src/gromacs/nbnxm/opencl/nbnxm_ocl_kernel.clh

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012-2018, The GROMACS development team.
   5  * Copyright (c) 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 /*! \internal \file
  38  *  \brief OpenCL non-bonded kernel.
  39  *
  40  *  OpenCL 1.2 support is expected.
  41  *
  42  *  \author Anca Hamuraru <anca@streamcomputing.eu>
  43  *  \author Szilárd Páll <pall.szilard@gmail.com>
  44  *  \ingroup module_nbnxm
  45  */
  46
  47 /* Currently we enable CJ prefetch for AMD/NVIDIA and disable it for the "nowarp" kernel
  48  * Note that this should precede the kernel_utils include.
  49  */
  50 #include "nbnxm_ocl_kernel_utils.clh"
  51
  52 /////////////////////////////////////////////////////////////////////////////////////////////////
  53
  54 #if defined EL_EWALD_ANA || defined EL_EWALD_TAB
  55 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
  56 #    define EL_EWALD_ANY
  57 #endif
  58
  59 #if defined LJ_EWALD_COMB_GEOM || defined LJ_EWALD_COMB_LB
  60 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
  61 #    define LJ_EWALD
  62 #endif
  63
  64 #if defined EL_EWALD_ANY || defined EL_RF || defined LJ_EWALD \
  65         || (defined EL_CUTOFF && defined CALC_ENERGIES)
  66 /* Macro to control the calculation of exclusion forces in the kernel
  67  * We do that with Ewald (elec/vdw) and RF. Cut-off only has exclusion
  68  * energy terms.
  69  *
  70  * Note: convenience macro, needs to be undef-ed at the end of the file.
  71  */
  72 #    define EXCLUSION_FORCES
  73 #endif
  74
  75 #if defined LJ_COMB_GEOM || defined LJ_COMB_LB
  76 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
  77 #    define LJ_COMB
  78 #endif
  79
  80 /*
  81    Kernel launch parameters:
  82     - #blocks   = #pair lists, blockId = pair list Id
  83     - #threads  = CL_SIZE^2
  84     - shmem     = CL_SIZE^2 * sizeof(float)
  85
  86     Each thread calculates an i force-component taking one pair of i-j atoms.
  87
  88    TODO: implement 128 threads/wavefront by porting over the NTHREAD_Z/j4 loop
  89    "horizontal splitting" over threads.
  90  */
  91
  92 /* NOTE:
  93    NB_KERNEL_FUNC_NAME differs from the CUDA equivalent as it is not a variadic macro due to OpenCL
  94    not having a support for them, this version only takes exactly 2 arguments. Thus if more strings
  95    need to be appended a new macro must be written or it must be directly appended here.
  96  */
  97 __attribute__((reqd_work_group_size(CL_SIZE, CL_SIZE, 1)))
  98 #ifdef cl_intel_required_subgroup_size
  99 __attribute__((intel_reqd_sub_group_size(SUBGROUP_SIZE)))
 100 #endif
 101 #ifdef PRUNE_NBL
 102 #    ifdef CALC_ENERGIES
 103 __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_prune_opencl)
 104 #    else
 105 __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_prune_opencl)
 106 #    endif
 107 #else
 108 #    ifdef CALC_ENERGIES
 109 __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_opencl)
 110 #    else
 111 __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_opencl)
 112 #    endif
 113 #endif
 114         (
 115 #ifndef LJ_COMB
 116                 int ntypes, /* IN  */
 117 #endif
 118                 cl_nbparam_params_t nbparam_params,       /* IN  */
 119                 const __global float4* restrict xq,       /* IN  */
 120                 __global float* restrict f,               /* OUT stores float3 values */
 121                 __global float* restrict gmx_unused e_lj, /* OUT */
 122                 __global float* restrict gmx_unused e_el, /* OUT */
 123                 __global float* restrict fshift,          /* OUT stores float3 values */
 124 #ifdef LJ_COMB
 125                 const __global float2* restrict lj_comb, /* IN stores float2 values */
 126 #else
 127                 const __global int* restrict atom_types, /* IN  */
 128 #endif
 129                 const __global float* restrict shift_vec,          /* IN stores float3 values */
 130                 __constant const float2* restrict gmx_unused nbfp, /* IN  */
 131                 __constant const float2* restrict gmx_unused nbfp_comb,  /* IN  */
 132                 __constant const float* restrict gmx_unused coulomb_tab, /* IN  */
 133                 const __global nbnxn_sci_t* pl_sci,                      /* IN  */
 134 #ifndef PRUNE_NBL
 135                 const
 136 #endif
 137                 __global nbnxn_cj4_t* pl_cj4,             /* OUT / IN */
 138                 const __global nbnxn_excl_t* excl,        /* IN  */
 139                 int                          bCalcFshift, /* IN  */
 140                 __local float4* xqib                      /* Pointer to dyn alloc'ed shmem */
 141         )
 142 {
 143     /* convenience variables */
 144     const cl_nbparam_params_t* const nbparam = &nbparam_params;
 145
 146     const float rcoulomb_sq = nbparam->rcoulomb_sq;
 147 #ifdef VDW_CUTOFF_CHECK
 148     const float rvdw_sq = nbparam_params.rvdw_sq;
 149 #endif
 150 #ifdef EL_RF
 151     const float two_k_rf = nbparam->two_k_rf;
 152 #endif
 153 #ifdef EL_EWALD_TAB
 154     const float coulomb_tab_scale = nbparam->coulomb_tab_scale;
 155 #endif
 156 #ifdef EL_EWALD_ANA
 157     const float beta2 = nbparam->ewald_beta * nbparam->ewald_beta;
 158     const float beta3 = nbparam->ewald_beta * nbparam->ewald_beta * nbparam->ewald_beta;
 159 #endif
 160 #ifdef PRUNE_NBL
 161     const float rlist_sq = nbparam->rlistOuter_sq;
 162 #endif
 163
 164 #ifdef CALC_ENERGIES
 165 #    ifdef EL_EWALD_ANY
 166     const float beta        = nbparam->ewald_beta;
 167     const float ewald_shift = nbparam->sh_ewald;
 168 #    else
 169     const float gmx_unused c_rf = nbparam->c_rf;
 170 #    endif /* EL_EWALD_ANY */
 171 #endif     /* CALC_ENERGIES */
 172
 173     /* thread/block/warp id-s */
 174     const int tidxi = get_local_id(0);
 175     const int tidxj = get_local_id(1);
 176     const int tidx  = (int)(get_local_id(1) * get_local_size(0) + get_local_id(0));
 177     const int bidx  = get_group_id(0);
 178     const int widx  = tidx / WARP_SIZE; /* warp index */
 179
 180     /*! i-cluster interaction mask for a super-cluster with all c_nbnxnGpuNumClusterPerSupercluster=8 bits set */
 181     const unsigned superClInteractionMask = ((1U << c_nbnxnGpuNumClusterPerSupercluster) - 1U);
 182
 183 #define LOCAL_OFFSET (xqib + c_nbnxnGpuNumClusterPerSupercluster * CL_SIZE)
 184     CjType cjs = 0;
 185 #if USE_CJ_PREFETCH
 186     /* shmem buffer for cj, for both warps separately */
 187     cjs = (__local int*)(LOCAL_OFFSET);
 188 #    undef LOCAL_OFFSET
 189 #    define LOCAL_OFFSET (cjs + 2 * c_nbnxnGpuJgroupSize)
 190 #endif // USE_CJ_PREFETCH
 191
 192 #ifdef IATYPE_SHMEM
 193 #    ifndef LJ_COMB
 194     /* shmem buffer for i atom-type pre-loading */
 195     __local int* atib = (__local int*)(LOCAL_OFFSET); //NOLINT(google-readability-casting)
 196 #        undef LOCAL_OFFSET
 197 #        define LOCAL_OFFSET (atib + c_nbnxnGpuNumClusterPerSupercluster * CL_SIZE)
 198 #    else
 199     __local float2* ljcpib      = (__local float2*)(LOCAL_OFFSET);
 200 #        undef LOCAL_OFFSET
 201 #        define LOCAL_OFFSET (ljcpib + c_nbnxnGpuNumClusterPerSupercluster * CL_SIZE)
 202 #    endif
 203 #endif
 204
 205 #if !REDUCE_SHUFFLE
 206     /* shmem j force buffer */
 207     __local float* f_buf = (__local float*)(LOCAL_OFFSET);
 208 #    undef LOCAL_OFFSET
 209 #    define LOCAL_OFFSET (f_buf + CL_SIZE * CL_SIZE * 3)
 210 #else
 211     __local float* f_buf             = 0;
 212 #endif
 213 #if !USE_SUBGROUP_ANY
 214     /* Local buffer used to implement __any warp vote function from CUDA.
 215        volatile is used to avoid compiler optimizations for AMD builds. */
 216     //NOLINTNEXTLINE(google-readability-casting)
 217     volatile __local int* warp_any = (__local int*)(LOCAL_OFFSET);
 218 #else
 219     __local int gmx_unused* warp_any = 0;
 220 #endif
 221 #undef LOCAL_OFFSET
 222
 223     const nbnxn_sci_t nb_sci     = pl_sci[bidx]; /* my i super-cluster's index = current bidx */
 224     const int         sci        = nb_sci.sci;   /* super-cluster */
 225     const int         cij4_start = nb_sci.cj4_ind_start; /* first ...*/
 226     const int         cij4_end   = nb_sci.cj4_ind_end;   /* and last index of j clusters */
 227
 228     for (int i = 0; i < c_nbnxnGpuNumClusterPerSupercluster; i += CL_SIZE)
 229     {
 230         /* Pre-load i-atom x and q into shared memory */
 231         const int ci = sci * c_nbnxnGpuNumClusterPerSupercluster + tidxj + i;
 232         const int ai = ci * CL_SIZE + tidxi;
 233
 234         float4 xqbuf = xq[ai]
 235                        + (float4)(shift_vec[3 * nb_sci.shift],
 236                                   shift_vec[3 * nb_sci.shift + 1],
 237                                   shift_vec[3 * nb_sci.shift + 2],
 238                                   0.0F);
 239         xqbuf.w *= nbparam->epsfac;
 240         xqib[(tidxj + i) * CL_SIZE + tidxi] = xqbuf;
 241 #ifdef IATYPE_SHMEM
 242 #    ifndef LJ_COMB
 243         /* Pre-load the i-atom types into shared memory */
 244         atib[(tidxj + i) * CL_SIZE + tidxi] = atom_types[ai];
 245 #    else
 246         ljcpib[(tidxj + i) * CL_SIZE + tidxi] = lj_comb[ai];
 247 #    endif
 248 #endif
 249     }
 250 #if !USE_SUBGROUP_ANY
 251     /* Initialise warp vote. (8x8 block) 2 warps for nvidia */
 252     if (tidx == 0 || tidx == WARP_SIZE)
 253     {
 254         warp_any[widx] = 0;
 255     }
 256 #endif
 257     barrier(CLK_LOCAL_MEM_FENCE);
 258
 259     fvec fci_buf[c_nbnxnGpuNumClusterPerSupercluster]; /* i force buffer */
 260     for (int ci_offset = 0; ci_offset < c_nbnxnGpuNumClusterPerSupercluster; ci_offset++)
 261     {
 262         fci_buf[ci_offset][0] = 0.0F;
 263         fci_buf[ci_offset][1] = 0.0F;
 264         fci_buf[ci_offset][2] = 0.0F;
 265     }
 266
 267 #ifdef LJ_EWALD
 268     /* TODO: we are trading registers with flops by keeping lje_coeff-s, try re-calculating it later */
 269     const float lje_coeff2   = nbparam->ewaldcoeff_lj * nbparam->ewaldcoeff_lj;
 270     const float lje_coeff6_6 = lje_coeff2 * lje_coeff2 * lje_coeff2 * ONE_SIXTH_F;
 271 #endif /* LJ_EWALD */
 272
 273
 274 #ifdef CALC_ENERGIES
 275     float E_lj = 0.0F;
 276     float E_el = 0.0F;
 277
 278 #    if defined EXCLUSION_FORCES /* Ewald or RF */
 279     if (nb_sci.shift == c_centralShiftIndex
 280         && pl_cj4[cij4_start].cj[0] == sci * c_nbnxnGpuNumClusterPerSupercluster)
 281     {
 282         /* we have the diagonal: add the charge and LJ self interaction energy term */
 283         for (int i = 0; i < c_nbnxnGpuNumClusterPerSupercluster; i++)
 284         {
 285 #        if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
 286             const float qi = xqib[i * CL_SIZE + tidxi].w;
 287             E_el += qi * qi;
 288 #        endif
 289 #        if defined LJ_EWALD
 290             E_lj += nbfp[atom_types[(sci * c_nbnxnGpuNumClusterPerSupercluster + i) * CL_SIZE + tidxi] * (ntypes + 1)]
 291                             .x;
 292 #        endif /* LJ_EWALD */
 293         }
 294
 295         /* divide the self term(s) equally over the j-threads, then multiply with the coefficients. */
 296 #        ifdef LJ_EWALD
 297         E_lj /= CL_SIZE;
 298         E_lj *= HALF_F * ONE_SIXTH_F * lje_coeff6_6;
 299 #        endif /* LJ_EWALD */
 300
 301 #        if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
 302         /* Correct for epsfac^2 due to adding qi^2 */
 303         E_el /= nbparam->epsfac * CL_SIZE;
 304 #            if defined EL_RF || defined EL_CUTOFF
 305         E_el *= -HALF_F * c_rf;
 306 #            else
 307         E_el *= -beta * M_FLOAT_1_SQRTPI; /* last factor 1/sqrt(pi) */
 308 #            endif
 309 #        endif /* EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF */
 310     }
 311 #    endif /* EXCLUSION_FORCES */
 312
 313 #endif /* CALC_ENERGIES */
 314
 315 #ifdef EXCLUSION_FORCES
 316     const int nonSelfInteraction = !(nb_sci.shift == c_centralShiftIndex & tidxj <= tidxi);
 317 #endif
 318
 319     /* loop over the j clusters = seen by any of the atoms in the current super-cluster */
 320     for (int j4 = cij4_start; j4 < cij4_end; j4++)
 321     {
 322         const int          wexcl_idx = pl_cj4[j4].imei[widx].excl_ind;
 323         unsigned int       imask     = pl_cj4[j4].imei[widx].imask;
 324         const unsigned int wexcl     = excl[wexcl_idx].pair[(tidx) & (WARP_SIZE - 1)];
 325
 326         preloadCj4(&cjs, pl_cj4[j4].cj, tidxi, tidxj, imask != 0U);
 327
 328 #ifndef PRUNE_NBL
 329         if (imask)
 330 #endif
 331         {
 332             /* Unrolling this loop improves performance without pruning but
 333              * with pruning it leads to slowdown.
 334              *
 335              * Tested with driver 1800.5
 336              *
 337              * TODO: check loop unrolling with NVIDIA OpenCL
 338              */
 339 #if !defined PRUNE_NBL && !defined _NVIDIA_SOURCE_
 340 #    pragma unroll 4
 341 #endif
 342             for (int jm = 0; jm < c_nbnxnGpuJgroupSize; jm++)
 343             {
 344                 if (imask & (superClInteractionMask << (jm * c_nbnxnGpuNumClusterPerSupercluster)))
 345                 {
 346                     unsigned int mask_ji = (1U << (jm * c_nbnxnGpuNumClusterPerSupercluster));
 347
 348                     const int cj = loadCj(cjs, pl_cj4[j4].cj, jm, tidxi, tidxj);
 349                     const int aj = cj * CL_SIZE + tidxj;
 350
 351                     /* load j atom data */
 352                     const float4 xjqbuf = xq[aj];
 353                     const float3 xj     = (float3)(xjqbuf.xyz);
 354                     const float  qj_f   = xjqbuf.w;
 355 #ifndef LJ_COMB
 356                     const int typej = atom_types[aj];
 357 #else
 358                     const float2 ljcp_j = lj_comb[aj];
 359 #endif
 360
 361                     float3 fcj_buf = (float3)(0.0F);
 362
 363 #if !defined PRUNE_NBL
 364 #    pragma unroll 8
 365 #endif
 366                     for (int i = 0; i < c_nbnxnGpuNumClusterPerSupercluster; i++)
 367                     {
 368                         if (imask & mask_ji)
 369                         {
 370                             const int gmx_unused ci = sci * c_nbnxnGpuNumClusterPerSupercluster + i; /* i cluster index */
 371
 372                             /* all threads load an atom from i cluster ci into shmem! */
 373                             const float4 xiqbuf = xqib[i * CL_SIZE + tidxi];
 374                             const float3 xi     = (float3)(xiqbuf.xyz);
 375
 376                             /* distance between i and j atoms */
 377                             const float3 rv = xi - xj;
 378                             float        r2 = norm2(rv);
 379
 380 #ifdef PRUNE_NBL
 381                             if (!gmx_sub_group_any(warp_any, widx, r2 < rlist_sq))
 382                             {
 383                                 imask &= ~mask_ji;
 384                             }
 385 #endif
 386
 387                             const float int_bit = (wexcl & mask_ji) ? 1.0F : 0.0F;
 388
 389                             /* cutoff & exclusion check */
 390 #ifdef EXCLUSION_FORCES
 391                             if ((r2 < rcoulomb_sq) * (nonSelfInteraction | (ci != cj)))
 392 #else
 393                             if ((float)(r2 < rcoulomb_sq) * int_bit != 0.0F)
 394 #endif
 395                             {
 396                                 /* load the rest of the i-atom parameters */
 397                                 const float qi = xiqbuf.w;
 398 #ifdef IATYPE_SHMEM
 399 #    ifndef LJ_COMB
 400                                 const int typei = atib[i * CL_SIZE + tidxi];
 401 #    else
 402                                 const float2 ljcp_i = ljcpib[i * CL_SIZE + tidxi];
 403 #    endif
 404 #else /* IATYPE_SHMEM */
 405                                 const int   ai     = ci * CL_SIZE + tidxi; /* i atom index */
 406
 407 #    ifndef LJ_COMB
 408                                 const int   typei  = atom_types[ai];
 409 #    else
 410                                 const float2 ljcp_i = lj_comb[ai];
 411 #    endif
 412 #endif                          /* IATYPE_SHMEM */
 413                                 /* LJ 6*C6 and 12*C12 */
 414 #ifndef LJ_COMB
 415                                 const float2 c6c12 = nbfp[ntypes * typei + typej];
 416
 417                                 const float c6  = c6c12.x;
 418                                 const float c12 = c6c12.y;
 419 #else /* LJ_COMB */
 420 #    ifdef LJ_COMB_GEOM
 421                                 const float c6     = ljcp_i.x * ljcp_j.x;
 422                                 const float c12    = ljcp_i.y * ljcp_j.y;
 423 #    else
 424                                 /* LJ 2^(1/6)*sigma and 12*epsilon */
 425                                 const float  sigma   = ljcp_i.x + ljcp_j.x;
 426                                 const float  epsilon = ljcp_i.y * ljcp_j.y;
 427 #        if defined CALC_ENERGIES || defined LJ_FORCE_SWITCH || defined LJ_POT_SWITCH
 428                                 const float2 c6c12 = convert_sigma_epsilon_to_c6_c12(sigma, epsilon);
 429                                 const float  c6    = c6c12.x;
 430                                 const float  c12   = c6c12.y;
 431 #        endif
 432 #    endif /* LJ_COMB_GEOM */
 433 #endif     /* LJ_COMB */
 434
 435                                 // Ensure distance do not become so small that r^-12 overflows.
 436                                 r2 = max(r2, c_nbnxnMinDistanceSquared);
 437
 438                                 const float inv_r  = rsqrt(r2);
 439                                 const float inv_r2 = inv_r * inv_r;
 440 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
 441                                 float inv_r6 = inv_r2 * inv_r2 * inv_r2;
 442 #    if defined EXCLUSION_FORCES
 443                                 /* We could mask inv_r2, but with Ewald
 444                                  * masking both inv_r6 and F_invr is faster */
 445                                 inv_r6 *= int_bit;
 446 #    endif /* EXCLUSION_FORCES */
 447
 448                                 float F_invr = inv_r6 * (c12 * inv_r6 - c6) * inv_r2;
 449 #    if defined CALC_ENERGIES || defined LJ_POT_SWITCH
 450                                 float E_lj_p =
 451                                         int_bit
 452                                         * (c12 * (inv_r6 * inv_r6 + nbparam->repulsion_shift.cpot) * ONE_TWELVETH_F
 453                                            - c6 * (inv_r6 + nbparam->dispersion_shift.cpot) * ONE_SIXTH_F);
 454
 455 #    endif
 456 #else /* ! LJ_COMB_LB || CALC_ENERGIES */
 457                                 const float sig_r  = sigma * inv_r;
 458                                 const float sig_r2 = sig_r * sig_r;
 459                                 float       sig_r6 = sig_r2 * sig_r2 * sig_r2;
 460 #    if defined EXCLUSION_FORCES
 461                                 sig_r6 *= int_bit;
 462 #    endif /* EXCLUSION_FORCES */
 463
 464                                 float F_invr = epsilon * sig_r6 * (sig_r6 - 1.0F) * inv_r2;
 465 #endif     /* ! LJ_COMB_LB || CALC_ENERGIES */
 466
 467
 468 #ifdef LJ_FORCE_SWITCH
 469 #    ifdef CALC_ENERGIES
 470                                 calculate_force_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
 471 #    else
 472                                 calculate_force_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr);
 473 #    endif /* CALC_ENERGIES */
 474 #endif     /* LJ_FORCE_SWITCH */
 475
 476
 477 #ifdef LJ_EWALD
 478 #    ifdef LJ_EWALD_COMB_GEOM
 479 #        ifdef CALC_ENERGIES
 480                                 calculate_lj_ewald_comb_geom_F_E(nbfp_comb,
 481                                                                  nbparam,
 482                                                                  typei,
 483                                                                  typej,
 484                                                                  r2,
 485                                                                  inv_r2,
 486                                                                  lje_coeff2,
 487                                                                  lje_coeff6_6,
 488                                                                  int_bit,
 489                                                                  &F_invr,
 490                                                                  &E_lj_p);
 491 #        else
 492                                 calculate_lj_ewald_comb_geom_F(
 493                                         nbfp_comb, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6, &F_invr);
 494 #        endif /* CALC_ENERGIES */
 495 #    elif defined LJ_EWALD_COMB_LB
 496                                 calculate_lj_ewald_comb_LB_F_E(nbfp_comb,
 497                                                                nbparam,
 498                                                                typei,
 499                                                                typej,
 500                                                                r2,
 501                                                                inv_r2,
 502                                                                lje_coeff2,
 503                                                                lje_coeff6_6,
 504 #        ifdef CALC_ENERGIES
 505                                                                int_bit,
 506                                                                true,
 507                                                                &F_invr,
 508                                                                &E_lj_p
 509 #        else
 510                                                                0,
 511                                                                false,
 512                                                                &F_invr,
 513                                                                0
 514 #        endif /* CALC_ENERGIES */
 515                                 );
 516 #    endif     /* LJ_EWALD_COMB_GEOM */
 517 #endif         /* LJ_EWALD */
 518
 519 #ifdef LJ_POT_SWITCH
 520 #    ifdef CALC_ENERGIES
 521                                 calculate_potential_switch_F_E(nbparam, inv_r, r2, &F_invr, &E_lj_p);
 522 #    else
 523                                 calculate_potential_switch_F(nbparam, inv_r, r2, &F_invr, &E_lj_p);
 524 #    endif /* CALC_ENERGIES */
 525 #endif     /* LJ_POT_SWITCH */
 526
 527 #ifdef VDW_CUTOFF_CHECK
 528                                 /* Separate VDW cut-off check to enable twin-range cut-offs
 529                                  * (rvdw < rcoulomb <= rlist)
 530                                  */
 531                                 const float vdw_in_range = (r2 < rvdw_sq) ? 1.0F : 0.0F;
 532                                 F_invr *= vdw_in_range;
 533 #    ifdef CALC_ENERGIES
 534                                 E_lj_p *= vdw_in_range;
 535 #    endif
 536 #endif /* VDW_CUTOFF_CHECK */
 537
 538 #ifdef CALC_ENERGIES
 539                                 E_lj += E_lj_p;
 540
 541 #endif
 542
 543
 544 #ifdef EL_CUTOFF
 545 #    ifdef EXCLUSION_FORCES
 546                                 F_invr += qi * qj_f * int_bit * inv_r2 * inv_r;
 547 #    else
 548                                 F_invr += qi * qj_f * inv_r2 * inv_r;
 549 #    endif
 550 #endif
 551 #ifdef EL_RF
 552                                 F_invr += qi * qj_f * (int_bit * inv_r2 * inv_r - two_k_rf);
 553 #endif
 554 #if defined EL_EWALD_ANA
 555                                 F_invr += qi * qj_f
 556                                           * (int_bit * inv_r2 * inv_r + pmecorrF(beta2 * r2) * beta3);
 557 #elif defined EL_EWALD_TAB
 558                                 F_invr += qi * qj_f
 559                                           * (int_bit * inv_r2
 560                                              - interpolate_coulomb_force_r(
 561                                                        coulomb_tab, r2 * inv_r, coulomb_tab_scale))
 562                                           * inv_r;
 563 #endif /* EL_EWALD_ANA/TAB */
 564
 565 #ifdef CALC_ENERGIES
 566 #    ifdef EL_CUTOFF
 567                                 E_el += qi * qj_f * (int_bit * inv_r - c_rf);
 568 #    endif
 569 #    ifdef EL_RF
 570                                 E_el += qi * qj_f * (int_bit * inv_r + HALF_F * two_k_rf * r2 - c_rf);
 571 #    endif
 572 #    ifdef EL_EWALD_ANY
 573                                 /* 1.0F - erff is faster than erfcf */
 574                                 E_el += qi * qj_f
 575                                         * (inv_r * (int_bit - erf(r2 * inv_r * beta)) - int_bit * ewald_shift);
 576 #    endif /* EL_EWALD_ANY */
 577 #endif
 578                                 const float3 f_ij = rv * F_invr;
 579
 580                                 /* accumulate j forces in registers */
 581                                 fcj_buf -= f_ij;
 582
 583                                 /* accumulate i forces in registers */
 584                                 fci_buf[i][0] += f_ij.x;
 585                                 fci_buf[i][1] += f_ij.y;
 586                                 fci_buf[i][2] += f_ij.z;
 587                             }
 588                         }
 589
 590                         /* shift the mask bit by 1 */
 591                         mask_ji += mask_ji;
 592                     }
 593
 594                     /* reduce j forces */
 595                     reduce_force_j(f_buf, fcj_buf, f, tidxi, tidxj, aj);
 596                 }
 597             }
 598 #ifdef PRUNE_NBL
 599             /* Update the imask with the new one which does not contain the
 600                out of range clusters anymore. */
 601
 602             pl_cj4[j4].imei[widx].imask = imask;
 603 #endif
 604         }
 605     }
 606
 607     /* skip central shifts when summing shift forces */
 608     if (nb_sci.shift == c_centralShiftIndex)
 609     {
 610         bCalcFshift = 0;
 611     }
 612     /* reduce i forces */
 613     reduce_force_i_and_shift(f_buf, fci_buf, f, bCalcFshift != 0, tidxi, tidxj, sci, nb_sci.shift, fshift);
 614
 615 #ifdef CALC_ENERGIES
 616     reduce_energy(f_buf, E_lj, E_el, e_lj, e_el, tidx);
 617 #endif
 618 }
 619
 620 #undef EL_EWALD_ANY
 621 #undef EXCLUSION_FORCES
 622 #undef LJ_EWALD
 623
 624 #undef LJ_COMB
 625
 626 #undef USE_CJ_PREFETCH