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/gmxlib/nonbonded/nb_free_energy.h"
  41 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  42 #include "gromacs/math/vectypes.h"
  43 #include "gromacs/mdlib/enerdata_utils.h"
  44 #include "gromacs/mdlib/force.h"
  45 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  46 #include "gromacs/mdtypes/enerdata.h"
  47 #include "gromacs/mdtypes/forceoutput.h"
  48 #include "gromacs/mdtypes/forcerec.h"
  49 #include "gromacs/mdtypes/inputrec.h"
  50 #include "gromacs/mdtypes/interaction_const.h"
  51 #include "gromacs/mdtypes/md_enums.h"
  52 #include "gromacs/mdtypes/mdatom.h"
  53 #include "gromacs/mdtypes/nblist.h"
  54 #include "gromacs/mdtypes/simulation_workload.h"
  55 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  56 #include "gromacs/nbnxm/nbnxm.h"
  57 #include "gromacs/simd/simd.h"
  58 #include "gromacs/timing/wallcycle.h"
  59 #include "gromacs/utility/enumerationhelpers.h"
  60 #include "gromacs/utility/fatalerror.h"
  61 #include "gromacs/utility/gmxassert.h"
  62 #include "gromacs/utility/real.h"
  63
  64 #include "kernel_common.h"
  65 #include "nbnxm_gpu.h"
  66 #include "nbnxm_simd.h"
  67 #include "pairlistset.h"
  68 #include "pairlistsets.h"
  69 #include "kernels_reference/kernel_gpu_ref.h"
  70 #define INCLUDE_KERNELFUNCTION_TABLES
  71 #include "kernels_reference/kernel_ref.h"
  72 #ifdef GMX_NBNXN_SIMD_2XNN
  73 #    include "kernels_simd_2xmm/kernels.h"
  74 #endif
  75 #ifdef GMX_NBNXN_SIMD_4XN
  76 #    include "kernels_simd_4xm/kernels.h"
  77 #endif
  78 #undef INCLUDE_FUNCTION_TABLES
  79
  80 /*! \brief Clears the energy group output buffers
  81  *
  82  * \param[in,out] out  nbnxn kernel output struct
  83  */
  84 static void clearGroupEnergies(nbnxn_atomdata_output_t* out)
  85 {
  86     std::fill(out->Vvdw.begin(), out->Vvdw.end(), 0.0_real);
  87     std::fill(out->Vc.begin(), out->Vc.end(), 0.0_real);
  88     std::fill(out->VSvdw.begin(), out->VSvdw.end(), 0.0_real);
  89     std::fill(out->VSc.begin(), out->VSc.end(), 0.0_real);
  90 }
  91
  92 /*! \brief Reduce the group-pair energy buffers produced by a SIMD kernel
  93  * to single terms in the output buffers.
  94  *
  95  * The SIMD kernels produce a large number of energy buffer in SIMD registers
  96  * to avoid scattered reads and writes.
  97  *
  98  * \tparam        unrollj         The unroll size for j-particles in the SIMD kernel
  99  * \param[in]     numGroups       The number of energy groups
 100  * \param[in]     numGroups_2log  Log2 of numGroups, rounded up
 101  * \param[in,out] out             Struct with energy buffers
 102  */
 103 template<int unrollj>
 104 static void reduceGroupEnergySimdBuffers(int numGroups, int numGroups_2log, nbnxn_atomdata_output_t* out)
 105 {
 106     const int unrollj_half = unrollj / 2;
 107     /* Energies are stored in SIMD registers with size 2^numGroups_2log */
 108     const int numGroupsStorage = (1 << numGroups_2log);
 109
 110     const real* gmx_restrict vVdwSimd     = out->VSvdw.data();
 111     const real* gmx_restrict vCoulombSimd = out->VSc.data();
 112     real* gmx_restrict       vVdw         = out->Vvdw.data();
 113     real* gmx_restrict       vCoulomb     = out->Vc.data();
 114
 115     /* The size of the SIMD energy group buffer array is:
 116      * numGroups*numGroups*numGroupsStorage*unrollj_half*simd_width
 117      */
 118     for (int i = 0; i < numGroups; i++)
 119     {
 120         for (int j1 = 0; j1 < numGroups; j1++)
 121         {
 122             for (int j0 = 0; j0 < numGroups; j0++)
 123             {
 124                 int c = ((i * numGroups + j1) * numGroupsStorage + j0) * unrollj_half * unrollj;
 125                 for (int s = 0; s < unrollj_half; s++)
 126                 {
 127                     vVdw[i * numGroups + j0] += vVdwSimd[c + 0];
 128                     vVdw[i * numGroups + j1] += vVdwSimd[c + 1];
 129                     vCoulomb[i * numGroups + j0] += vCoulombSimd[c + 0];
 130                     vCoulomb[i * numGroups + j1] += vCoulombSimd[c + 1];
 131                     c += unrollj + 2;
 132                 }
 133             }
 134         }
 135     }
 136 }
 137
 138 CoulombKernelType getCoulombKernelType(const Nbnxm::EwaldExclusionType ewaldExclusionType,
 139                                        const CoulombInteractionType    coulombInteractionType,
 140                                        const bool                      haveEqualCoulombVwdRadii)
 141 {
 142
 143     if (EEL_RF(coulombInteractionType) || coulombInteractionType == CoulombInteractionType::Cut)
 144     {
 145         return CoulombKernelType::ReactionField;
 146     }
 147     else
 148     {
 149         if (ewaldExclusionType == Nbnxm::EwaldExclusionType::Table)
 150         {
 151             if (haveEqualCoulombVwdRadii)
 152             {
 153                 return CoulombKernelType::Table;
 154             }
 155             else
 156             {
 157                 return CoulombKernelType::TableTwin;
 158             }
 159         }
 160         else
 161         {
 162             if (haveEqualCoulombVwdRadii)
 163             {
 164                 return CoulombKernelType::Ewald;
 165             }
 166             else
 167             {
 168                 return CoulombKernelType::EwaldTwin;
 169             }
 170         }
 171     }
 172 }
 173
 174 static int getVdwKernelType(const Nbnxm::KernelSetup&       kernelSetup,
 175                             const nbnxn_atomdata_t::Params& nbatParams,
 176                             const interaction_const_t&      ic)
 177 {
 178     if (ic.vdwtype == VanDerWaalsType::Cut)
 179     {
 180         switch (ic.vdw_modifier)
 181         {
 182             case InteractionModifiers::None:
 183             case InteractionModifiers::PotShift:
 184                 switch (nbatParams.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_incons("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(ic.vdw_modifier),
 197                                           static_cast<int>(ic.vdw_modifier));
 198                 gmx_incons(errorMsg);
 199         }
 200     }
 201     else if (ic.vdwtype == VanDerWaalsType::Pme)
 202     {
 203         if (ic.ljpme_comb_rule == 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(kernelSetup.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(ic.vdwtype),
 220                                                  static_cast<int>(ic.vdwtype));
 221         gmx_incons(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(kernelSetup, nbatParams, ic);
 259
 260     gmx::ArrayRef<const NbnxnPairlistCpu> pairlists = pairlistSet.cpuLists();
 261
 262     const auto* shiftVecPointer = as_rvec_array(shiftVectors.data());
 263
 264     int gmx_unused nthreads = gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded);
 265     wallcycle_sub_start(wcycle, WallCycleSubCounter::NonbondedClear);
 266 #pragma omp parallel for schedule(static) num_threads(nthreads)
 267     for (gmx::index nb = 0; nb < pairlists.ssize(); nb++)
 268     {
 269         // Presently, the kernels do not call C++ code that can throw,
 270         // so no need for a try/catch pair in this OpenMP region.
 271         nbnxn_atomdata_output_t* out = &nbat->out[nb];
 272
 273         if (clearF == enbvClearFYes)
 274         {
 275             clearForceBuffer(nbat, nb);
 276
 277             clear_fshift(out->fshift.data());
 278         }
 279
 280         if (nb == 0)
 281         {
 282             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NonbondedClear);
 283             wallcycle_sub_start(wcycle, WallCycleSubCounter::NonbondedKernel);
 284         }
 285
 286         // TODO: Change to reference
 287         const NbnxnPairlistCpu* pairlist = &pairlists[nb];
 288
 289         if (!stepWork.computeEnergy)
 290         {
 291             /* Don't calculate energies */
 292             switch (kernelSetup.kernelType)
 293             {
 294                 case Nbnxm::KernelType::Cpu4x4_PlainC:
 295                     nbnxn_kernel_noener_ref[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 296                     break;
 297 #ifdef GMX_NBNXN_SIMD_2XNN
 298                 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 299                     nbnxm_kernel_noener_simd_2xmm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 300                     break;
 301 #endif
 302 #ifdef GMX_NBNXN_SIMD_4XN
 303                 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 304                     nbnxm_kernel_noener_simd_4xm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 305                     break;
 306 #endif
 307                 default: GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 308             }
 309         }
 310         else if (out->Vvdw.size() == 1)
 311         {
 312             /* A single energy group (pair) */
 313             out->Vvdw[0] = 0;
 314             out->Vc[0]   = 0;
 315
 316             switch (kernelSetup.kernelType)
 317             {
 318                 case Nbnxm::KernelType::Cpu4x4_PlainC:
 319                     nbnxn_kernel_ener_ref[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 320                     break;
 321 #ifdef GMX_NBNXN_SIMD_2XNN
 322                 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 323                     nbnxm_kernel_ener_simd_2xmm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 324                     break;
 325 #endif
 326 #ifdef GMX_NBNXN_SIMD_4XN
 327                 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 328                     nbnxm_kernel_ener_simd_4xm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 329                     break;
 330 #endif
 331                 default: GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 332             }
 333         }
 334         else
 335         {
 336             /* Calculate energy group contributions */
 337             clearGroupEnergies(out);
 338
 339             int unrollj = 0;
 340
 341             switch (kernelSetup.kernelType)
 342             {
 343                 case Nbnxm::KernelType::Cpu4x4_PlainC:
 344                     unrollj = c_nbnxnCpuIClusterSize;
 345                     nbnxn_kernel_energrp_ref[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 346                     break;
 347 #ifdef GMX_NBNXN_SIMD_2XNN
 348                 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 349                     unrollj = GMX_SIMD_REAL_WIDTH / 2;
 350                     nbnxm_kernel_energrp_simd_2xmm[coulkt][vdwkt](
 351                             pairlist, nbat, &ic, shiftVecPointer, out);
 352                     break;
 353 #endif
 354 #ifdef GMX_NBNXN_SIMD_4XN
 355                 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 356                     unrollj = GMX_SIMD_REAL_WIDTH;
 357                     nbnxm_kernel_energrp_simd_4xm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVecPointer, out);
 358                     break;
 359 #endif
 360                 default: GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 361             }
 362
 363             if (kernelSetup.kernelType != Nbnxm::KernelType::Cpu4x4_PlainC)
 364             {
 365                 switch (unrollj)
 366                 {
 367                     case 2:
 368                         reduceGroupEnergySimdBuffers<2>(nbatParams.nenergrp, nbatParams.neg_2log, out);
 369                         break;
 370                     case 4:
 371                         reduceGroupEnergySimdBuffers<4>(nbatParams.nenergrp, nbatParams.neg_2log, out);
 372                         break;
 373                     case 8:
 374                         reduceGroupEnergySimdBuffers<8>(nbatParams.nenergrp, nbatParams.neg_2log, out);
 375                         break;
 376                     default: GMX_RELEASE_ASSERT(false, "Unsupported j-unroll size");
 377                 }
 378             }
 379         }
 380     }
 381     wallcycle_sub_stop(wcycle, WallCycleSubCounter::NonbondedKernel);
 382
 383     if (stepWork.computeEnergy)
 384     {
 385         reduce_energies_over_lists(nbat, pairlists.ssize(), vVdw, vCoulomb);
 386     }
 387 }
 388
 389 static void accountFlops(t_nrnb*                    nrnb,
 390                          const PairlistSet&         pairlistSet,
 391                          const nonbonded_verlet_t&  nbv,
 392                          const interaction_const_t& ic,
 393                          const gmx::StepWorkload&   stepWork)
 394 {
 395     const bool usingGpuKernels = nbv.useGpu();
 396
 397     int enr_nbnxn_kernel_ljc = eNRNB;
 398     if (EEL_RF(ic.eeltype) || ic.eeltype == CoulombInteractionType::Cut)
 399     {
 400         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
 401     }
 402     else if ((!usingGpuKernels && nbv.kernelSetup().ewaldExclusionType == Nbnxm::EwaldExclusionType::Analytical)
 403              || (usingGpuKernels && Nbnxm::gpu_is_kernel_ewald_analytical(nbv.gpu_nbv)))
 404     {
 405         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
 406     }
 407     else
 408     {
 409         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
 410     }
 411     int enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
 412     if (stepWork.computeEnergy)
 413     {
 414         /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 415         enr_nbnxn_kernel_ljc += 1;
 416         enr_nbnxn_kernel_lj += 1;
 417     }
 418
 419     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc, pairlistSet.natpair_ljq_);
 420     inc_nrnb(nrnb, enr_nbnxn_kernel_lj, pairlistSet.natpair_lj_);
 421     /* The Coulomb-only kernels are offset -eNR_NBNXN_LJ_RF+eNR_NBNXN_RF */
 422     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc - eNR_NBNXN_LJ_RF + eNR_NBNXN_RF, pairlistSet.natpair_q_);
 423
 424     if (ic.vdw_modifier == InteractionModifiers::ForceSwitch)
 425     {
 426         /* We add up the switch cost separately */
 427         inc_nrnb(nrnb,
 428                  eNR_NBNXN_ADD_LJ_FSW + (stepWork.computeEnergy ? 1 : 0),
 429                  pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_);
 430     }
 431     if (ic.vdw_modifier == InteractionModifiers::PotSwitch)
 432     {
 433         /* We add up the switch cost separately */
 434         inc_nrnb(nrnb,
 435                  eNR_NBNXN_ADD_LJ_PSW + (stepWork.computeEnergy ? 1 : 0),
 436                  pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_);
 437     }
 438     if (ic.vdwtype == VanDerWaalsType::Pme)
 439     {
 440         /* We add up the LJ Ewald cost separately */
 441         inc_nrnb(nrnb,
 442                  eNR_NBNXN_ADD_LJ_EWALD + (stepWork.computeEnergy ? 1 : 0),
 443                  pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_);
 444     }
 445 }
 446
 447 void nonbonded_verlet_t::dispatchNonbondedKernel(gmx::InteractionLocality       iLocality,
 448                                                  const interaction_const_t&     ic,
 449                                                  const gmx::StepWorkload&       stepWork,
 450                                                  int                            clearF,
 451                                                  gmx::ArrayRef<const gmx::RVec> shiftvec,
 452                                                  gmx::ArrayRef<real> repulsionDispersionSR,
 453                                                  gmx::ArrayRef<real> CoulombSR,
 454                                                  t_nrnb*             nrnb) const
 455 {
 456     const PairlistSet& pairlistSet = pairlistSets().pairlistSet(iLocality);
 457
 458     switch (kernelSetup().kernelType)
 459     {
 460         case Nbnxm::KernelType::Cpu4x4_PlainC:
 461         case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 462         case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 463             nbnxn_kernel_cpu(pairlistSet,
 464                              kernelSetup(),
 465                              nbat.get(),
 466                              ic,
 467                              shiftvec,
 468                              stepWork,
 469                              clearF,
 470                              CoulombSR.data(),
 471                              repulsionDispersionSR.data(),
 472                              wcycle_);
 473             break;
 474
 475         case Nbnxm::KernelType::Gpu8x8x8:
 476             Nbnxm::gpu_launch_kernel(gpu_nbv, stepWork, iLocality);
 477             break;
 478
 479         case Nbnxm::KernelType::Cpu8x8x8_PlainC:
 480             nbnxn_kernel_gpu_ref(pairlistSet.gpuList(),
 481                                  nbat.get(),
 482                                  &ic,
 483                                  shiftvec,
 484                                  stepWork,
 485                                  clearF,
 486                                  nbat->out[0].f,
 487                                  nbat->out[0].fshift.data(),
 488                                  CoulombSR.data(),
 489                                  repulsionDispersionSR.data());
 490             break;
 491
 492         default: GMX_RELEASE_ASSERT(false, "Invalid nonbonded kernel type passed!");
 493     }
 494
 495     accountFlops(nrnb, pairlistSet, *this, ic, stepWork);
 496 }
 497
 498 void nonbonded_verlet_t::dispatchFreeEnergyKernel(gmx::InteractionLocality       iLocality,
 499                                                   gmx::ArrayRef<const gmx::RVec> coords,
 500                                                   gmx::ForceWithShiftForces* forceWithShiftForces,
 501                                                   bool                       useSimd,
 502                                                   int                        ntype,
 503                                                   real                       rlist,
 504                                                   const interaction_const_t& ic,
 505                                                   gmx::ArrayRef<const gmx::RVec> shiftvec,
 506                                                   gmx::ArrayRef<const real>      nbfp,
 507                                                   gmx::ArrayRef<const real>      nbfp_grid,
 508                                                   gmx::ArrayRef<const real>      chargeA,
 509                                                   gmx::ArrayRef<const real>      chargeB,
 510                                                   gmx::ArrayRef<const int>       typeA,
 511                                                   gmx::ArrayRef<const int>       typeB,
 512                                                   t_lambda*                      fepvals,
 513                                                   gmx::ArrayRef<const real>      lambda,
 514                                                   gmx_enerdata_t*                enerd,
 515                                                   const gmx::StepWorkload&       stepWork,
 516                                                   t_nrnb*                        nrnb)
 517 {
 518     const auto nbl_fep = pairlistSets().pairlistSet(iLocality).fepLists();
 519
 520     /* When the first list is empty, all are empty and there is nothing to do */
 521     if (!pairlistSets().params().haveFep || nbl_fep[0]->nrj == 0)
 522     {
 523         return;
 524     }
 525
 526     int donb_flags = 0;
 527     /* Add short-range interactions */
 528     donb_flags |= GMX_NONBONDED_DO_SR;
 529
 530     if (stepWork.computeForces)
 531     {
 532         donb_flags |= GMX_NONBONDED_DO_FORCE;
 533     }
 534     if (stepWork.computeVirial)
 535     {
 536         donb_flags |= GMX_NONBONDED_DO_SHIFTFORCE;
 537     }
 538     if (stepWork.computeEnergy)
 539     {
 540         donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
 541     }
 542
 543     gmx::EnumerationArray<FreeEnergyPerturbationCouplingType, real> dvdl_nb      = { 0 };
 544     int                                                             kernelFlags  = donb_flags;
 545     gmx::ArrayRef<const real>                                       kernelLambda = lambda;
 546     gmx::ArrayRef<real>                                             kernelDvdl   = dvdl_nb;
 547
 548     gmx::ArrayRef<real> energygrp_elec = enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::CoulombSR];
 549     gmx::ArrayRef<real> energygrp_vdw = enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::LJSR];
 550
 551     GMX_ASSERT(gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded) == nbl_fep.ssize(),
 552                "Number of lists should be same as number of NB threads");
 553
 554     wallcycle_sub_start(wcycle_, WallCycleSubCounter::NonbondedFep);
 555 #pragma omp parallel for schedule(static) num_threads(nbl_fep.ssize())
 556     for (gmx::index th = 0; th < nbl_fep.ssize(); th++)
 557     {
 558         try
 559         {
 560             gmx_nb_free_energy_kernel(*nbl_fep[th],
 561                                       coords,
 562                                       forceWithShiftForces,
 563                                       useSimd,
 564                                       ntype,
 565                                       rlist,
 566                                       ic,
 567                                       shiftvec,
 568                                       nbfp,
 569                                       nbfp_grid,
 570                                       chargeA,
 571                                       chargeB,
 572                                       typeA,
 573                                       typeB,
 574                                       kernelFlags,
 575                                       kernelLambda,
 576                                       kernelDvdl,
 577                                       energygrp_elec,
 578                                       energygrp_vdw,
 579                                       nrnb);
 580         }
 581         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 582     }
 583
 584     if (fepvals->sc_alpha != 0)
 585     {
 586         enerd->dvdl_nonlin[FreeEnergyPerturbationCouplingType::Vdw] +=
 587                 dvdl_nb[FreeEnergyPerturbationCouplingType::Vdw];
 588         enerd->dvdl_nonlin[FreeEnergyPerturbationCouplingType::Coul] +=
 589                 dvdl_nb[FreeEnergyPerturbationCouplingType::Coul];
 590     }
 591     else
 592     {
 593         enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Vdw] +=
 594                 dvdl_nb[FreeEnergyPerturbationCouplingType::Vdw];
 595         enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Coul] +=
 596                 dvdl_nb[FreeEnergyPerturbationCouplingType::Coul];
 597     }
 598
 599     /* If we do foreign lambda and we have soft-core interactions
 600      * we have to recalculate the (non-linear) energies contributions.
 601      */
 602     if (fepvals->n_lambda > 0 && stepWork.computeDhdl && fepvals->sc_alpha != 0)
 603     {
 604         gmx::EnumerationArray<FreeEnergyPerturbationCouplingType, real> lam_i;
 605         kernelFlags = (donb_flags & ~(GMX_NONBONDED_DO_FORCE | GMX_NONBONDED_DO_SHIFTFORCE))
 606                       | GMX_NONBONDED_DO_FOREIGNLAMBDA;
 607         kernelLambda = lam_i;
 608         kernelDvdl   = dvdl_nb;
 609         gmx::ArrayRef<real> energygrp_elec =
 610                 foreignEnergyGroups_->energyGroupPairTerms[NonBondedEnergyTerms::CoulombSR];
 611         gmx::ArrayRef<real> energygrp_vdw =
 612                 foreignEnergyGroups_->energyGroupPairTerms[NonBondedEnergyTerms::LJSR];
 613
 614         for (gmx::index i = 0; i < 1 + enerd->foreignLambdaTerms.numLambdas(); i++)
 615         {
 616             std::fill(std::begin(dvdl_nb), std::end(dvdl_nb), 0);
 617             for (int j = 0; j < static_cast<int>(FreeEnergyPerturbationCouplingType::Count); j++)
 618             {
 619                 lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i - 1]);
 620             }
 621             foreignEnergyGroups_->clear();
 622 #pragma omp parallel for schedule(static) num_threads(nbl_fep.ssize())
 623             for (gmx::index th = 0; th < nbl_fep.ssize(); th++)
 624             {
 625                 try
 626                 {
 627                     gmx_nb_free_energy_kernel(*nbl_fep[th],
 628                                               coords,
 629                                               forceWithShiftForces,
 630                                               useSimd,
 631                                               ntype,
 632                                               rlist,
 633                                               ic,
 634                                               shiftvec,
 635                                               nbfp,
 636                                               nbfp_grid,
 637                                               chargeA,
 638                                               chargeB,
 639                                               typeA,
 640                                               typeB,
 641                                               kernelFlags,
 642                                               kernelLambda,
 643                                               kernelDvdl,
 644                                               energygrp_elec,
 645                                               energygrp_vdw,
 646                                               nrnb);
 647                 }
 648                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 649             }
 650             std::array<real, F_NRE> foreign_term = { 0 };
 651             sum_epot(*foreignEnergyGroups_, foreign_term.data());
 652             enerd->foreignLambdaTerms.accumulate(
 653                     i,
 654                     foreign_term[F_EPOT],
 655                     dvdl_nb[FreeEnergyPerturbationCouplingType::Vdw]
 656                             + dvdl_nb[FreeEnergyPerturbationCouplingType::Coul]);
 657         }
 658     }
 659     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::NonbondedFep);
 660 }