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