src/gromacs/mdlib/nbnxn_ocl/nbnxn_ocl_kernel_nvidia.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 kernel for NVIDIA GPUs.
  38  *
  39  *  OpenCL 1.2 support is expected (CUDA driver API 7.5 and later).
  40  *
  41  *  \author Anca Hamuraru <anca@streamcomputing.eu>
  42  *  \author Szilárd Páll <pall.szilard@gmail.com>
  43  *  \ingroup module_mdlib
  44  */
  45
  46
  47 #include "nbnxn_ocl_kernel_utils.clh"
  48
  49 /////////////////////////////////////////////////////////////////////////////////////////////////
  50
  51 #if defined EL_EWALD_ANA || defined EL_EWALD_TAB
  52 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
  53 #define EL_EWALD_ANY
  54 #endif
  55
  56 #if defined EL_EWALD_ANY || defined EL_RF || defined LJ_EWALD || (defined EL_CUTOFF && defined CALC_ENERGIES)
  57 /* Macro to control the calculation of exclusion forces in the kernel
  58  * We do that with Ewald (elec/vdw) and RF. Cut-off only has exclusion
  59  * energy terms.
  60  *
  61  * Note: convenience macro, needs to be undef-ed at the end of the file.
  62  */
  63 #define EXCLUSION_FORCES
  64 #endif
  65
  66 #if defined LJ_EWALD_COMB_GEOM || defined LJ_EWALD_COMB_LB
  67 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
  68 #define LJ_EWALD
  69 #endif
  70
  71 #if defined LJ_COMB_GEOM || defined LJ_COMB_LB
  72 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
  73 #define LJ_COMB
  74 #endif
  75
  76 /*
  77    Kernel launch parameters:
  78     - #blocks   = #pair lists, blockId = pair list Id
  79     - #threads  = CL_SIZE^2
  80     - shmem     = CL_SIZE^2 * sizeof(float)
  81
  82     Each thread calculates an i force-component taking one pair of i-j atoms.
  83  */
  84
  85 /* NOTE:
  86  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.
  87  Thus if more strings need to be appended a new macro must be written or it must be directly appended here.
  88 */
  89 __attribute__((reqd_work_group_size(CL_SIZE, CL_SIZE, 1)))
  90 #ifdef PRUNE_NBL
  91     #ifdef CALC_ENERGIES
  92         __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_prune_opencl)
  93     #else
  94         __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_prune_opencl)
  95     #endif
  96 #else
  97     #ifdef CALC_ENERGIES
  98         __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_opencl)
  99     #else
 100         __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_opencl)
 101     #endif
 102 #endif
 103 (
 104 #ifndef LJ_COMB
 105  int ntypes,                                                               /* IN  */
 106 #endif
 107  cl_nbparam_params_t nbparam_params,                                       /* IN  */
 108  const __global float4 *restrict xq,                                       /* IN  */
 109  __global float *restrict f,                /* stores float3 values */     /* OUT */
 110  __global float *restrict e_lj,                                            /* OUT */
 111  __global float *restrict e_el,                                            /* OUT */
 112 __global float *restrict fshift,            /* stores float3 values */     /* OUT */
 113 #ifdef LJ_COMB
 114  const __global float2 *restrict lj_comb,      /* stores float2 values */  /* IN  */
 115 #else
 116  const __global int *restrict atom_types,                                  /* IN  */
 117 #endif
 118  const __global float *restrict shift_vec,  /* stores float3 values */     /* IN  */
 119  __constant float* nbfp_climg2d,                                           /* IN  */
 120  __constant float* nbfp_comb_climg2d,                                      /* IN  */
 121  __constant float* coulomb_tab_climg2d,                                    /* IN  */
 122  const __global nbnxn_sci_t* pl_sci,                                       /* IN  */
 123 #ifndef PRUNE_NBL
 124     const
 125 #endif
 126  __global nbnxn_cj4_t* pl_cj4,                                             /* OUT / IN */
 127  const __global nbnxn_excl_t* excl,                                        /* IN  */
 128  int bCalcFshift,                                                          /* IN  */
 129  __local  float4   *xqib,                                                  /* Pointer to dyn alloc'ed shmem */
 130  __global float *debug_buffer                                              /* Debug buffer, can be used with print_to_debug_buffer_f */
 131  )
 132 {
 133     /* convenience variables */
 134     cl_nbparam_params_t *nbparam = &nbparam_params;
 135
 136     float               rcoulomb_sq = nbparam->rcoulomb_sq;
 137 #ifdef LJ_COMB
 138     float2              ljcp_i, ljcp_j;
 139 #endif
 140 #ifdef VDW_CUTOFF_CHECK
 141     float               rvdw_sq     = nbparam_params.rvdw_sq;
 142     float               vdw_in_range;
 143 #endif
 144 #ifdef LJ_EWALD
 145     float               lje_coeff2, lje_coeff6_6;
 146 #endif
 147 #ifdef EL_RF
 148     float two_k_rf              = nbparam->two_k_rf;
 149 #endif
 150 #ifdef EL_EWALD_TAB
 151     float coulomb_tab_scale     = nbparam->coulomb_tab_scale;
 152 #endif
 153 #ifdef EL_EWALD_ANA
 154     float beta2                 = nbparam->ewald_beta*nbparam->ewald_beta;
 155     float beta3                 = nbparam->ewald_beta*nbparam->ewald_beta*nbparam->ewald_beta;
 156 #endif
 157 #ifdef PRUNE_NBL
 158     float rlist_sq              = nbparam->rlistOuter_sq;
 159 #endif
 160
 161 #ifdef CALC_ENERGIES
 162 #ifdef EL_EWALD_ANY
 163     float  beta        = nbparam->ewald_beta;
 164     float  ewald_shift = nbparam->sh_ewald;
 165 #else
 166     float  c_rf        = nbparam->c_rf;
 167 #endif /* EL_EWALD_ANY */
 168 #endif /* CALC_ENERGIES */
 169
 170     /* thread/block/warp id-s */
 171     unsigned int tidxi  = get_local_id(0);
 172     unsigned int tidxj  = get_local_id(1);
 173     unsigned int tidx   = get_local_id(1) * get_local_size(0) + get_local_id(0);
 174     unsigned int bidx   = get_group_id(0);
 175     unsigned int widx   = tidx / WARP_SIZE; /* warp index */
 176     int          sci, ci, cj, ci_offset,
 177                  ai, aj,
 178                  cij4_start, cij4_end;
 179 #ifndef LJ_COMB
 180     int          typei, typej;
 181 #endif
 182     int          i, jm, j4, wexcl_idx;
 183     float        qi, qj_f,
 184                  r2, inv_r, inv_r2;
 185 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
 186     float        inv_r6, c6, c12;
 187 #endif
 188 #if defined LJ_COMB_LB
 189     float        sigma, epsilon;
 190 #endif
 191     float        int_bit,
 192                  F_invr;
 193
 194 #ifdef CALC_ENERGIES
 195     float        E_lj, E_el;
 196 #endif
 197 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
 198     float        E_lj_p;
 199 #endif
 200     unsigned int wexcl, imask, mask_ji;
 201     float4       xqbuf;
 202     float3       xi, xj, rv, f_ij, fcj_buf/*, fshift_buf*/;
 203     float        fshift_buf;
 204     float3       fci_buf[NCL_PER_SUPERCL]; /* i force buffer */
 205     nbnxn_sci_t  nb_sci;
 206
 207     /*! i-cluster interaction mask for a super-cluster with all NCL_PER_SUPERCL=8 bits set */
 208     const unsigned superClInteractionMask = ((1U << NCL_PER_SUPERCL) - 1U);
 209
 210     /* shmem buffer for cj, for both warps separately */
 211     __local int *cjs       = (__local int *)(xqib + NCL_PER_SUPERCL * CL_SIZE);
 212     #define LOCAL_OFFSET cjs + 2 * NBNXN_GPU_JGROUP_SIZE
 213
 214 #ifdef IATYPE_SHMEM
 215 #ifndef LJ_COMB
 216     /* shmem buffer for i atom-type pre-loading */
 217     __local int *atib      = (__local int *)(LOCAL_OFFSET);
 218     #undef LOCAL_OFFSET
 219     #define LOCAL_OFFSET atib + NCL_PER_SUPERCL * CL_SIZE
 220 #else
 221     __local float2 *ljcpib = (__local float2 *)(LOCAL_OFFSET);
 222     #undef LOCAL_OFFSET
 223     #define LOCAL_OFFSET ljcpib + NCL_PER_SUPERCL * CL_SIZE
 224 #endif
 225 #endif
 226
 227 #ifndef REDUCE_SHUFFLE
 228     /* shmem j force buffer */
 229     __local float *f_buf   = (__local float *)(LOCAL_OFFSET);
 230     #undef LOCAL_OFFSET
 231     #define LOCAL_OFFSET f_buf + CL_SIZE * CL_SIZE * 3
 232 #endif
 233     /* Local buffer used to implement __any warp vote function from CUDA.
 234        volatile is used to avoid compiler optimizations for AMD builds. */
 235     volatile __local uint *warp_any = (__local uint*)(LOCAL_OFFSET);
 236 #undef LOCAL_OFFSET
 237
 238     nb_sci      = pl_sci[bidx];         /* my i super-cluster's index = current bidx */
 239     sci         = nb_sci.sci;           /* super-cluster */
 240     cij4_start  = nb_sci.cj4_ind_start; /* first ...*/
 241     cij4_end    = nb_sci.cj4_ind_end;   /* and last index of j clusters */
 242
 243     /* Pre-load i-atom x and q into shared memory */
 244     ci = sci * NCL_PER_SUPERCL + tidxj;
 245     ai = ci * CL_SIZE + tidxi;
 246
 247     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);
 248     xqbuf.w *= nbparam->epsfac;
 249     xqib[tidxj * CL_SIZE + tidxi] = xqbuf;
 250
 251 #ifdef IATYPE_SHMEM
 252 #ifndef LJ_COMB
 253     /* Pre-load the i-atom types into shared memory */
 254     atib[tidxj * CL_SIZE + tidxi]   = atom_types[ai];
 255 #else
 256     ljcpib[tidxj * CL_SIZE + tidxi] = lj_comb[ai];
 257 #endif
 258 #endif
 259     /* Initialise warp vote. (8x8 block) 2 warps for nvidia */
 260     if(tidx==0 || tidx==32)
 261         warp_any[widx] = 0;
 262
 263     barrier(CLK_LOCAL_MEM_FENCE);
 264
 265     for (ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
 266     {
 267         fci_buf[ci_offset] = (float3)(0.0f);
 268     }
 269
 270 #ifdef LJ_EWALD
 271     /* TODO: we are trading registers with flops by keeping lje_coeff-s, try re-calculating it later */
 272     lje_coeff2   = nbparam->ewaldcoeff_lj*nbparam->ewaldcoeff_lj;
 273     lje_coeff6_6 = lje_coeff2*lje_coeff2*lje_coeff2*ONE_SIXTH_F;
 274 #endif /* LJ_EWALD */
 275
 276
 277 #ifdef CALC_ENERGIES
 278     E_lj = 0.0f;
 279     E_el = 0.0f;
 280
 281 #if defined EXCLUSION_FORCES /* Ewald or RF */
 282     if (nb_sci.shift == CENTRAL && pl_cj4[cij4_start].cj[0] == sci*NCL_PER_SUPERCL)
 283     {
 284         /* we have the diagonal: add the charge and LJ self interaction energy term */
 285         for (i = 0; i < NCL_PER_SUPERCL; i++)
 286         {
 287 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
 288             qi    = xqib[i * CL_SIZE + tidxi].w;
 289             E_el += qi*qi;
 290 #endif
 291 #if defined LJ_EWALD
 292             E_lj += nbfp_climg2d[atom_types[(sci*NCL_PER_SUPERCL + i)*CL_SIZE + tidxi]*(ntypes + 1)*2];
 293 #endif /* LJ_EWALD */
 294         }
 295
 296         /* divide the self term(s) equally over the j-threads, then multiply with the coefficients. */
 297 #ifdef LJ_EWALD
 298         E_lj /= CL_SIZE;
 299         E_lj *= 0.5f*ONE_SIXTH_F*lje_coeff6_6;
 300 #endif  /* LJ_EWALD */
 301
 302 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
 303         /* Correct for epsfac^2 due to adding qi^2 */
 304         E_el /= nbparam->epsfac*CL_SIZE;
 305 #if defined EL_RF || defined EL_CUTOFF
 306         E_el *= -0.5f*c_rf;
 307 #else
 308         E_el *= -beta*M_FLOAT_1_SQRTPI; /* last factor 1/sqrt(pi) */
 309 #endif
 310 #endif                                  /* EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF */
 311     }
 312 #endif                                  /* EXCLUSION_FORCES */
 313
 314 #endif                                  /* CALC_ENERGIES */
 315
 316 #ifdef EXCLUSION_FORCES
 317     const int nonSelfInteraction  = !(nb_sci.shift == CENTRAL & tidxj <= tidxi);
 318 #endif
 319
 320     /* loop over the j clusters = seen by any of the atoms in the current super-cluster */
 321     for (j4 = cij4_start; j4 < cij4_end; j4++)
 322     {
 323         wexcl_idx   = pl_cj4[j4].imei[widx].excl_ind;
 324         imask       = pl_cj4[j4].imei[widx].imask;
 325         wexcl       = excl[wexcl_idx].pair[(tidx) & (WARP_SIZE - 1)];
 326
 327 #ifndef PRUNE_NBL
 328         if (imask)
 329 #endif
 330         {
 331             /* Pre-load cj into shared memory on both warps separately */
 332             if ((tidxj == 0 | tidxj == 4) & (tidxi < NBNXN_GPU_JGROUP_SIZE))
 333             {
 334                 cjs[tidxi + tidxj * NBNXN_GPU_JGROUP_SIZE / 4] = pl_cj4[j4].cj[tidxi];
 335             }
 336
 337             /* TODO: check loop unrolling with NVIDIA OpenCL */
 338 #if !defined PRUNE_NBL
 339 //#pragma unroll 4
 340 #endif
 341
 342             for (jm = 0; jm < NBNXN_GPU_JGROUP_SIZE; jm++)
 343             {
 344                 if (imask & (superClInteractionMask << (jm * NCL_PER_SUPERCL)))
 345                 {
 346                     mask_ji = (1U << (jm * NCL_PER_SUPERCL));
 347
 348                     cj      = cjs[jm + (tidxj & 4) * NBNXN_GPU_JGROUP_SIZE / 4];
 349                     aj      = cj * CL_SIZE + tidxj;
 350
 351                     /* load j atom data */
 352                     xqbuf   = xq[aj];
 353                     xj      = (float3)(xqbuf.xyz);
 354                     qj_f    = xqbuf.w;
 355 #ifndef LJ_COMB
 356                     typej   = atom_types[aj];
 357 #else
 358                     ljcp_j  = lj_comb[aj];
 359 #endif
 360
 361                     fcj_buf = (float3)(0.0f);
 362
 363 #if !defined PRUNE_NBL
 364 #pragma unroll 8
 365 #endif
 366                     for (i = 0; i < NCL_PER_SUPERCL; i++)
 367                     {
 368                         if (imask & mask_ji)
 369                         {
 370                             ci_offset   = i;                     /* i force buffer offset */
 371
 372                             ci      = sci * NCL_PER_SUPERCL + i; /* i cluster index */
 373                             ai      = ci * CL_SIZE + tidxi;      /* i atom index */
 374
 375                             /* all threads load an atom from i cluster ci into shmem! */
 376                             xqbuf   = xqib[i * CL_SIZE + tidxi];
 377                             xi      = (float3)(xqbuf.xyz);
 378
 379                             /* distance between i and j atoms */
 380                             rv      = xi - xj;
 381                             r2      = norm2(rv);
 382
 383 #ifdef PRUNE_NBL
 384                             /* vote.. should code shmem serialisation, wonder what the hit will be */
 385                             if (r2 < rlist_sq)
 386                                 warp_any[widx]=1;
 387
 388                             /* If _none_ of the atoms pairs are in cutoff range,
 389                                the bit corresponding to the current
 390                                cluster-pair in imask gets set to 0. */
 391                             if (!warp_any[widx])
 392                                 imask &= ~mask_ji;
 393
 394                             warp_any[widx]=0;
 395 #endif
 396
 397                             int_bit = (wexcl & mask_ji) ? 1.0f : 0.0f;
 398
 399                             /* cutoff & exclusion check */
 400 #ifdef EXCLUSION_FORCES
 401                             if ((r2 < rcoulomb_sq) * (nonSelfInteraction | (ci != cj)))
 402 #else
 403                             if ((r2 < rcoulomb_sq) * int_bit)
 404 #endif
 405                             {
 406                                 /* load the rest of the i-atom parameters */
 407                                 qi      = xqbuf.w;
 408 #ifdef IATYPE_SHMEM
 409 #ifndef LJ_COMB
 410                                 typei   = atib[i * CL_SIZE + tidxi];
 411 #else
 412                                 ljcp_i  = ljcpib[i * CL_SIZE + tidxi];
 413 #endif
 414 #else /* IATYPE_SHMEM */
 415 #ifndef LJ_COMB
 416                                 typei   = atom_types[ai];
 417 #else
 418                                 ljcp_i  = lj_comb[ai];
 419 #endif
 420 #endif /* IATYPE_SHMEM */
 421                                 /* LJ 6*C6 and 12*C12 */
 422 #ifndef LJ_COMB
 423                                 c6      = nbfp_climg2d[2 * (ntypes * typei + typej)];
 424                                 c12     = nbfp_climg2d[2 * (ntypes * typei + typej)+1];
 425 #else   /* LJ_COMB */
 426 #ifdef LJ_COMB_GEOM
 427                                 c6      = ljcp_i.x * ljcp_j.x;
 428                                 c12     = ljcp_i.y * ljcp_j.y;
 429 #else
 430                                 /* LJ 2^(1/6)*sigma and 12*epsilon */
 431                                 sigma   = ljcp_i.x + ljcp_j.x;
 432                                 epsilon = ljcp_i.y * ljcp_j.y;
 433 #if defined CALC_ENERGIES || defined LJ_FORCE_SWITCH || defined LJ_POT_SWITCH
 434                                 convert_sigma_epsilon_to_c6_c12(sigma, epsilon, &c6, &c12);
 435 #endif
 436 #endif                          /* LJ_COMB_GEOM */
 437 #endif                          /* LJ_COMB */
 438
 439                                 // Ensure distance do not become so small that r^-12 overflows
 440                                 r2      = max(r2,NBNXN_MIN_RSQ);
 441
 442                                 inv_r   = rsqrt(r2);
 443                                 inv_r2  = inv_r * inv_r;
 444 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
 445                                 inv_r6  = inv_r2 * inv_r2 * inv_r2;
 446 #if defined EXCLUSION_FORCES
 447                                 /* We could mask inv_r2, but with Ewald
 448                                  * masking both inv_r6 and F_invr is faster */
 449                                 inv_r6  *= int_bit;
 450 #endif                          /* EXCLUSION_FORCES */
 451
 452                                 F_invr  = inv_r6 * (c12 * inv_r6 - c6) * inv_r2;
 453 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
 454                                 E_lj_p  = int_bit * (c12 * (inv_r6 * inv_r6 + nbparam->repulsion_shift.cpot)*ONE_TWELVETH_F -
 455                                                      c6 * (inv_r6 + nbparam->dispersion_shift.cpot)*ONE_SIXTH_F);
 456
 457 #endif
 458 #else                           /* ! LJ_COMB_LB || CALC_ENERGIES */
 459                                 float sig_r  = sigma*inv_r;
 460                                 float sig_r2 = sig_r*sig_r;
 461                                 float sig_r6 = sig_r2*sig_r2*sig_r2;
 462 #if defined EXCLUSION_FORCES
 463                                 sig_r6 *= int_bit;
 464 #endif                          /* EXCLUSION_FORCES */
 465
 466                                 F_invr  = epsilon * sig_r6 * (sig_r6 - 1.0f) * inv_r2;
 467 #endif                          /* ! LJ_COMB_LB || CALC_ENERGIES */
 468
 469
 470 #ifdef LJ_FORCE_SWITCH
 471 #ifdef CALC_ENERGIES
 472                                 calculate_force_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
 473 #else
 474                                 calculate_force_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr);
 475 #endif /* CALC_ENERGIES */
 476 #endif /* LJ_FORCE_SWITCH */
 477
 478
 479 #ifdef LJ_EWALD
 480 #ifdef LJ_EWALD_COMB_GEOM
 481 #ifdef CALC_ENERGIES
 482                                 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);
 483 #else
 484                                 calculate_lj_ewald_comb_geom_F(nbfp_comb_climg2d, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6, &F_invr);
 485 #endif                          /* CALC_ENERGIES */
 486 #elif defined LJ_EWALD_COMB_LB
 487                                 calculate_lj_ewald_comb_LB_F_E(nbfp_comb_climg2d, nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6,
 488 #ifdef CALC_ENERGIES
 489                                                                int_bit, true, &F_invr, &E_lj_p
 490 #else
 491                                                                0, false, &F_invr, 0
 492 #endif /* CALC_ENERGIES */
 493                                                                );
 494 #endif /* LJ_EWALD_COMB_GEOM */
 495 #endif /* LJ_EWALD */
 496
 497 #ifdef LJ_POT_SWITCH
 498 #ifdef CALC_ENERGIES
 499                                 calculate_potential_switch_F_E(nbparam, inv_r, r2, &F_invr, &E_lj_p);
 500 #else
 501                                 calculate_potential_switch_F(nbparam, inv_r, r2, &F_invr, &E_lj_p);
 502 #endif /* CALC_ENERGIES */
 503 #endif /* LJ_POT_SWITCH */
 504
 505 #ifdef VDW_CUTOFF_CHECK
 506                                 /* Separate VDW cut-off check to enable twin-range cut-offs
 507                                  * (rvdw < rcoulomb <= rlist)
 508                                  */
 509                                 vdw_in_range  = (r2 < rvdw_sq) ? 1.0f : 0.0f;
 510                                 F_invr       *= vdw_in_range;
 511 #ifdef CALC_ENERGIES
 512                                 E_lj_p       *= vdw_in_range;
 513 #endif
 514 #endif                          /* VDW_CUTOFF_CHECK */
 515
 516 #ifdef CALC_ENERGIES
 517                                 E_lj    += E_lj_p;
 518
 519 #endif
 520
 521
 522 #ifdef EL_CUTOFF
 523 #ifdef EXCLUSION_FORCES
 524                                 F_invr  += qi * qj_f * int_bit * inv_r2 * inv_r;
 525 #else
 526                                 F_invr  += qi * qj_f * inv_r2 * inv_r;
 527 #endif
 528 #endif
 529 #ifdef EL_RF
 530                                 F_invr  += qi * qj_f * (int_bit*inv_r2 * inv_r - two_k_rf);
 531 #endif
 532 #if defined EL_EWALD_ANA
 533                                 F_invr  += qi * qj_f * (int_bit*inv_r2*inv_r + pmecorrF(beta2*r2)*beta3);
 534 #elif defined EL_EWALD_TAB
 535                                 F_invr  += qi * qj_f * (int_bit*inv_r2 -
 536 #ifdef USE_TEXOBJ
 537                                                         interpolate_coulomb_force_r(nbparam->coulomb_tab_texobj, r2 * inv_r, coulomb_tab_scale)
 538 #else
 539                                                         interpolate_coulomb_force_r(coulomb_tab_climg2d, r2 * inv_r, coulomb_tab_scale)
 540 #endif /* USE_TEXOBJ */
 541                                                         ) * inv_r;
 542 #endif /* EL_EWALD_ANA/TAB */
 543
 544 #ifdef CALC_ENERGIES
 545 #ifdef EL_CUTOFF
 546                                 E_el    += qi * qj_f * (int_bit*inv_r - c_rf);
 547 #endif
 548 #ifdef EL_RF
 549                                 E_el    += qi * qj_f * (int_bit*inv_r + 0.5f * two_k_rf * r2 - c_rf);
 550 #endif
 551 #ifdef EL_EWALD_ANY
 552                                 /* 1.0f - erff is faster than erfcf */
 553                                 E_el    += qi * qj_f * (inv_r * (int_bit - erf(r2 * inv_r * beta)) - int_bit * ewald_shift);
 554 #endif                          /* EL_EWALD_ANY */
 555 #endif
 556                                 f_ij    = rv * F_invr;
 557
 558                                 /* accumulate j forces in registers */
 559                                 fcj_buf -= f_ij;
 560
 561                                 /* accumulate i forces in registers */
 562                                 fci_buf[ci_offset] += f_ij;
 563                             }
 564                         }
 565
 566                         /* shift the mask bit by 1 */
 567                         mask_ji += mask_ji;
 568                     }
 569
 570                     /* reduce j forces */
 571
 572                     /* store j forces in shmem */
 573                     f_buf[                  tidx] = fcj_buf.x;
 574                     f_buf[    FBUF_STRIDE + tidx] = fcj_buf.y;
 575                     f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
 576
 577                     reduce_force_j_generic(f_buf, f, tidxi, tidxj, aj);
 578                 }
 579             }
 580 #ifdef PRUNE_NBL
 581             /* Update the imask with the new one which does not contain the
 582                out of range clusters anymore. */
 583
 584             pl_cj4[j4].imei[widx].imask = imask;
 585 #endif
 586         }
 587     }
 588
 589     /* skip central shifts when summing shift forces */
 590     if (nb_sci.shift == CENTRAL)
 591     {
 592         bCalcFshift = false;
 593     }
 594
 595     fshift_buf = 0.0f;
 596
 597     /* reduce i forces */
 598     for (ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
 599     {
 600         ai  = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
 601
 602         f_buf[                  tidx] = fci_buf[ci_offset].x;
 603         f_buf[    FBUF_STRIDE + tidx] = fci_buf[ci_offset].y;
 604         f_buf[2 * FBUF_STRIDE + tidx] = fci_buf[ci_offset].z;
 605         barrier(CLK_LOCAL_MEM_FENCE);
 606         reduce_force_i(f_buf, f,
 607                        &fshift_buf, bCalcFshift,
 608                        tidxi, tidxj, ai);
 609         barrier(CLK_LOCAL_MEM_FENCE);
 610     }
 611
 612     /* add up local shift forces into global mem */
 613     if (bCalcFshift)
 614     {
 615         /* Only threads with tidxj < 3 will update fshift.
 616            The threads performing the update must be the same with the threads
 617            which stored the reduction result in reduce_force_i function
 618         */
 619         if (tidxj < 3)
 620             atomicAdd_g_f(&(fshift[3 * nb_sci.shift + tidxj]), fshift_buf);
 621     }
 622
 623 #ifdef CALC_ENERGIES
 624     /* flush the energies to shmem and reduce them */
 625     f_buf[              tidx] = E_lj;
 626     f_buf[FBUF_STRIDE + tidx] = E_el;
 627     reduce_energy_pow2(f_buf + (tidx & WARP_SIZE), e_lj, e_el, tidx & ~WARP_SIZE);
 628
 629 #endif
 630 }
 631
 632 #undef EL_EWALD_ANY
 633 #undef EXCLUSION_FORCES
 634 #undef LJ_EWALD
 635
 636 #undef LJ_COMB