374                     case 8:
 375                         reduceGroupEnergySimdBuffers<8>(nbatParams.nenergrp, nbatParams.neg_2log, out);
 376                         break;
 377                     default: GMX_RELEASE_ASSERT(false, "Unsupported j-unroll size");
 378                 }
 379             }
 380         }
 381     }
 382     wallcycle_sub_stop(wcycle, WallCycleSubCounter::NonbondedKernel);
 383
 384     if (stepWork.computeEnergy)
 385     {
 386         reduce_energies_over_lists(nbat, pairlists.ssize(), vVdw, vCoulomb);
 387     }
 388 }
 389
 390 static void accountFlops(t_nrnb*                    nrnb,
 391                          const PairlistSet&         pairlistSet,
 392                          const nonbonded_verlet_t&  nbv,
 393                          const interaction_const_t& ic,
 394                          const gmx::StepWorkload&   stepWork)
 395 {
 396     const bool usingGpuKernels = nbv.useGpu();
 397
 398     int enr_nbnxn_kernel_ljc = eNRNB;
 399     if (EEL_RF(ic.eeltype) || ic.eeltype == CoulombInteractionType::Cut)
 400     {
 401         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
 402     }
 403     else if ((!usingGpuKernels && nbv.kernelSetup().ewaldExclusionType == Nbnxm::EwaldExclusionType::Analytical)
 404              || (usingGpuKernels && Nbnxm::gpu_is_kernel_ewald_analytical(nbv.gpu_nbv)))
 405     {
 406         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
 407     }
 408     else
 409     {
 410         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
 411     }
 412     int enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
 413     if (stepWork.computeEnergy)
 414     {
 415         /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 416         enr_nbnxn_kernel_ljc += 1;
 417         enr_nbnxn_kernel_lj += 1;
 418     }
 419
 420     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc, pairlistSet.natpair_ljq_);
 421     inc_nrnb(nrnb, enr_nbnxn_kernel_lj, pairlistSet.natpair_lj_);
 422     /* The Coulomb-only kernels are offset -eNR_NBNXN_LJ_RF+eNR_NBNXN_RF */
 423     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc - eNR_NBNXN_LJ_RF + eNR_NBNXN_RF, pairlistSet.natpair_q_);
 424
 425     if (ic.vdw_modifier == InteractionModifiers::ForceSwitch)
 426     {
 427         /* We add up the switch cost separately */
 428         inc_nrnb(nrnb,
 429                  eNR_NBNXN_ADD_LJ_FSW + (stepWork.computeEnergy ? 1 : 0),
 430                  pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_);
 431     }
 432     if (ic.vdw_modifier == InteractionModifiers::PotSwitch)
 433     {
 434         /* We add up the switch cost separately */
 435         inc_nrnb(nrnb,
 436                  eNR_NBNXN_ADD_LJ_PSW + (stepWork.computeEnergy ? 1 : 0),
 437                  pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_);
 438     }
 439     if (ic.vdwtype == VanDerWaalsType::Pme)
 440     {
 441         /* We add up the LJ Ewald cost separately */
 442         inc_nrnb(nrnb,
 443                  eNR_NBNXN_ADD_LJ_EWALD + (stepWork.computeEnergy ? 1 : 0),
 444                  pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_);
 445     }
 446 }
 447
 448 void nonbonded_verlet_t::dispatchNonbondedKernel(gmx::InteractionLocality       iLocality,
 449                                                  const interaction_const_t&     ic,
 450                                                  const gmx::StepWorkload&       stepWork,
 451                                                  int                            clearF,
 452                                                  gmx::ArrayRef<const gmx::RVec> shiftvec,
 453                                                  gmx::ArrayRef<real> repulsionDispersionSR,
 454                                                  gmx::ArrayRef<real> CoulombSR,
 455                                                  t_nrnb*             nrnb) const
 456 {
 457     const PairlistSet& pairlistSet = pairlistSets().pairlistSet(iLocality);
 458
 459     switch (kernelSetup().kernelType)
 460     {
 461         case Nbnxm::KernelType::Cpu4x4_PlainC:
 462         case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 463         case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 464             nbnxn_kernel_cpu(pairlistSet,
 465                              kernelSetup(),
 466                              nbat.get(),
 467                              ic,
 468                              shiftvec,
 469                              stepWork,
 470                              clearF,
 471                              CoulombSR.data(),
 472                              repulsionDispersionSR.data(),
 473                              wcycle_);
 474             break;
 475
 476         case Nbnxm::KernelType::Gpu8x8x8:
 477             Nbnxm::gpu_launch_kernel(gpu_nbv, stepWork, iLocality);
 478             break;
 479
 480         case Nbnxm::KernelType::Cpu8x8x8_PlainC:
 481             nbnxn_kernel_gpu_ref(pairlistSet.gpuList(),
 482                                  nbat.get(),
 483                                  &ic,
 484                                  shiftvec,
 485                                  stepWork,
 486                                  clearF,
 487                                  nbat->out[0].f,
 488                                  nbat->out[0].fshift.data(),
 489                                  CoulombSR.data(),
 490                                  repulsionDispersionSR.data());
 491             break;
 492
 493         default: GMX_RELEASE_ASSERT(false, "Invalid nonbonded kernel type passed!");
 494     }
 495
 496     accountFlops(nrnb, pairlistSet, *this, ic, stepWork);
 497 }