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/nb_kernel.h"
  42 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  43 #include "gromacs/math/vectypes.h"
  44 #include "gromacs/mdlib/enerdata_utils.h"
  45 #include "gromacs/mdlib/force.h"
  46 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  47 #include "gromacs/mdtypes/enerdata.h"
  48 #include "gromacs/mdtypes/forceoutput.h"
  49 #include "gromacs/mdtypes/forcerec.h"
  50 #include "gromacs/mdtypes/inputrec.h"
  51 #include "gromacs/mdtypes/interaction_const.h"
  52 #include "gromacs/mdtypes/md_enums.h"
  53 #include "gromacs/mdtypes/mdatom.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/gmxassert.h"
  60 #include "gromacs/utility/real.h"
  61
  62 #include "kernel_common.h"
  63 #include "nbnxm_gpu.h"
  64 #include "nbnxm_gpu_data_mgmt.h"
  65 #include "nbnxm_simd.h"
  66 #include "pairlistset.h"
  67 #include "pairlistsets.h"
  68 #include "kernels_reference/kernel_gpu_ref.h"
  69 #define INCLUDE_KERNELFUNCTION_TABLES
  70 #include "kernels_reference/kernel_ref.h"
  71 #ifdef GMX_NBNXN_SIMD_2XNN
  72 #    include "kernels_simd_2xmm/kernels.h"
  73 #endif
  74 #ifdef GMX_NBNXN_SIMD_4XN
  75 #    include "kernels_simd_4xm/kernels.h"
  76 #endif
  77 #undef INCLUDE_FUNCTION_TABLES
  78
  79 /*! \brief Clears the energy group output buffers
  80  *
  81  * \param[in,out] out  nbnxn kernel output struct
  82  */
  83 static void clearGroupEnergies(nbnxn_atomdata_output_t* out)
  84 {
  85     std::fill(out->Vvdw.begin(), out->Vvdw.end(), 0.0_real);
  86     std::fill(out->Vc.begin(), out->Vc.end(), 0.0_real);
  87     std::fill(out->VSvdw.begin(), out->VSvdw.end(), 0.0_real);
  88     std::fill(out->VSc.begin(), out->VSc.end(), 0.0_real);
  89 }
  90
  91 /*! \brief Reduce the group-pair energy buffers produced by a SIMD kernel
  92  * to single terms in the output buffers.
  93  *
  94  * The SIMD kernels produce a large number of energy buffer in SIMD registers
  95  * to avoid scattered reads and writes.
  96  *
  97  * \tparam        unrollj         The unroll size for j-particles in the SIMD kernel
  98  * \param[in]     numGroups       The number of energy groups
  99  * \param[in]     numGroups_2log  Log2 of numGroups, rounded up
 100  * \param[in,out] out             Struct with energy buffers
 101  */
 102 template<int unrollj>
 103 static void reduceGroupEnergySimdBuffers(int numGroups, int numGroups_2log, nbnxn_atomdata_output_t* out)
 104 {
 105     const int unrollj_half = unrollj / 2;
 106     /* Energies are stored in SIMD registers with size 2^numGroups_2log */
 107     const int numGroupsStorage = (1 << numGroups_2log);
 108
 109     const real* gmx_restrict vVdwSimd     = out->VSvdw.data();
 110     const real* gmx_restrict vCoulombSimd = out->VSc.data();
 111     real* gmx_restrict vVdw               = out->Vvdw.data();
 112     real* gmx_restrict vCoulomb           = out->Vc.data();
 113
 114     /* The size of the SIMD energy group buffer array is:
 115      * numGroups*numGroups*numGroupsStorage*unrollj_half*simd_width
 116      */
 117     for (int i = 0; i < numGroups; i++)
 118     {
 119         for (int j1 = 0; j1 < numGroups; j1++)
 120         {
 121             for (int j0 = 0; j0 < numGroups; j0++)
 122             {
 123                 int c = ((i * numGroups + j1) * numGroupsStorage + j0) * unrollj_half * unrollj;
 124                 for (int s = 0; s < unrollj_half; s++)
 125                 {
 126                     vVdw[i * numGroups + j0] += vVdwSimd[c + 0];
 127                     vVdw[i * numGroups + j1] += vVdwSimd[c + 1];
 128                     vCoulomb[i * numGroups + j0] += vCoulombSimd[c + 0];
 129                     vCoulomb[i * numGroups + j1] += vCoulombSimd[c + 1];
 130                     c += unrollj + 2;
 131                 }
 132             }
 133         }
 134     }
 135 }
 136
 137 /*! \brief Dispatches the non-bonded N versus M atom cluster CPU kernels.
 138  *
 139  * OpenMP parallelization is performed within this function.
 140  * Energy reduction, but not force and shift force reduction, is performed
 141  * within this function.
 142  *
 143  * \param[in]     pairlistSet   Pairlists with local or non-local interactions to compute
 144  * \param[in]     kernelSetup   The non-bonded kernel setup
 145  * \param[in,out] nbat          The atomdata for the interactions
 146  * \param[in]     ic            Non-bonded interaction constants
 147  * \param[in]     shiftVectors  The PBC shift vectors
 148  * \param[in]     stepWork      Flags that tell what to compute
 149  * \param[in]     clearF        Enum that tells if to clear the force output buffer
 150  * \param[out]    vCoulomb      Output buffer for Coulomb energies
 151  * \param[out]    vVdw          Output buffer for Van der Waals energies
 152  * \param[in]     wcycle        Pointer to cycle counting data structure.
 153  */
 154 static void nbnxn_kernel_cpu(const PairlistSet&         pairlistSet,
 155                              const Nbnxm::KernelSetup&  kernelSetup,
 156                              nbnxn_atomdata_t*          nbat,
 157                              const interaction_const_t& ic,
 158                              rvec*                      shiftVectors,
 159                              const gmx::StepWorkload&   stepWork,
 160                              int                        clearF,
 161                              real*                      vCoulomb,
 162                              real*                      vVdw,
 163                              gmx_wallcycle*             wcycle)
 164 {
 165
 166     int coulkt;
 167     if (EEL_RF(ic.eeltype) || ic.eeltype == eelCUT)
 168     {
 169         coulkt = coulktRF;
 170     }
 171     else
 172     {
 173         if (kernelSetup.ewaldExclusionType == Nbnxm::EwaldExclusionType::Table)
 174         {
 175             if (ic.rcoulomb == ic.rvdw)
 176             {
 177                 coulkt = coulktTAB;
 178             }
 179             else
 180             {
 181                 coulkt = coulktTAB_TWIN;
 182             }
 183         }
 184         else
 185         {
 186             if (ic.rcoulomb == ic.rvdw)
 187             {
 188                 coulkt = coulktEWALD;
 189             }
 190             else
 191             {
 192                 coulkt = coulktEWALD_TWIN;
 193             }
 194         }
 195     }
 196
 197     const nbnxn_atomdata_t::Params& nbatParams = nbat->params();
 198
 199     int vdwkt = 0;
 200     if (ic.vdwtype == evdwCUT)
 201     {
 202         switch (ic.vdw_modifier)
 203         {
 204             case eintmodNONE:
 205             case eintmodPOTSHIFT:
 206                 switch (nbatParams.ljCombinationRule)
 207                 {
 208                     case LJCombinationRule::Geometric: vdwkt = vdwktLJCUT_COMBGEOM; break;
 209                     case LJCombinationRule::LorentzBerthelot: vdwkt = vdwktLJCUT_COMBLB; break;
 210                     case LJCombinationRule::None: vdwkt = vdwktLJCUT_COMBNONE; break;
 211                     default: GMX_RELEASE_ASSERT(false, "Unknown combination rule");
 212                 }
 213                 break;
 214             case eintmodFORCESWITCH: vdwkt = vdwktLJFORCESWITCH; break;
 215             case eintmodPOTSWITCH: vdwkt = vdwktLJPOTSWITCH; break;
 216             default: GMX_RELEASE_ASSERT(false, "Unsupported VdW interaction modifier");
 217         }
 218     }
 219     else if (ic.vdwtype == evdwPME)
 220     {
 221         if (ic.ljpme_comb_rule == eljpmeGEOM)
 222         {
 223             vdwkt = vdwktLJEWALDCOMBGEOM;
 224         }
 225         else
 226         {
 227             vdwkt = vdwktLJEWALDCOMBLB;
 228             /* At setup we (should have) selected the C reference kernel */
 229             GMX_RELEASE_ASSERT(kernelSetup.kernelType == Nbnxm::KernelType::Cpu4x4_PlainC,
 230                                "Only the C reference nbnxn SIMD kernel supports LJ-PME with LB "
 231                                "combination rules");
 232         }
 233     }
 234     else
 235     {
 236         GMX_RELEASE_ASSERT(false, "Unsupported VdW interaction type");
 237     }
 238
 239     gmx::ArrayRef<const NbnxnPairlistCpu> pairlists = pairlistSet.cpuLists();
 240
 241     int gmx_unused nthreads = gmx_omp_nthreads_get(emntNonbonded);
 242     wallcycle_sub_start(wcycle, ewcsNONBONDED_CLEAR);
 243 #pragma omp parallel for schedule(static) num_threads(nthreads)
 244     for (gmx::index nb = 0; nb < pairlists.ssize(); nb++)
 245     {
 246         // Presently, the kernels do not call C++ code that can throw,
 247         // so no need for a try/catch pair in this OpenMP region.
 248         nbnxn_atomdata_output_t* out = &nbat->out[nb];
 249
 250         if (clearF == enbvClearFYes)
 251         {
 252             clearForceBuffer(nbat, nb);
 253
 254             clear_fshift(out->fshift.data());
 255         }
 256
 257         if (nb == 0)
 258         {
 259             wallcycle_sub_stop(wcycle, ewcsNONBONDED_CLEAR);
 260             wallcycle_sub_start(wcycle, ewcsNONBONDED_KERNEL);
 261         }
 262
 263         // TODO: Change to reference
 264         const NbnxnPairlistCpu* pairlist = &pairlists[nb];
 265
 266         if (!stepWork.computeEnergy)
 267         {
 268             /* Don't calculate energies */
 269             switch (kernelSetup.kernelType)
 270             {
 271                 case Nbnxm::KernelType::Cpu4x4_PlainC:
 272                     nbnxn_kernel_noener_ref[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 273                     break;
 274 #ifdef GMX_NBNXN_SIMD_2XNN
 275                 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 276                     nbnxm_kernel_noener_simd_2xmm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 277                     break;
 278 #endif
 279 #ifdef GMX_NBNXN_SIMD_4XN
 280                 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 281                     nbnxm_kernel_noener_simd_4xm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 282                     break;
 283 #endif
 284                 default: GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 285             }
 286         }
 287         else if (out->Vvdw.size() == 1)
 288         {
 289             /* A single energy group (pair) */
 290             out->Vvdw[0] = 0;
 291             out->Vc[0]   = 0;
 292
 293             switch (kernelSetup.kernelType)
 294             {
 295                 case Nbnxm::KernelType::Cpu4x4_PlainC:
 296                     nbnxn_kernel_ener_ref[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 297                     break;
 298 #ifdef GMX_NBNXN_SIMD_2XNN
 299                 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 300                     nbnxm_kernel_ener_simd_2xmm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 301                     break;
 302 #endif
 303 #ifdef GMX_NBNXN_SIMD_4XN
 304                 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 305                     nbnxm_kernel_ener_simd_4xm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 306                     break;
 307 #endif
 308                 default: GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 309             }
 310         }
 311         else
 312         {
 313             /* Calculate energy group contributions */
 314             clearGroupEnergies(out);
 315
 316             int unrollj = 0;
 317
 318             switch (kernelSetup.kernelType)
 319             {
 320                 case Nbnxm::KernelType::Cpu4x4_PlainC:
 321                     unrollj = c_nbnxnCpuIClusterSize;
 322                     nbnxn_kernel_energrp_ref[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 323                     break;
 324 #ifdef GMX_NBNXN_SIMD_2XNN
 325                 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 326                     unrollj = GMX_SIMD_REAL_WIDTH / 2;
 327                     nbnxm_kernel_energrp_simd_2xmm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 328                     break;
 329 #endif
 330 #ifdef GMX_NBNXN_SIMD_4XN
 331                 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 332                     unrollj = GMX_SIMD_REAL_WIDTH;
 333                     nbnxm_kernel_energrp_simd_4xm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 334                     break;
 335 #endif
 336                 default: GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 337             }
 338
 339             if (kernelSetup.kernelType != Nbnxm::KernelType::Cpu4x4_PlainC)
 340             {
 341                 switch (unrollj)
 342                 {
 343                     case 2:
 344                         reduceGroupEnergySimdBuffers<2>(nbatParams.nenergrp, nbatParams.neg_2log, out);
 345                         break;
 346                     case 4:
 347                         reduceGroupEnergySimdBuffers<4>(nbatParams.nenergrp, nbatParams.neg_2log, out);
 348                         break;
 349                     case 8:
 350                         reduceGroupEnergySimdBuffers<8>(nbatParams.nenergrp, nbatParams.neg_2log, out);
 351                         break;
 352                     default: GMX_RELEASE_ASSERT(false, "Unsupported j-unroll size");
 353                 }
 354             }
 355         }
 356     }
 357     wallcycle_sub_stop(wcycle, ewcsNONBONDED_KERNEL);
 358
 359     if (stepWork.computeEnergy)
 360     {
 361         reduce_energies_over_lists(nbat, pairlists.ssize(), vVdw, vCoulomb);
 362     }
 363 }
 364
 365 static void accountFlops(t_nrnb*                    nrnb,
 366                          const PairlistSet&         pairlistSet,
 367                          const nonbonded_verlet_t&  nbv,
 368                          const interaction_const_t& ic,
 369                          const gmx::StepWorkload&   stepWork)
 370 {
 371     const bool usingGpuKernels = nbv.useGpu();
 372
 373     int enr_nbnxn_kernel_ljc;
 374     if (EEL_RF(ic.eeltype) || ic.eeltype == eelCUT)
 375     {
 376         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
 377     }
 378     else if ((!usingGpuKernels && nbv.kernelSetup().ewaldExclusionType == Nbnxm::EwaldExclusionType::Analytical)
 379              || (usingGpuKernels && Nbnxm::gpu_is_kernel_ewald_analytical(nbv.gpu_nbv)))
 380     {
 381         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
 382     }
 383     else
 384     {
 385         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
 386     }
 387     int enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
 388     if (stepWork.computeEnergy)
 389     {
 390         /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 391         enr_nbnxn_kernel_ljc += 1;
 392         enr_nbnxn_kernel_lj += 1;
 393     }
 394
 395     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc, pairlistSet.natpair_ljq_);
 396     inc_nrnb(nrnb, enr_nbnxn_kernel_lj, pairlistSet.natpair_lj_);
 397     /* The Coulomb-only kernels are offset -eNR_NBNXN_LJ_RF+eNR_NBNXN_RF */
 398     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc - eNR_NBNXN_LJ_RF + eNR_NBNXN_RF, pairlistSet.natpair_q_);
 399
 400     if (ic.vdw_modifier == eintmodFORCESWITCH)
 401     {
 402         /* We add up the switch cost separately */
 403         inc_nrnb(nrnb,
 404                  eNR_NBNXN_ADD_LJ_FSW + (stepWork.computeEnergy ? 1 : 0),
 405                  pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_);
 406     }
 407     if (ic.vdw_modifier == eintmodPOTSWITCH)
 408     {
 409         /* We add up the switch cost separately */
 410         inc_nrnb(nrnb,
 411                  eNR_NBNXN_ADD_LJ_PSW + (stepWork.computeEnergy ? 1 : 0),
 412                  pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_);
 413     }
 414     if (ic.vdwtype == evdwPME)
 415     {
 416         /* We add up the LJ Ewald cost separately */
 417         inc_nrnb(nrnb,
 418                  eNR_NBNXN_ADD_LJ_EWALD + (stepWork.computeEnergy ? 1 : 0),
 419                  pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_);
 420     }
 421 }
 422
 423 void nonbonded_verlet_t::dispatchNonbondedKernel(gmx::InteractionLocality   iLocality,
 424                                                  const interaction_const_t& ic,
 425                                                  const gmx::StepWorkload&   stepWork,
 426                                                  int                        clearF,
 427                                                  const t_forcerec&          fr,
 428                                                  gmx_enerdata_t*            enerd,
 429                                                  t_nrnb*                    nrnb)
 430 {
 431     const PairlistSet& pairlistSet = pairlistSets().pairlistSet(iLocality);
 432
 433     switch (kernelSetup().kernelType)
 434     {
 435         case Nbnxm::KernelType::Cpu4x4_PlainC:
 436         case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 437         case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 438             nbnxn_kernel_cpu(pairlistSet,
 439                              kernelSetup(),
 440                              nbat.get(),
 441                              ic,
 442                              fr.shift_vec,
 443                              stepWork,
 444                              clearF,
 445                              enerd->grpp.ener[egCOULSR].data(),
 446                              fr.bBHAM ? enerd->grpp.ener[egBHAMSR].data() : enerd->grpp.ener[egLJSR].data(),
 447                              wcycle_);
 448             break;
 449
 450         case Nbnxm::KernelType::Gpu8x8x8:
 451             Nbnxm::gpu_launch_kernel(gpu_nbv, stepWork, iLocality);
 452             break;
 453
 454         case Nbnxm::KernelType::Cpu8x8x8_PlainC:
 455             nbnxn_kernel_gpu_ref(
 456                     pairlistSet.gpuList(),
 457                     nbat.get(),
 458                     &ic,
 459                     fr.shift_vec,
 460                     stepWork,
 461                     clearF,
 462                     nbat->out[0].f,
 463                     nbat->out[0].fshift.data(),
 464                     enerd->grpp.ener[egCOULSR].data(),
 465                     fr.bBHAM ? enerd->grpp.ener[egBHAMSR].data() : enerd->grpp.ener[egLJSR].data());
 466             break;
 467
 468         default: GMX_RELEASE_ASSERT(false, "Invalid nonbonded kernel type passed!");
 469     }
 470
 471     accountFlops(nrnb, pairlistSet, *this, ic, stepWork);
 472 }
 473
 474 void nonbonded_verlet_t::dispatchFreeEnergyKernel(gmx::InteractionLocality   iLocality,
 475                                                   const t_forcerec*          fr,
 476                                                   rvec                       x[],
 477                                                   gmx::ForceWithShiftForces* forceWithShiftForces,
 478                                                   const t_mdatoms&           mdatoms,
 479                                                   t_lambda*                  fepvals,
 480                                                   gmx::ArrayRef<real const>  lambda,
 481                                                   gmx_enerdata_t*            enerd,
 482                                                   const gmx::StepWorkload&   stepWork,
 483                                                   t_nrnb*                    nrnb)
 484 {
 485     const auto nbl_fep = pairlistSets().pairlistSet(iLocality).fepLists();
 486
 487     /* When the first list is empty, all are empty and there is nothing to do */
 488     if (!pairlistSets().params().haveFep || nbl_fep[0]->nrj == 0)
 489     {
 490         return;
 491     }
 492
 493     int donb_flags = 0;
 494     /* Add short-range interactions */
 495     donb_flags |= GMX_NONBONDED_DO_SR;
 496
 497     if (stepWork.computeForces)
 498     {
 499         donb_flags |= GMX_NONBONDED_DO_FORCE;
 500     }
 501     if (stepWork.computeVirial)
 502     {
 503         donb_flags |= GMX_NONBONDED_DO_SHIFTFORCE;
 504     }
 505     if (stepWork.computeEnergy)
 506     {
 507         donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
 508     }
 509
 510     nb_kernel_data_t kernel_data;
 511     real             dvdl_nb[efptNR] = { 0 };
 512     kernel_data.flags                = donb_flags;
 513     kernel_data.lambda               = lambda.data();
 514     kernel_data.dvdl                 = dvdl_nb;
 515
 516     kernel_data.energygrp_elec = enerd->grpp.ener[egCOULSR].data();
 517     kernel_data.energygrp_vdw  = enerd->grpp.ener[egLJSR].data();
 518
 519     GMX_ASSERT(gmx_omp_nthreads_get(emntNonbonded) == nbl_fep.ssize(),
 520                "Number of lists should be same as number of NB threads");
 521
 522     wallcycle_sub_start(wcycle_, ewcsNONBONDED_FEP);
 523 #pragma omp parallel for schedule(static) num_threads(nbl_fep.ssize())
 524     for (gmx::index th = 0; th < nbl_fep.ssize(); th++)
 525     {
 526         try
 527         {
 528             gmx_nb_free_energy_kernel(
 529                     nbl_fep[th].get(), x, forceWithShiftForces, fr, &mdatoms, &kernel_data, nrnb);
 530         }
 531         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 532     }
 533
 534     if (fepvals->sc_alpha != 0)
 535     {
 536         enerd->dvdl_nonlin[efptVDW] += dvdl_nb[efptVDW];
 537         enerd->dvdl_nonlin[efptCOUL] += dvdl_nb[efptCOUL];
 538     }
 539     else
 540     {
 541         enerd->dvdl_lin[efptVDW] += dvdl_nb[efptVDW];
 542         enerd->dvdl_lin[efptCOUL] += dvdl_nb[efptCOUL];
 543     }
 544
 545     /* If we do foreign lambda and we have soft-core interactions
 546      * we have to recalculate the (non-linear) energies contributions.
 547      */
 548     if (fepvals->n_lambda > 0 && stepWork.computeDhdl && fepvals->sc_alpha != 0)
 549     {
 550         real lam_i[efptNR];
 551         kernel_data.flags = (donb_flags & ~(GMX_NONBONDED_DO_FORCE | GMX_NONBONDED_DO_SHIFTFORCE))
 552                             | GMX_NONBONDED_DO_FOREIGNLAMBDA;
 553         kernel_data.lambda         = lam_i;
 554         kernel_data.dvdl           = dvdl_nb;
 555         kernel_data.energygrp_elec = enerd->foreign_grpp.ener[egCOULSR].data();
 556         kernel_data.energygrp_vdw  = enerd->foreign_grpp.ener[egLJSR].data();
 557
 558         for (gmx::index i = 0; i < 1 + enerd->foreignLambdaTerms.numLambdas(); i++)
 559         {
 560             std::fill(std::begin(dvdl_nb), std::end(dvdl_nb), 0);
 561             for (int j = 0; j < efptNR; j++)
 562             {
 563                 lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i - 1]);
 564             }
 565             reset_foreign_enerdata(enerd);
 566 #pragma omp parallel for schedule(static) num_threads(nbl_fep.ssize())
 567             for (gmx::index th = 0; th < nbl_fep.ssize(); th++)
 568             {
 569                 try
 570                 {
 571                     gmx_nb_free_energy_kernel(
 572                             nbl_fep[th].get(), x, forceWithShiftForces, fr, &mdatoms, &kernel_data, nrnb);
 573                 }
 574                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 575             }
 576
 577             sum_epot(enerd->foreign_grpp, enerd->foreign_term);
 578             enerd->foreignLambdaTerms.accumulate(
 579                     i, enerd->foreign_term[F_EPOT], dvdl_nb[efptVDW] + dvdl_nb[efptCOUL]);
 580         }
 581     }
 582     wallcycle_sub_stop(wcycle_, ewcsNONBONDED_FEP);
 583 }