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