src/gromacs/mdlib/nbnxn_ocl/nbnxn_ocl_kernel_nowarp.clh

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2016,2017, 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 /*! \internal \file
  37  *  \brief OpenCL non-bonded "generic" kernel for architectures other than AMD/NVIDIA.
  38  *
  39  *  This kernel is use by default on all architectures other than the ones we
  40  *  explicitly support/optimize for.
  41  *
  42  *  OpenCL 1.1 support is expected.
  43  *
  44  *  \author Anca Hamuraru <anca@streamcomputing.eu>
  45  *  \author Szilárd Páll <pall.szilard@gmail.com>
  46  *  \ingroup module_mdlib
  47  */
  48
  49
  50 #include "nbnxn_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 EL_EWALD_ANY || defined EL_RF || defined LJ_EWALD || (defined EL_CUTOFF && defined CALC_ENERGIES)
  60 /* Macro to control the calculation of exclusion forces in the kernel
  61  * We do that with Ewald (elec/vdw) and RF. Cut-off only has exclusion
  62  * energy terms.
  63  *
  64  * Note: convenience macro, needs to be undef-ed at the end of the file.
  65  */
  66 #define EXCLUSION_FORCES
  67 #endif
  68
  69 #if defined LJ_EWALD_COMB_GEOM || defined LJ_EWALD_COMB_LB
  70 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
  71 #define LJ_EWALD
  72 #endif
  73
  74 #if defined LJ_COMB_GEOM || defined LJ_COMB_LB
  75 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
  76 #define LJ_COMB
  77 #endif
  78
  79 /*
  80    Kernel launch parameters:
  81     - #blocks   = #pair lists, blockId = pair list Id
  82     - #threads  = CL_SIZE^2
  83     - shmem     = CL_SIZE^2 * sizeof(float)
  84
  85     Each thread calculates an i force-component taking one pair of i-j atoms.
  86
  87   TODO: implement 128 threads/wavefront by porting over the NTHREAD_Z/j4 loop
  88   "horizontal splitting" over threads.
  89  */
  90
  91 /* NOTE:
  92  NB_KERNEL_FUNC_NAME differs from the CUDA equivalent as it is not a variadic macro due to OpenCL not having a support for them, this version only takes exactly 2 arguments.
  93  Thus if more strings need to be appended a new macro must be written or it must be directly appended here.
  94 */
  95 __attribute__((reqd_work_group_size(CL_SIZE, CL_SIZE, 1)))
  96 #ifdef PRUNE_NBL
  97     #ifdef CALC_ENERGIES
  98         __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_prune_opencl)
  99     #else
 100         __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_prune_opencl)
 101     #endif
 102 #else
 103     #ifdef CALC_ENERGIES
 104         __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_opencl)
 105     #else
 106         __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_opencl)
 107     #endif
 108 #endif
 109 (
 110 #ifndef LJ_COMB
 111  int ntypes,                                                               /* IN  */
 112 #endif
 113  cl_nbparam_params_t nbparam_params,                                       /* IN  */
 114  const __global float4 *restrict xq,                                       /* IN  */
 115  __global float *restrict f,                /* stores float3 values */     /* OUT */
 116  __global float *restrict e_lj,                                            /* OUT */
 117  __global float *restrict e_el,                                            /* OUT */
 118 __global float *restrict fshift,            /* stores float3 values */     /* OUT */
 119 #ifdef LJ_COMB
 120  const __global float2 *restrict lj_comb,      /* stores float2 values */  /* IN  */
 121 #else
 122  const __global int *restrict atom_types,                                  /* IN  */
 123 #endif
 124  const __global float *restrict shift_vec,  /* stores float3 values */     /* IN  */
 125  __constant float* nbfp_climg2d,                                           /* IN  */
 126  __constant float* nbfp_comb_climg2d,                                      /* IN  */
 127  __constant float* coulomb_tab_climg2d,                                    /* IN  */
 128  const __global nbnxn_sci_t* pl_sci,                                       /* IN  */
 129 #ifndef PRUNE_NBL
 130     const
 131 #endif
 132  __global nbnxn_cj4_t* pl_cj4,                                             /* OUT / IN */
 133  const __global nbnxn_excl_t* excl,                                        /* IN  */
 134  int bCalcFshift,                                                          /* IN  */
 135  __local  float4   *xqib,                                                  /* Pointer to dyn alloc'ed shmem */
 136  __global float *debug_buffer                                              /* Debug buffer, can be used with print_to_debug_buffer_f */
 137  )
 138 {
 139     /* convenience variables */
 140     cl_nbparam_params_t *nbparam = &nbparam_params;
 141
 142     float               rcoulomb_sq = nbparam->rcoulomb_sq;
 143 #ifdef LJ_COMB
 144     float2              ljcp_i, ljcp_j;
 145 #endif
 146 #ifdef VDW_CUTOFF_CHECK
 147     float               rvdw_sq     = nbparam_params.rvdw_sq;
 148     float               vdw_in_range;
 149 #endif
 150 #ifdef LJ_EWALD
 151     float               lje_coeff2, lje_coeff6_6;
 152 #endif
 153 #ifdef EL_RF
 154     float two_k_rf              = nbparam->two_k_rf;
 155 #endif
 156 #ifdef EL_EWALD_TAB
 157     float coulomb_tab_scale     = nbparam->coulomb_tab_scale;
 158 #endif
 159 #ifdef EL_EWALD_ANA
 160     float beta2                 = nbparam->ewald_beta*nbparam->ewald_beta;
 161     float beta3                 = nbparam->ewald_beta*nbparam->ewald_beta*nbparam->ewald_beta;
 162 #endif
 163 #ifdef PRUNE_NBL
 164     float rlist_sq              = nbparam->rlistOuter_sq;
 165 #endif
 166
 167 #ifdef CALC_ENERGIES
 168 #ifdef EL_EWALD_ANY
 169     float  beta        = nbparam->ewald_beta;
 170     float  ewald_shift = nbparam->sh_ewald;
 171 #else
 172     float  c_rf        = nbparam->c_rf;
 173 #endif /* EL_EWALD_ANY */
 174 #endif /* CALC_ENERGIES */
 175
 176     /* thread/block/warp id-s */
 177     unsigned int tidxi  = get_local_id(0);
 178     unsigned int tidxj  = get_local_id(1);
 179     unsigned int tidx   = get_local_id(1) * get_local_size(0) + get_local_id(0);
 180     unsigned int bidx   = get_group_id(0);
 181     unsigned int widx   = tidx / WARP_SIZE; /* warp index */
 182     int          sci, ci, cj, ci_offset,
 183                  ai, aj,
 184                  cij4_start, cij4_end;
 185 #ifndef LJ_COMB
 186     int          typei, typej;
 187 #endif
 188     int          i, jm, j4, wexcl_idx;
 189     float        qi, qj_f,
 190                  r2, inv_r, inv_r2;
 191 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
 192     float        inv_r6, c6, c12;
 193 #endif
 194 #if defined LJ_COMB_LB
 195     float        sigma, epsilon;
 196 #endif
 197     float        int_bit,
 198                  F_invr;
 199
 200 #ifdef CALC_ENERGIES
 201     float        E_lj, E_el;
 202 #endif
 203 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
 204     float        E_lj_p;
 205 #endif
 206     unsigned int wexcl, imask, mask_ji;
 207     float4       xqbuf;
 208     float3       xi, xj, rv, f_ij, fcj_buf/*, fshift_buf*/;
 209     float        fshift_buf;
 210     float3       fci_buf[NCL_PER_SUPERCL]; /* i force buffer */
 211     nbnxn_sci_t  nb_sci;
 212
 213     /*! i-cluster interaction mask for a super-cluster with all NCL_PER_SUPERCL=8 bits set */
 214     const unsigned superClInteractionMask = ((1U << NCL_PER_SUPERCL) - 1U);
 215
 216     /* shmem buffer for cj, for both warps separately */
 217     __local int *cjs       = (__local int *)(xqib + NCL_PER_SUPERCL * CL_SIZE);
 218     #define LOCAL_OFFSET cjs + 2 * NBNXN_GPU_JGROUP_SIZE
 219
 220 #ifdef IATYPE_SHMEM
 221 #ifndef LJ_COMB
 222     /* shmem buffer for i atom-type pre-loading */
 223     __local int *atib      = (__local int *)(LOCAL_OFFSET);
 224     #undef LOCAL_OFFSET
 225     #define LOCAL_OFFSET atib + NCL_PER_SUPERCL * CL_SIZE
 226 #else
 227     __local float2 *ljcpib = (__local float2 *)(LOCAL_OFFSET);
 228     #undef LOCAL_OFFSET
 229     #define LOCAL_OFFSET ljcpib + NCL_PER_SUPERCL * CL_SIZE
 230 #endif
 231 #endif
 232
 233 #ifndef REDUCE_SHUFFLE
 234     /* shmem j force buffer */
 235     __local float *f_buf   = (__local float *)(LOCAL_OFFSET);
 236     #undef LOCAL_OFFSET
 237     #define LOCAL_OFFSET f_buf + CL_SIZE * CL_SIZE * 3
 238 #endif
 239     /* Local buffer used to implement __any warp vote function from CUDA.
 240        volatile is used to avoid compiler optimizations for AMD builds. */
 241     volatile __local uint *warp_any = (__local uint*)(LOCAL_OFFSET);
 242 #undef LOCAL_OFFSET
 243
 244     nb_sci      = pl_sci[bidx];         /* my i super-cluster's index = current bidx */
 245     sci         = nb_sci.sci;           /* super-cluster */
 246     cij4_start  = nb_sci.cj4_ind_start; /* first ...*/
 247     cij4_end    = nb_sci.cj4_ind_end;   /* and last index of j clusters */
 248
 249     /* Pre-load i-atom x and q into shared memory */
 250     ci = sci * NCL_PER_SUPERCL + tidxj;
 251     ai = ci * CL_SIZE + tidxi;
 252
 253     xqbuf    = xq[ai] + (float4)(shift_vec[3 * nb_sci.shift], shift_vec[3 * nb_sci.shift + 1], shift_vec[3 * nb_sci.shift + 2], 0.0f);
 254     xqbuf.w *= nbparam->epsfac;
 255     xqib[tidxj * CL_SIZE + tidxi] = xqbuf;
 256
 257 #ifdef IATYPE_SHMEM
 258 #ifndef LJ_COMB
 259     /* Pre-load the i-atom types into shared memory */
 260     atib[tidxj * CL_SIZE + tidxi]   = atom_types[ai];
 261 #else
 262     ljcpib[tidxj * CL_SIZE + tidxi] = lj_comb[ai];
 263 #endif
 264 #endif
 265     /* Initialise warp vote. (8x8 block) 2 warps for nvidia */
 266     if(tidx==0 || tidx==32)
 267         warp_any[widx] = 0;
 268
 269     barrier(CLK_LOCAL_MEM_FENCE);
 270
 271     for (ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
 272     {
 273         fci_buf[ci_offset] = (float3)(0.0f);
 274     }
 275
 276 #ifdef LJ_EWALD
 277     /* TODO: we are trading registers with flops by keeping lje_coeff-s, try re-calculating it later */
 278     lje_coeff2   = nbparam->ewaldcoeff_lj*nbparam->ewaldcoeff_lj;
 279     lje_coeff6_6 = lje_coeff2*lje_coeff2*lje_coeff2*ONE_SIXTH_F;
 280 #endif /* LJ_EWALD */
 281
 282
 283 #ifdef CALC_ENERGIES
 284     E_lj = 0.0f;
 285     E_el = 0.0f;
 286
 287 #if defined EXCLUSION_FORCES /* Ewald or RF */
 288     if (nb_sci.shift == CENTRAL && pl_cj4[cij4_start].cj[0] == sci*NCL_PER_SUPERCL)
 289     {
 290         /* we have the diagonal: add the charge and LJ self interaction energy term */
 291         for (i = 0; i < NCL_PER_SUPERCL; i++)
 292         {
 293 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
 294             qi    = xqib[i * CL_SIZE + tidxi].w;
 295             E_el += qi*qi;
 296 #endif
 297 #if defined LJ_EWALD
 298             E_lj += nbfp_climg2d[atom_types[(sci*NCL_PER_SUPERCL + i)*CL_SIZE + tidxi]*(ntypes + 1)*2];
 299 #endif /* LJ_EWALD */
 300         }
 301
 302         /* divide the self term(s) equally over the j-threads, then multiply with the coefficients. */
 303 #ifdef LJ_EWALD
 304         E_lj /= CL_SIZE;
 305         E_lj *= 0.5f*ONE_SIXTH_F*lje_coeff6_6;
 306 #endif  /* LJ_EWALD */
 307
 308 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
 309         /* Correct for epsfac^2 due to adding qi^2 */
 310         E_el /= nbparam->epsfac*CL_SIZE;
 311 #if defined EL_RF || defined EL_CUTOFF
 312         E_el *= -0.5f*c_rf;
 313 #else
 314         E_el *= -beta*M_FLOAT_1_SQRTPI; /* last factor 1/sqrt(pi) */
 315 #endif
 316 #endif                                  /* EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF */
 317     }
 318 #endif                                  /* EXCLUSION_FORCES */
 319
 320 #endif                                  /* CALC_ENERGIES */
 321
 322 #ifdef EXCLUSION_FORCES
 323     const int nonSelfInteraction  = !(nb_sci.shift == CENTRAL & tidxj <= tidxi);
 324 #endif
 325
 326     /* loop over the j clusters = seen by any of the atoms in the current super-cluster */
 327     for (j4 = cij4_start; j4 < cij4_end; j4++)
 328     {
 329         wexcl_idx   = pl_cj4[j4].imei[widx].excl_ind;
 330         imask       = pl_cj4[j4].imei[widx].imask;
 331         wexcl       = excl[wexcl_idx].pair[(tidx) & (WARP_SIZE - 1)];
 332
 333 #ifndef PRUNE_NBL
 334         if (imask)
 335 #endif
 336         {
 337             /* Pre-load cj into shared memory on both warps separately */
 338             if ((tidxj == 0 | tidxj == 4) & (tidxi < NBNXN_GPU_JGROUP_SIZE))
 339             {
 340                 cjs[tidxi + tidxj * NBNXN_GPU_JGROUP_SIZE / 4] = pl_cj4[j4].cj[tidxi];
 341             }
 342
 343             /* Unrolling this loop improves performance without pruning but
 344              * with pruning it leads to slowdown.
 345              *
 346              * Tested with driver 1800.5
 347              */
 348 #if !defined PRUNE_NBL
 349 #pragma unroll 4
 350 #endif
 351
 352             for (jm = 0; jm < NBNXN_GPU_JGROUP_SIZE; jm++)
 353             {
 354                 if (imask & (superClInteractionMask << (jm * NCL_PER_SUPERCL)))
 355                 {
 356                     mask_ji = (1U << (jm * NCL_PER_SUPERCL));
 357
 358                     cj      = cjs[jm + (tidxj & 4) * NBNXN_GPU_JGROUP_SIZE / 4];
 359                     aj      = cj * CL_SIZE + tidxj;
 360
 361                     /* load j atom data */
 362                     xqbuf   = xq[aj];
 363                     xj      = (float3)(xqbuf.xyz);
 364                     qj_f    = xqbuf.w;
 365 #ifndef LJ_COMB
 366                     typej   = atom_types[aj];
 367 #else
 368                     ljcp_j  = lj_comb[aj];
 369 #endif
 370
 371                     fcj_buf = (float3)(0.0f);
 372
 373 #if !defined PRUNE_NBL
 374 #pragma unroll 8
 375 #endif
 376                     for (i = 0; i < NCL_PER_SUPERCL; i++)
 377                     {
 378                         if (imask & mask_ji)
 379                         {
 380                             ci_offset   = i;                     /* i force buffer offset */
 381
 382                             ci      = sci * NCL_PER_SUPERCL + i; /* i cluster index */
 383                             ai      = ci * CL_SIZE + tidxi;      /* i atom index */
 384
 385                             /* all threads load an atom from i cluster ci into shmem! */
 386                             xqbuf   = xqib[i * CL_SIZE + tidxi];
 387                             xi      = (float3)(xqbuf.xyz);
 388
 389                             /* distance between i and j atoms */
 390                             rv      = xi - xj;
 391                             r2      = norm2(rv);
 392
 393 #ifdef PRUNE_NBL
 394                             /* vote.. should code shmem serialisation, wonder what the hit will be */
 395                             if (r2 < rlist_sq)
 396                                 warp_any[widx]=1;
 397
 398                             /* If _none_ of the atoms pairs are in cutoff range,
 399                                the bit corresponding to the current
 400                                cluster-pair in imask gets set to 0. */
 401                             if (!warp_any[widx])
 402                                 imask &= ~mask_ji;
 403
 404                             warp_any[widx]=0;
 405 #endif
 406
 407                             int_bit = (wexcl & mask_ji) ? 1.0f : 0.0f;
 408
 409                             /* cutoff & exclusion check */
 410 #ifdef EXCLUSION_FORCES
 411                             if ((r2 < rcoulomb_sq) * (nonSelfInteraction | (ci != cj)))
 412 #else
 413                             if ((r2 < rcoulomb_sq) * int_bit)
 414 #endif
 415                             {
 416                                 /* load the rest of the i-atom parameters */
 417                                 qi      = xqbuf.w;
 418 #ifdef IATYPE_SHMEM
 419 #ifndef LJ_COMB
 420                                 typei   = atib[i * CL_SIZE + tidxi];
 421 #else
 422                                 ljcp_i  = ljcpib[i * CL_SIZE + tidxi];
 423 #endif
 424 #else /* IATYPE_SHMEM */
 425 #ifndef LJ_COMB
 426                                 typei   = atom_types[ai];
 427 #else
 428                                 ljcp_i  = lj_comb[ai];
 429 #endif
 430 #endif /* IATYPE_SHMEM */
 431                                 /* LJ 6*C6 and 12*C12 */
 432 #ifndef LJ_COMB
 433                                 c6      = nbfp_climg2d[2 * (ntypes * typei + typej)];
 434                                 c12     = nbfp_climg2d[2 * (ntypes * typei + typej)+1];
 435 #else   /* LJ_COMB */
 436 #ifdef LJ_COMB_GEOM
 437                                 c6      = ljcp_i.x * ljcp_j.x;
 438                                 c12     = ljcp_i.y * ljcp_j.y;
 439 #else
 440                                 /* LJ 2^(1/6)*sigma and 12*epsilon */
 441                                 sigma   = ljcp_i.x + ljcp_j.x;
 442                                 epsilon = ljcp_i.y * ljcp_j.y;
 443 #if defined CALC_ENERGIES || defined LJ_FORCE_SWITCH || defined LJ_POT_SWITCH
 444                                 convert_sigma_epsilon_to_c6_c12(sigma, epsilon, &c6, &c12);
 445 #endif
 446 #endif                          /* LJ_COMB_GEOM */
 447 #endif                          /* LJ_COMB */
 448
 449                                 // Ensure distance do not become so small that r^-12 overflows
 450                                 r2      = max(r2,NBNXN_MIN_RSQ);
 451
 452                                 inv_r   = rsqrt(r2);
 453                                 inv_r2  = inv_r * inv_r;
 454 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
 455                                 inv_r6  = inv_r2 * inv_r2 * inv_r2;
 456 #if defined EXCLUSION_FORCES
 457                                 /* We could mask inv_r2, but with Ewald
 458                                  * masking both inv_r6 and F_invr is faster */
 459                                 inv_r6  *= int_bit;
 460 #endif                          /* EXCLUSION_FORCES */
 461
 462                                 F_invr  = inv_r6 * (c12 * inv_r6 - c6) * inv_r2;
 463 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
 464                                 E_lj_p  = int_bit * (c12 * (inv_r6 * inv_r6 + nbparam->repulsion_shift.cpot)*ONE_TWELVETH_F -
 465                                                      c6 * (inv_r6 + nbparam->dispersion_shift.cpot)*ONE_SIXTH_F);
 466
 467 #endif
 468 #else                           /* ! LJ_COMB_LB || CALC_ENERGIES */
 469                                 float sig_r  = sigma*inv_r;
 470                                 float sig_r2 = sig_r*sig_r;
 471                                 float sig_r6 = sig_r2*sig_r2*sig_r2;
 472 #if defined EXCLUSION_FORCES
 473                                 sig_r6 *= int_bit;
 474 #endif                          /* EXCLUSION_FORCES */
 475
 476                                 F_invr  = epsilon * sig_r6 * (sig_r6 - 1.0f) * inv_r2;
 477 #endif                          /* ! LJ_COMB_LB || CALC_ENERGIES */
 478
 479
 480 #ifdef LJ_FORCE_SWITCH
 481 #ifdef CALC_ENERGIES
 482                                 calculate_force_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
 483 #else
 484                                 calculate_force_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr);
 485 #endif /* CALC_ENERGIES */
 486 #endif /* LJ_FORCE_SWITCH */
 487
 488
 489 #ifdef LJ_EWALD
 490 #ifdef LJ_EWALD_COMB_GEOM
 491 #ifdef CALC_ENERGIES
 492                                 calculate_lj_ewald_comb_geom_F_E(nbfp_comb_climg2d, nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6, int_bit, &F_invr, &E_lj_p);
 493 #else
 494                                 calculate_lj_ewald_comb_geom_F(nbfp_comb_climg2d, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6, &F_invr);
 495 #endif                          /* CALC_ENERGIES */
 496 #elif defined LJ_EWALD_COMB_LB
 497                                 calculate_lj_ewald_comb_LB_F_E(nbfp_comb_climg2d, nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6,
 498 #ifdef CALC_ENERGIES
 499                                                                int_bit, true, &F_invr, &E_lj_p
 500 #else
 501                                                                0, false, &F_invr, 0
 502 #endif /* CALC_ENERGIES */
 503                                                                );
 504 #endif /* LJ_EWALD_COMB_GEOM */
 505 #endif /* LJ_EWALD */
 506
 507 #ifdef LJ_POT_SWITCH
 508 #ifdef CALC_ENERGIES
 509                                 calculate_potential_switch_F_E(nbparam, inv_r, r2, &F_invr, &E_lj_p);
 510 #else
 511                                 calculate_potential_switch_F(nbparam, inv_r, r2, &F_invr, &E_lj_p);
 512 #endif /* CALC_ENERGIES */
 513 #endif /* LJ_POT_SWITCH */
 514
 515 #ifdef VDW_CUTOFF_CHECK
 516                                 /* Separate VDW cut-off check to enable twin-range cut-offs
 517                                  * (rvdw < rcoulomb <= rlist)
 518                                  */
 519                                 vdw_in_range  = (r2 < rvdw_sq) ? 1.0f : 0.0f;
 520                                 F_invr       *= vdw_in_range;
 521 #ifdef CALC_ENERGIES
 522                                 E_lj_p       *= vdw_in_range;
 523 #endif
 524 #endif                          /* VDW_CUTOFF_CHECK */
 525
 526 #ifdef CALC_ENERGIES
 527                                 E_lj    += E_lj_p;
 528
 529 #endif
 530
 531
 532 #ifdef EL_CUTOFF
 533 #ifdef EXCLUSION_FORCES
 534                                 F_invr  += qi * qj_f * int_bit * inv_r2 * inv_r;
 535 #else
 536                                 F_invr  += qi * qj_f * inv_r2 * inv_r;
 537 #endif
 538 #endif
 539 #ifdef EL_RF
 540                                 F_invr  += qi * qj_f * (int_bit*inv_r2 * inv_r - two_k_rf);
 541 #endif
 542 #if defined EL_EWALD_ANA
 543                                 F_invr  += qi * qj_f * (int_bit*inv_r2*inv_r + pmecorrF(beta2*r2)*beta3);
 544 #elif defined EL_EWALD_TAB
 545                                 F_invr  += qi * qj_f * (int_bit*inv_r2 -
 546 #ifdef USE_TEXOBJ
 547                                                         interpolate_coulomb_force_r(nbparam->coulomb_tab_texobj, r2 * inv_r, coulomb_tab_scale)
 548 #else
 549                                                         interpolate_coulomb_force_r(coulomb_tab_climg2d, r2 * inv_r, coulomb_tab_scale)
 550 #endif /* USE_TEXOBJ */
 551                                                         ) * inv_r;
 552 #endif /* EL_EWALD_ANA/TAB */
 553
 554 #ifdef CALC_ENERGIES
 555 #ifdef EL_CUTOFF
 556                                 E_el    += qi * qj_f * (int_bit*inv_r - c_rf);
 557 #endif
 558 #ifdef EL_RF
 559                                 E_el    += qi * qj_f * (int_bit*inv_r + 0.5f * two_k_rf * r2 - c_rf);
 560 #endif
 561 #ifdef EL_EWALD_ANY
 562                                 /* 1.0f - erff is faster than erfcf */
 563                                 E_el    += qi * qj_f * (inv_r * (int_bit - erf(r2 * inv_r * beta)) - int_bit * ewald_shift);
 564 #endif                          /* EL_EWALD_ANY */
 565 #endif
 566                                 f_ij    = rv * F_invr;
 567
 568                                 /* accumulate j forces in registers */
 569                                 fcj_buf -= f_ij;
 570
 571                                 /* accumulate i forces in registers */
 572                                 fci_buf[ci_offset] += f_ij;
 573                             }
 574                         }
 575
 576                         /* shift the mask bit by 1 */
 577                         mask_ji += mask_ji;
 578                     }
 579
 580                     /* reduce j forces */
 581
 582                     /* store j forces in shmem */
 583                     f_buf[                  tidx] = fcj_buf.x;
 584                     f_buf[    FBUF_STRIDE + tidx] = fcj_buf.y;
 585                     f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
 586
 587                     reduce_force_j_generic(f_buf, f, tidxi, tidxj, aj);
 588                 }
 589             }
 590 #ifdef PRUNE_NBL
 591             /* Update the imask with the new one which does not contain the
 592                out of range clusters anymore. */
 593
 594             pl_cj4[j4].imei[widx].imask = imask;
 595 #endif
 596         }
 597     }
 598
 599     /* skip central shifts when summing shift forces */
 600     if (nb_sci.shift == CENTRAL)
 601     {
 602         bCalcFshift = false;
 603     }
 604
 605     fshift_buf = 0.0f;
 606
 607     /* reduce i forces */
 608     for (ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
 609     {
 610         ai  = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
 611
 612         f_buf[                  tidx] = fci_buf[ci_offset].x;
 613         f_buf[    FBUF_STRIDE + tidx] = fci_buf[ci_offset].y;
 614         f_buf[2 * FBUF_STRIDE + tidx] = fci_buf[ci_offset].z;
 615         barrier(CLK_LOCAL_MEM_FENCE);
 616         reduce_force_i(f_buf, f,
 617                        &fshift_buf, bCalcFshift,
 618                        tidxi, tidxj, ai);
 619         barrier(CLK_LOCAL_MEM_FENCE);
 620     }
 621
 622     /* add up local shift forces into global mem */
 623     if (bCalcFshift)
 624     {
 625         /* Only threads with tidxj < 3 will update fshift.
 626            The threads performing the update must be the same with the threads
 627            which stored the reduction result in reduce_force_i function
 628         */
 629         if (tidxj < 3)
 630             atomicAdd_g_f(&(fshift[3 * nb_sci.shift + tidxj]), fshift_buf);
 631     }
 632
 633 #ifdef CALC_ENERGIES
 634     /* flush the energies to shmem and reduce them */
 635     f_buf[              tidx] = E_lj;
 636     f_buf[FBUF_STRIDE + tidx] = E_el;
 637     reduce_energy_pow2(f_buf + (tidx & WARP_SIZE), e_lj, e_el, tidx & ~WARP_SIZE);
 638
 639 #endif
 640 }
 641
 642 #undef EL_EWALD_ANY
 643 #undef EXCLUSION_FORCES
 644 #undef LJ_EWALD
 645
 646 #undef LJ_COMB