src/gromacs/nbnxm/kerneldispatch.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35
  36 #include "gmxpre.h"
  37
  38 #include "gromacs/gmxlib/nrnb.h"
  39 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  40 #include "gromacs/gmxlib/nonbonded/nb_kernel.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/inputrec.h"
  49 #include "gromacs/mdtypes/interaction_const.h"
  50 #include "gromacs/mdtypes/md_enums.h"
  51 #include "gromacs/mdtypes/mdatom.h"
  52 #include "gromacs/mdtypes/simulation_workload.h"
  53 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  54 #include "gromacs/nbnxm/nbnxm.h"
  55 #include "gromacs/simd/simd.h"
  56 #include "gromacs/timing/wallcycle.h"
  57 #include "gromacs/utility/gmxassert.h"
  58 #include "gromacs/utility/real.h"
  59
  60 #include "kernel_common.h"
  61 #include "nbnxm_simd.h"
  62 #include "pairlistset.h"
  63 #include "pairlistsets.h"
  64 #include "kernels_reference/kernel_gpu_ref.h"
  65 #define INCLUDE_KERNELFUNCTION_TABLES
  66 #include "kernels_reference/kernel_ref.h"
  67 #ifdef GMX_NBNXN_SIMD_2XNN
  68 #    include "kernels_simd_2xmm/kernels.h"
  69 #endif
  70 #ifdef GMX_NBNXN_SIMD_4XN
  71 #    include "kernels_simd_4xm/kernels.h"
  72 #endif
  73 #undef INCLUDE_FUNCTION_TABLES
  74
  75 /*! \brief Clears the energy group output buffers
  76  *
  77  * \param[in,out] out  nbnxn kernel output struct
  78  */
  79 static void clearGroupEnergies(nbnxn_atomdata_output_t* out)
  80 {
  81     std::fill(out->Vvdw.begin(), out->Vvdw.end(), 0.0_real);
  82     std::fill(out->Vc.begin(), out->Vc.end(), 0.0_real);
  83     std::fill(out->VSvdw.begin(), out->VSvdw.end(), 0.0_real);
  84     std::fill(out->VSc.begin(), out->VSc.end(), 0.0_real);
  85 }
  86
  87 /*! \brief Reduce the group-pair energy buffers produced by a SIMD kernel
  88  * to single terms in the output buffers.
  89  *
  90  * The SIMD kernels produce a large number of energy buffer in SIMD registers
  91  * to avoid scattered reads and writes.
  92  *
  93  * \tparam        unrollj         The unroll size for j-particles in the SIMD kernel
  94  * \param[in]     numGroups       The number of energy groups
  95  * \param[in]     numGroups_2log  Log2 of numGroups, rounded up
  96  * \param[in,out] out             Struct with energy buffers
  97  */
  98 template<int unrollj>
  99 static void reduceGroupEnergySimdBuffers(int numGroups, int numGroups_2log, nbnxn_atomdata_output_t* out)
 100 {
 101     const int unrollj_half = unrollj / 2;
 102     /* Energies are stored in SIMD registers with size 2^numGroups_2log */
 103     const int numGroupsStorage = (1 << numGroups_2log);
 104
 105     const real* gmx_restrict vVdwSimd     = out->VSvdw.data();
 106     const real* gmx_restrict vCoulombSimd = out->VSc.data();
 107     real* gmx_restrict vVdw               = out->Vvdw.data();
 108     real* gmx_restrict vCoulomb           = out->Vc.data();
 109
 110     /* The size of the SIMD energy group buffer array is:
 111      * numGroups*numGroups*numGroupsStorage*unrollj_half*simd_width
 112      */
 113     for (int i = 0; i < numGroups; i++)
 114     {
 115         for (int j1 = 0; j1 < numGroups; j1++)
 116         {
 117             for (int j0 = 0; j0 < numGroups; j0++)
 118             {
 119                 int c = ((i * numGroups + j1) * numGroupsStorage + j0) * unrollj_half * unrollj;
 120                 for (int s = 0; s < unrollj_half; s++)
 121                 {
 122                     vVdw[i * numGroups + j0] += vVdwSimd[c + 0];
 123                     vVdw[i * numGroups + j1] += vVdwSimd[c + 1];
 124                     vCoulomb[i * numGroups + j0] += vCoulombSimd[c + 0];
 125                     vCoulomb[i * numGroups + j1] += vCoulombSimd[c + 1];
 126                     c += unrollj + 2;
 127                 }
 128             }
 129         }
 130     }
 131 }
 132
 133 /*! \brief Dispatches the non-bonded N versus M atom cluster CPU kernels.
 134  *
 135  * OpenMP parallelization is performed within this function.
 136  * Energy reduction, but not force and shift force reduction, is performed
 137  * within this function.
 138  *
 139  * \param[in]     pairlistSet   Pairlists with local or non-local interactions to compute
 140  * \param[in]     kernelSetup   The non-bonded kernel setup
 141  * \param[in,out] nbat          The atomdata for the interactions
 142  * \param[in]     ic            Non-bonded interaction constants
 143  * \param[in]     shiftVectors  The PBC shift vectors
 144  * \param[in]     stepWork      Flags that tell what to compute
 145  * \param[in]     clearF        Enum that tells if to clear the force output buffer
 146  * \param[out]    vCoulomb      Output buffer for Coulomb energies
 147  * \param[out]    vVdw          Output buffer for Van der Waals energies
 148  * \param[in]     wcycle        Pointer to cycle counting data structure.
 149  */
 150 static void nbnxn_kernel_cpu(const PairlistSet&         pairlistSet,
 151                              const Nbnxm::KernelSetup&  kernelSetup,
 152                              nbnxn_atomdata_t*          nbat,
 153                              const interaction_const_t& ic,
 154                              rvec*                      shiftVectors,
 155                              const gmx::StepWorkload&   stepWork,
 156                              int                        clearF,
 157                              real*                      vCoulomb,
 158                              real*                      vVdw,
 159                              gmx_wallcycle*             wcycle)
 160 {
 161
 162     int coulkt;
 163     if (EEL_RF(ic.eeltype) || ic.eeltype == eelCUT)
 164     {
 165         coulkt = coulktRF;
 166     }
 167     else
 168     {
 169         if (kernelSetup.ewaldExclusionType == Nbnxm::EwaldExclusionType::Table)
 170         {
 171             if (ic.rcoulomb == ic.rvdw)
 172             {
 173                 coulkt = coulktTAB;
 174             }
 175             else
 176             {
 177                 coulkt = coulktTAB_TWIN;
 178             }
 179         }
 180         else
 181         {
 182             if (ic.rcoulomb == ic.rvdw)
 183             {
 184                 coulkt = coulktEWALD;
 185             }
 186             else
 187             {
 188                 coulkt = coulktEWALD_TWIN;
 189             }
 190         }
 191     }
 192
 193     const nbnxn_atomdata_t::Params& nbatParams = nbat->params();
 194
 195     int vdwkt = 0;
 196     if (ic.vdwtype == evdwCUT)
 197     {
 198         switch (ic.vdw_modifier)
 199         {
 200             case eintmodNONE:
 201             case eintmodPOTSHIFT:
 202                 switch (nbatParams.comb_rule)
 203                 {
 204                     case ljcrGEOM: vdwkt = vdwktLJCUT_COMBGEOM; break;
 205                     case ljcrLB: vdwkt = vdwktLJCUT_COMBLB; break;
 206                     case ljcrNONE: vdwkt = vdwktLJCUT_COMBNONE; break;
 207                     default: GMX_RELEASE_ASSERT(false, "Unknown combination rule");
 208                 }
 209                 break;
 210             case eintmodFORCESWITCH: vdwkt = vdwktLJFORCESWITCH; break;
 211             case eintmodPOTSWITCH: vdwkt = vdwktLJPOTSWITCH; break;
 212             default: GMX_RELEASE_ASSERT(false, "Unsupported VdW interaction modifier");
 213         }
 214     }
 215     else if (ic.vdwtype == evdwPME)
 216     {
 217         if (ic.ljpme_comb_rule == eljpmeGEOM)
 218         {
 219             vdwkt = vdwktLJEWALDCOMBGEOM;
 220         }
 221         else
 222         {
 223             vdwkt = vdwktLJEWALDCOMBLB;
 224             /* At setup we (should have) selected the C reference kernel */
 225             GMX_RELEASE_ASSERT(kernelSetup.kernelType == Nbnxm::KernelType::Cpu4x4_PlainC,
 226                                "Only the C reference nbnxn SIMD kernel supports LJ-PME with LB "
 227                                "combination rules");
 228         }
 229     }
 230     else
 231     {
 232         GMX_RELEASE_ASSERT(false, "Unsupported VdW interaction type");
 233     }
 234
 235     gmx::ArrayRef<const NbnxnPairlistCpu> pairlists = pairlistSet.cpuLists();
 236
 237     int gmx_unused nthreads = gmx_omp_nthreads_get(emntNonbonded);
 238     wallcycle_sub_start(wcycle, ewcsNONBONDED_CLEAR);
 239 #pragma omp parallel for schedule(static) num_threads(nthreads)
 240     for (gmx::index nb = 0; nb < pairlists.ssize(); nb++)
 241     {
 242         // Presently, the kernels do not call C++ code that can throw,
 243         // so no need for a try/catch pair in this OpenMP region.
 244         nbnxn_atomdata_output_t* out = &nbat->out[nb];
 245
 246         if (clearF == enbvClearFYes)
 247         {
 248             clearForceBuffer(nbat, nb);
 249
 250             clear_fshift(out->fshift.data());
 251         }
 252
 253         if (nb == 0)
 254         {
 255             wallcycle_sub_stop(wcycle, ewcsNONBONDED_CLEAR);
 256             wallcycle_sub_start(wcycle, ewcsNONBONDED_KERNEL);
 257         }
 258
 259         // TODO: Change to reference
 260         const NbnxnPairlistCpu* pairlist = &pairlists[nb];
 261
 262         if (!stepWork.computeEnergy)
 263         {
 264             /* Don't calculate energies */
 265             switch (kernelSetup.kernelType)
 266             {
 267                 case Nbnxm::KernelType::Cpu4x4_PlainC:
 268                     nbnxn_kernel_noener_ref[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 269                     break;
 270 #ifdef GMX_NBNXN_SIMD_2XNN
 271                 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 272                     nbnxm_kernel_noener_simd_2xmm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 273                     break;
 274 #endif
 275 #ifdef GMX_NBNXN_SIMD_4XN
 276                 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 277                     nbnxm_kernel_noener_simd_4xm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 278                     break;
 279 #endif
 280                 default: GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 281             }
 282         }
 283         else if (out->Vvdw.size() == 1)
 284         {
 285             /* A single energy group (pair) */
 286             out->Vvdw[0] = 0;
 287             out->Vc[0]   = 0;
 288
 289             switch (kernelSetup.kernelType)
 290             {
 291                 case Nbnxm::KernelType::Cpu4x4_PlainC:
 292                     nbnxn_kernel_ener_ref[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 293                     break;
 294 #ifdef GMX_NBNXN_SIMD_2XNN
 295                 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 296                     nbnxm_kernel_ener_simd_2xmm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 297                     break;
 298 #endif
 299 #ifdef GMX_NBNXN_SIMD_4XN
 300                 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 301                     nbnxm_kernel_ener_simd_4xm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 302                     break;
 303 #endif
 304                 default: GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 305             }
 306         }
 307         else
 308         {
 309             /* Calculate energy group contributions */
 310             clearGroupEnergies(out);
 311
 312             int unrollj = 0;
 313
 314             switch (kernelSetup.kernelType)
 315             {
 316                 case Nbnxm::KernelType::Cpu4x4_PlainC:
 317                     unrollj = c_nbnxnCpuIClusterSize;
 318                     nbnxn_kernel_energrp_ref[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 319                     break;
 320 #ifdef GMX_NBNXN_SIMD_2XNN
 321                 case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 322                     unrollj = GMX_SIMD_REAL_WIDTH / 2;
 323                     nbnxm_kernel_energrp_simd_2xmm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 324                     break;
 325 #endif
 326 #ifdef GMX_NBNXN_SIMD_4XN
 327                 case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 328                     unrollj = GMX_SIMD_REAL_WIDTH;
 329                     nbnxm_kernel_energrp_simd_4xm[coulkt][vdwkt](pairlist, nbat, &ic, shiftVectors, out);
 330                     break;
 331 #endif
 332                 default: GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 333             }
 334
 335             if (kernelSetup.kernelType != Nbnxm::KernelType::Cpu4x4_PlainC)
 336             {
 337                 switch (unrollj)
 338                 {
 339                     case 2:
 340                         reduceGroupEnergySimdBuffers<2>(nbatParams.nenergrp, nbatParams.neg_2log, out);
 341                         break;
 342                     case 4:
 343                         reduceGroupEnergySimdBuffers<4>(nbatParams.nenergrp, nbatParams.neg_2log, out);
 344                         break;
 345                     case 8:
 346                         reduceGroupEnergySimdBuffers<8>(nbatParams.nenergrp, nbatParams.neg_2log, out);
 347                         break;
 348                     default: GMX_RELEASE_ASSERT(false, "Unsupported j-unroll size");
 349                 }
 350             }
 351         }
 352     }
 353     wallcycle_sub_stop(wcycle, ewcsNONBONDED_KERNEL);
 354
 355     if (stepWork.computeEnergy)
 356     {
 357         reduce_energies_over_lists(nbat, pairlists.ssize(), vVdw, vCoulomb);
 358     }
 359 }
 360
 361 static void accountFlops(t_nrnb*                    nrnb,
 362                          const PairlistSet&         pairlistSet,
 363                          const nonbonded_verlet_t&  nbv,
 364                          const interaction_const_t& ic,
 365                          const gmx::StepWorkload&   stepWork)
 366 {
 367     const bool usingGpuKernels = nbv.useGpu();
 368
 369     int enr_nbnxn_kernel_ljc;
 370     if (EEL_RF(ic.eeltype) || ic.eeltype == eelCUT)
 371     {
 372         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
 373     }
 374     else if ((!usingGpuKernels && nbv.kernelSetup().ewaldExclusionType == Nbnxm::EwaldExclusionType::Analytical)
 375              || (usingGpuKernels && Nbnxm::gpu_is_kernel_ewald_analytical(nbv.gpu_nbv)))
 376     {
 377         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
 378     }
 379     else
 380     {
 381         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
 382     }
 383     int enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
 384     if (stepWork.computeEnergy)
 385     {
 386         /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 387         enr_nbnxn_kernel_ljc += 1;
 388         enr_nbnxn_kernel_lj += 1;
 389     }
 390
 391     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc, pairlistSet.natpair_ljq_);
 392     inc_nrnb(nrnb, enr_nbnxn_kernel_lj, pairlistSet.natpair_lj_);
 393     /* The Coulomb-only kernels are offset -eNR_NBNXN_LJ_RF+eNR_NBNXN_RF */
 394     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc - eNR_NBNXN_LJ_RF + eNR_NBNXN_RF, pairlistSet.natpair_q_);
 395
 396     if (ic.vdw_modifier == eintmodFORCESWITCH)
 397     {
 398         /* We add up the switch cost separately */
 399         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_FSW + (stepWork.computeEnergy ? 1 : 0),
 400                  pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_);
 401     }
 402     if (ic.vdw_modifier == eintmodPOTSWITCH)
 403     {
 404         /* We add up the switch cost separately */
 405         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_PSW + (stepWork.computeEnergy ? 1 : 0),
 406                  pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_);
 407     }
 408     if (ic.vdwtype == evdwPME)
 409     {
 410         /* We add up the LJ Ewald cost separately */
 411         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_EWALD + (stepWork.computeEnergy ? 1 : 0),
 412                  pairlistSet.natpair_ljq_ + pairlistSet.natpair_lj_);
 413     }
 414 }
 415
 416 void nonbonded_verlet_t::dispatchNonbondedKernel(gmx::InteractionLocality   iLocality,
 417                                                  const interaction_const_t& ic,
 418                                                  const gmx::StepWorkload&   stepWork,
 419                                                  int                        clearF,
 420                                                  const t_forcerec&          fr,
 421                                                  gmx_enerdata_t*            enerd,
 422                                                  t_nrnb*                    nrnb)
 423 {
 424     const PairlistSet& pairlistSet = pairlistSets().pairlistSet(iLocality);
 425
 426     switch (kernelSetup().kernelType)
 427     {
 428         case Nbnxm::KernelType::Cpu4x4_PlainC:
 429         case Nbnxm::KernelType::Cpu4xN_Simd_4xN:
 430         case Nbnxm::KernelType::Cpu4xN_Simd_2xNN:
 431             nbnxn_kernel_cpu(pairlistSet, kernelSetup(), nbat.get(), ic, fr.shift_vec, stepWork,
 432                              clearF, enerd->grpp.ener[egCOULSR].data(),
 433                              fr.bBHAM ? enerd->grpp.ener[egBHAMSR].data() : enerd->grpp.ener[egLJSR].data(),
 434                              wcycle_);
 435             break;
 436
 437         case Nbnxm::KernelType::Gpu8x8x8:
 438             Nbnxm::gpu_launch_kernel(gpu_nbv, stepWork, iLocality);
 439             break;
 440
 441         case Nbnxm::KernelType::Cpu8x8x8_PlainC:
 442             nbnxn_kernel_gpu_ref(
 443                     pairlistSet.gpuList(), nbat.get(), &ic, fr.shift_vec, stepWork, clearF,
 444                     nbat->out[0].f, nbat->out[0].fshift.data(), enerd->grpp.ener[egCOULSR].data(),
 445                     fr.bBHAM ? enerd->grpp.ener[egBHAMSR].data() : enerd->grpp.ener[egLJSR].data());
 446             break;
 447
 448         default: GMX_RELEASE_ASSERT(false, "Invalid nonbonded kernel type passed!");
 449     }
 450
 451     accountFlops(nrnb, pairlistSet, *this, ic, stepWork);
 452 }
 453
 454 void nonbonded_verlet_t::dispatchFreeEnergyKernel(gmx::InteractionLocality   iLocality,
 455                                                   const t_forcerec*          fr,
 456                                                   rvec                       x[],
 457                                                   gmx::ForceWithShiftForces* forceWithShiftForces,
 458                                                   const t_mdatoms&           mdatoms,
 459                                                   t_lambda*                  fepvals,
 460                                                   real*                      lambda,
 461                                                   gmx_enerdata_t*            enerd,
 462                                                   const gmx::StepWorkload&   stepWork,
 463                                                   t_nrnb*                    nrnb)
 464 {
 465     const auto nbl_fep = pairlistSets().pairlistSet(iLocality).fepLists();
 466
 467     /* When the first list is empty, all are empty and there is nothing to do */
 468     if (!pairlistSets().params().haveFep || nbl_fep[0]->nrj == 0)
 469     {
 470         return;
 471     }
 472
 473     int donb_flags = 0;
 474     /* Add short-range interactions */
 475     donb_flags |= GMX_NONBONDED_DO_SR;
 476
 477     if (stepWork.computeForces)
 478     {
 479         donb_flags |= GMX_NONBONDED_DO_FORCE;
 480     }
 481     if (stepWork.computeVirial)
 482     {
 483         donb_flags |= GMX_NONBONDED_DO_SHIFTFORCE;
 484     }
 485     if (stepWork.computeEnergy)
 486     {
 487         donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
 488     }
 489
 490     nb_kernel_data_t kernel_data;
 491     real             dvdl_nb[efptNR] = { 0 };
 492     kernel_data.flags                = donb_flags;
 493     kernel_data.lambda               = lambda;
 494     kernel_data.dvdl                 = dvdl_nb;
 495
 496     kernel_data.energygrp_elec = enerd->grpp.ener[egCOULSR].data();
 497     kernel_data.energygrp_vdw  = enerd->grpp.ener[egLJSR].data();
 498
 499     GMX_ASSERT(gmx_omp_nthreads_get(emntNonbonded) == nbl_fep.ssize(),
 500                "Number of lists should be same as number of NB threads");
 501
 502     wallcycle_sub_start(wcycle_, ewcsNONBONDED_FEP);
 503 #pragma omp parallel for schedule(static) num_threads(nbl_fep.ssize())
 504     for (gmx::index th = 0; th < nbl_fep.ssize(); th++)
 505     {
 506         try
 507         {
 508             gmx_nb_free_energy_kernel(nbl_fep[th].get(), x, forceWithShiftForces, fr, &mdatoms,
 509                                       &kernel_data, nrnb);
 510         }
 511         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 512     }
 513
 514     if (fepvals->sc_alpha != 0)
 515     {
 516         enerd->dvdl_nonlin[efptVDW] += dvdl_nb[efptVDW];
 517         enerd->dvdl_nonlin[efptCOUL] += dvdl_nb[efptCOUL];
 518     }
 519     else
 520     {
 521         enerd->dvdl_lin[efptVDW] += dvdl_nb[efptVDW];
 522         enerd->dvdl_lin[efptCOUL] += dvdl_nb[efptCOUL];
 523     }
 524
 525     /* If we do foreign lambda and we have soft-core interactions
 526      * we have to recalculate the (non-linear) energies contributions.
 527      */
 528     if (fepvals->n_lambda > 0 && stepWork.computeDhdl && fepvals->sc_alpha != 0)
 529     {
 530         real lam_i[efptNR];
 531         kernel_data.flags = (donb_flags & ~(GMX_NONBONDED_DO_FORCE | GMX_NONBONDED_DO_SHIFTFORCE))
 532                             | GMX_NONBONDED_DO_FOREIGNLAMBDA;
 533         kernel_data.lambda         = lam_i;
 534         kernel_data.energygrp_elec = enerd->foreign_grpp.ener[egCOULSR].data();
 535         kernel_data.energygrp_vdw  = enerd->foreign_grpp.ener[egLJSR].data();
 536         /* Note that we add to kernel_data.dvdl, but ignore the result */
 537
 538         for (size_t i = 0; i < enerd->enerpart_lambda.size(); i++)
 539         {
 540             for (int j = 0; j < efptNR; j++)
 541             {
 542                 lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i - 1]);
 543             }
 544             reset_foreign_enerdata(enerd);
 545 #pragma omp parallel for schedule(static) num_threads(nbl_fep.ssize())
 546             for (gmx::index th = 0; th < nbl_fep.ssize(); th++)
 547             {
 548                 try
 549                 {
 550                     gmx_nb_free_energy_kernel(nbl_fep[th].get(), x, forceWithShiftForces, fr,
 551                                               &mdatoms, &kernel_data, nrnb);
 552                 }
 553                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 554             }
 555
 556             sum_epot(&(enerd->foreign_grpp), enerd->foreign_term);
 557             enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
 558         }
 559     }
 560     wallcycle_sub_stop(wcycle_, ewcsNONBONDED_FEP);
 561 }