src/gromacs/mdlib/nbnxn_ocl/nbnxn_ocl_kernel_amd.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 AMD GPUs.
  38  *
  39  *  OpenCL 1.2 support is expected.
  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   TODO: implement 128 threads/wavefront by porting over the NTHREAD_Z/j4 loop
  85   "horizontal splitting" over threads.
  86  */
  87
  88 /* NOTE:
  89  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.
  90  Thus if more strings need to be appended a new macro must be written or it must be directly appended here.
  91 */
  92 __attribute__((reqd_work_group_size(CL_SIZE, CL_SIZE, 1)))
  93 #ifdef PRUNE_NBL
  94     #ifdef CALC_ENERGIES
  95         __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_prune_opencl)
  96     #else
  97         __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_prune_opencl)
  98     #endif
  99 #else
 100     #ifdef CALC_ENERGIES
 101         __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_opencl)
 102     #else
 103         __kernel void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_opencl)
 104     #endif
 105 #endif
 106 (
 107 #ifndef LJ_COMB
 108  int ntypes,                                                               /* IN  */
 109 #endif
 110  cl_nbparam_params_t nbparam_params,                                       /* IN  */
 111  const __global float4 *restrict xq,                                       /* IN  */
 112  __global float *restrict f,                /* stores float3 values */     /* OUT */
 113  __global float *restrict e_lj,                                            /* OUT */
 114  __global float *restrict e_el,                                            /* OUT */
 115 __global float *restrict fshift,            /* stores float3 values */     /* OUT */
 116 #ifdef LJ_COMB
 117  const __global float2 *restrict lj_comb,      /* stores float2 values */  /* IN  */
 118 #else
 119  const __global int *restrict atom_types,                                  /* IN  */
 120 #endif
 121  const __global float *restrict shift_vec,  /* stores float3 values */     /* IN  */
 122  __constant float* nbfp_climg2d,                                           /* IN  */
 123  __constant float* nbfp_comb_climg2d,                                      /* IN  */
 124  __constant float* coulomb_tab_climg2d,                                    /* IN  */
 125  const __global nbnxn_sci_t* pl_sci,                                       /* IN  */
 126 #ifndef PRUNE_NBL
 127     const
 128 #endif
 129  __global nbnxn_cj4_t* pl_cj4,                                             /* OUT / IN */
 130  const __global nbnxn_excl_t* excl,                                        /* IN  */
 131  int bCalcFshift,                                                          /* IN  */
 132  __local  float4   *xqib,                                                  /* Pointer to dyn alloc'ed shmem */
 133  __global float *debug_buffer                                              /* Debug buffer, can be used with print_to_debug_buffer_f */
 134  )
 135 {
 136     /* convenience variables */
 137     cl_nbparam_params_t *nbparam = &nbparam_params;
 138
 139     float               rcoulomb_sq = nbparam->rcoulomb_sq;
 140 #ifdef LJ_COMB
 141     float2              ljcp_i, ljcp_j;
 142 #endif
 143 #ifdef VDW_CUTOFF_CHECK
 144     float               rvdw_sq     = nbparam_params.rvdw_sq;
 145     float               vdw_in_range;
 146 #endif
 147 #ifdef LJ_EWALD
 148     float               lje_coeff2, lje_coeff6_6;
 149 #endif
 150 #ifdef EL_RF
 151     float two_k_rf              = nbparam->two_k_rf;
 152 #endif
 153 #ifdef EL_EWALD_TAB
 154     float coulomb_tab_scale     = nbparam->coulomb_tab_scale;
 155 #endif
 156 #ifdef EL_EWALD_ANA
 157     float beta2                 = nbparam->ewald_beta*nbparam->ewald_beta;
 158     float beta3                 = nbparam->ewald_beta*nbparam->ewald_beta*nbparam->ewald_beta;
 159 #endif
 160 #ifdef PRUNE_NBL
 161     float rlist_sq              = nbparam->rlistOuter_sq;
 162 #endif
 163
 164 #ifdef CALC_ENERGIES
 165 #ifdef EL_EWALD_ANY
 166     float  beta        = nbparam->ewald_beta;
 167     float  ewald_shift = nbparam->sh_ewald;
 168 #else
 169     float  c_rf        = nbparam->c_rf;
 170 #endif /* EL_EWALD_ANY */
 171 #endif /* CALC_ENERGIES */
 172
 173     /* thread/block/warp id-s */
 174     unsigned int tidxi  = get_local_id(0);
 175     unsigned int tidxj  = get_local_id(1);
 176     unsigned int tidx   = get_local_id(1) * get_local_size(0) + get_local_id(0);
 177     unsigned int bidx   = get_group_id(0);
 178     unsigned int widx   = tidx / WARP_SIZE; /* warp index */
 179     int          sci, ci, cj, ci_offset,
 180                  ai, aj,
 181                  cij4_start, cij4_end;
 182 #ifndef LJ_COMB
 183     int          typei, typej;
 184 #endif
 185     int          i, jm, j4, wexcl_idx;
 186     float        qi, qj_f,
 187                  r2, inv_r, inv_r2;
 188 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
 189     float        inv_r6, c6, c12;
 190 #endif
 191 #if defined LJ_COMB_LB
 192     float        sigma, epsilon;
 193 #endif
 194     float        int_bit,
 195                  F_invr;
 196
 197 #ifdef CALC_ENERGIES
 198     float        E_lj, E_el;
 199 #endif
 200 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
 201     float        E_lj_p;
 202 #endif
 203     unsigned int wexcl, imask, mask_ji;
 204     float4       xqbuf;
 205     float3       xi, xj, rv, f_ij, fcj_buf/*, fshift_buf*/;
 206     float        fshift_buf;
 207     float3       fci_buf[NCL_PER_SUPERCL]; /* i force buffer */
 208     nbnxn_sci_t  nb_sci;
 209
 210     /*! i-cluster interaction mask for a super-cluster with all NCL_PER_SUPERCL=8 bits set */
 211     const unsigned superClInteractionMask = ((1U << NCL_PER_SUPERCL) - 1U);
 212
 213     /* shmem buffer for cj, for both warps separately */
 214     __local int *cjs       = (__local int *)(xqib + NCL_PER_SUPERCL * CL_SIZE);
 215     #define LOCAL_OFFSET cjs + 2 * NBNXN_GPU_JGROUP_SIZE
 216
 217 #ifdef IATYPE_SHMEM
 218 #ifndef LJ_COMB
 219     /* shmem buffer for i atom-type pre-loading */
 220     __local int *atib      = (__local int *)(LOCAL_OFFSET);
 221     #undef LOCAL_OFFSET
 222     #define LOCAL_OFFSET atib + NCL_PER_SUPERCL * CL_SIZE
 223 #else
 224     __local float2 *ljcpib = (__local float2 *)(LOCAL_OFFSET);
 225     #undef LOCAL_OFFSET
 226     #define LOCAL_OFFSET ljcpib + NCL_PER_SUPERCL * CL_SIZE
 227 #endif
 228 #endif
 229
 230 #ifndef REDUCE_SHUFFLE
 231     /* shmem j force buffer */
 232     __local float *f_buf   = (__local float *)(LOCAL_OFFSET);
 233     #undef LOCAL_OFFSET
 234     #define LOCAL_OFFSET f_buf + CL_SIZE * CL_SIZE * 3
 235 #endif
 236     /* Local buffer used to implement __any warp vote function from CUDA.
 237        volatile is used to avoid compiler optimizations for AMD builds. */
 238     volatile __local uint *warp_any = (__local uint*)(LOCAL_OFFSET);
 239 #undef LOCAL_OFFSET
 240
 241     nb_sci      = pl_sci[bidx];         /* my i super-cluster's index = current bidx */
 242     sci         = nb_sci.sci;           /* super-cluster */
 243     cij4_start  = nb_sci.cj4_ind_start; /* first ...*/
 244     cij4_end    = nb_sci.cj4_ind_end;   /* and last index of j clusters */
 245
 246     /* Pre-load i-atom x and q into shared memory */
 247     ci = sci * NCL_PER_SUPERCL + tidxj;
 248     ai = ci * CL_SIZE + tidxi;
 249
 250     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);
 251     xqbuf.w *= nbparam->epsfac;
 252     xqib[tidxj * CL_SIZE + tidxi] = xqbuf;
 253
 254 #ifdef IATYPE_SHMEM
 255 #ifndef LJ_COMB
 256     /* Pre-load the i-atom types into shared memory */
 257     atib[tidxj * CL_SIZE + tidxi]   = atom_types[ai];
 258 #else
 259     ljcpib[tidxj * CL_SIZE + tidxi] = lj_comb[ai];
 260 #endif
 261 #endif
 262     /* Initialise warp vote. (8x8 block) 2 warps for nvidia */
 263     if(tidx==0 || tidx==32)
 264         warp_any[widx] = 0;
 265
 266     barrier(CLK_LOCAL_MEM_FENCE);
 267
 268     for (ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
 269     {
 270         fci_buf[ci_offset] = (float3)(0.0f);
 271     }
 272
 273 #ifdef LJ_EWALD
 274     /* TODO: we are trading registers with flops by keeping lje_coeff-s, try re-calculating it later */
 275     lje_coeff2   = nbparam->ewaldcoeff_lj*nbparam->ewaldcoeff_lj;
 276     lje_coeff6_6 = lje_coeff2*lje_coeff2*lje_coeff2*ONE_SIXTH_F;
 277 #endif /* LJ_EWALD */
 278
 279
 280 #ifdef CALC_ENERGIES
 281     E_lj = 0.0f;
 282     E_el = 0.0f;
 283
 284 #if defined EXCLUSION_FORCES /* Ewald or RF */
 285     if (nb_sci.shift == CENTRAL && pl_cj4[cij4_start].cj[0] == sci*NCL_PER_SUPERCL)
 286     {
 287         /* we have the diagonal: add the charge and LJ self interaction energy term */
 288         for (i = 0; i < NCL_PER_SUPERCL; i++)
 289         {
 290 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
 291             qi    = xqib[i * CL_SIZE + tidxi].w;
 292             E_el += qi*qi;
 293 #endif
 294 #if defined LJ_EWALD
 295             E_lj += nbfp_climg2d[atom_types[(sci*NCL_PER_SUPERCL + i)*CL_SIZE + tidxi]*(ntypes + 1)*2];
 296 #endif /* LJ_EWALD */
 297         }
 298
 299         /* divide the self term(s) equally over the j-threads, then multiply with the coefficients. */
 300 #ifdef LJ_EWALD
 301         E_lj /= CL_SIZE;
 302         E_lj *= 0.5f*ONE_SIXTH_F*lje_coeff6_6;
 303 #endif  /* LJ_EWALD */
 304
 305 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
 306         /* Correct for epsfac^2 due to adding qi^2 */
 307         E_el /= nbparam->epsfac*CL_SIZE;
 308 #if defined EL_RF || defined EL_CUTOFF
 309         E_el *= -0.5f*c_rf;
 310 #else
 311         E_el *= -beta*M_FLOAT_1_SQRTPI; /* last factor 1/sqrt(pi) */
 312 #endif
 313 #endif                                  /* EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF */
 314     }
 315 #endif                                  /* EXCLUSION_FORCES */
 316
 317 #endif                                  /* CALC_ENERGIES */
 318
 319 #ifdef EXCLUSION_FORCES
 320     const int nonSelfInteraction  = !(nb_sci.shift == CENTRAL & tidxj <= tidxi);
 321 #endif
 322
 323     /* loop over the j clusters = seen by any of the atoms in the current super-cluster */
 324     for (j4 = cij4_start; j4 < cij4_end; j4++)
 325     {
 326         wexcl_idx   = pl_cj4[j4].imei[widx].excl_ind;
 327         imask       = pl_cj4[j4].imei[widx].imask;
 328         wexcl       = excl[wexcl_idx].pair[(tidx) & (WARP_SIZE - 1)];
 329
 330 #ifndef PRUNE_NBL
 331         if (imask)
 332 #endif
 333         {
 334             /* Pre-load cj into shared memory on both warps separately */
 335             if ((tidxj == 0 | tidxj == 4) & (tidxi < NBNXN_GPU_JGROUP_SIZE))
 336             {
 337                 cjs[tidxi + tidxj * NBNXN_GPU_JGROUP_SIZE / 4] = pl_cj4[j4].cj[tidxi];
 338             }
 339
 340             /* Unrolling this loop improves performance without pruning but
 341              * with pruning it leads to slowdown.
 342              *
 343              * Tested with driver 1800.5
 344              */
 345 #if !defined PRUNE_NBL
 346 #pragma unroll 4
 347 #endif
 348
 349             for (jm = 0; jm < NBNXN_GPU_JGROUP_SIZE; jm++)
 350             {
 351                 if (imask & (superClInteractionMask << (jm * NCL_PER_SUPERCL)))
 352                 {
 353                     mask_ji = (1U << (jm * NCL_PER_SUPERCL));
 354
 355                     cj      = cjs[jm + (tidxj & 4) * NBNXN_GPU_JGROUP_SIZE / 4];
 356                     aj      = cj * CL_SIZE + tidxj;
 357
 358                     /* load j atom data */
 359                     xqbuf   = xq[aj];
 360                     xj      = (float3)(xqbuf.xyz);
 361                     qj_f    = xqbuf.w;
 362 #ifndef LJ_COMB
 363                     typej   = atom_types[aj];
 364 #else
 365                     ljcp_j  = lj_comb[aj];
 366 #endif
 367
 368                     fcj_buf = (float3)(0.0f);
 369
 370 #if !defined PRUNE_NBL
 371 #pragma unroll 8
 372 #endif
 373                     for (i = 0; i < NCL_PER_SUPERCL; i++)
 374                     {
 375                         if (imask & mask_ji)
 376                         {
 377                             ci_offset   = i;                     /* i force buffer offset */
 378
 379                             ci      = sci * NCL_PER_SUPERCL + i; /* i cluster index */
 380                             ai      = ci * CL_SIZE + tidxi;      /* i atom index */
 381
 382                             /* all threads load an atom from i cluster ci into shmem! */
 383                             xqbuf   = xqib[i * CL_SIZE + tidxi];
 384                             xi      = (float3)(xqbuf.xyz);
 385
 386                             /* distance between i and j atoms */
 387                             rv      = xi - xj;
 388                             r2      = norm2(rv);
 389
 390 #ifdef PRUNE_NBL
 391                             /* vote.. should code shmem serialisation, wonder what the hit will be */
 392                             if (r2 < rlist_sq)
 393                                 warp_any[widx]=1;
 394
 395                             /* If _none_ of the atoms pairs are in cutoff range,
 396                                the bit corresponding to the current
 397                                cluster-pair in imask gets set to 0. */
 398                             if (!warp_any[widx])
 399                                 imask &= ~mask_ji;
 400
 401                             warp_any[widx]=0;
 402 #endif
 403
 404                             int_bit = (wexcl & mask_ji) ? 1.0f : 0.0f;
 405
 406                             /* cutoff & exclusion check */
 407 #ifdef EXCLUSION_FORCES
 408                             if ((r2 < rcoulomb_sq) * (nonSelfInteraction | (ci != cj)))
 409 #else
 410                             if ((r2 < rcoulomb_sq) * int_bit)
 411 #endif
 412                             {
 413                                 /* load the rest of the i-atom parameters */
 414                                 qi      = xqbuf.w;
 415 #ifdef IATYPE_SHMEM
 416 #ifndef LJ_COMB
 417                                 typei   = atib[i * CL_SIZE + tidxi];
 418 #else
 419                                 ljcp_i  = ljcpib[i * CL_SIZE + tidxi];
 420 #endif
 421 #else /* IATYPE_SHMEM */
 422 #ifndef LJ_COMB
 423                                 typei   = atom_types[ai];
 424 #else
 425                                 ljcp_i  = lj_comb[ai];
 426 #endif
 427 #endif /* IATYPE_SHMEM */
 428                                 /* LJ 6*C6 and 12*C12 */
 429 #ifndef LJ_COMB
 430                                 c6      = nbfp_climg2d[2 * (ntypes * typei + typej)];
 431                                 c12     = nbfp_climg2d[2 * (ntypes * typei + typej)+1];
 432 #else   /* LJ_COMB */
 433 #ifdef LJ_COMB_GEOM
 434                                 c6      = ljcp_i.x * ljcp_j.x;
 435                                 c12     = ljcp_i.y * ljcp_j.y;
 436 #else
 437                                 /* LJ 2^(1/6)*sigma and 12*epsilon */
 438                                 sigma   = ljcp_i.x + ljcp_j.x;
 439                                 epsilon = ljcp_i.y * ljcp_j.y;
 440 #if defined CALC_ENERGIES || defined LJ_FORCE_SWITCH || defined LJ_POT_SWITCH
 441                                 convert_sigma_epsilon_to_c6_c12(sigma, epsilon, &c6, &c12);
 442 #endif
 443 #endif                          /* LJ_COMB_GEOM */
 444 #endif                          /* LJ_COMB */
 445
 446                                 // Ensure distance do not become so small that r^-12 overflows
 447                                 r2      = max(r2,NBNXN_MIN_RSQ);
 448
 449                                 inv_r   = rsqrt(r2);
 450                                 inv_r2  = inv_r * inv_r;
 451 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
 452                                 inv_r6  = inv_r2 * inv_r2 * inv_r2;
 453 #if defined EXCLUSION_FORCES
 454                                 /* We could mask inv_r2, but with Ewald
 455                                  * masking both inv_r6 and F_invr is faster */
 456                                 inv_r6  *= int_bit;
 457 #endif                          /* EXCLUSION_FORCES */
 458
 459                                 F_invr  = inv_r6 * (c12 * inv_r6 - c6) * inv_r2;
 460 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
 461                                 E_lj_p  = int_bit * (c12 * (inv_r6 * inv_r6 + nbparam->repulsion_shift.cpot)*ONE_TWELVETH_F -
 462                                                      c6 * (inv_r6 + nbparam->dispersion_shift.cpot)*ONE_SIXTH_F);
 463
 464 #endif
 465 #else                           /* ! LJ_COMB_LB || CALC_ENERGIES */
 466                                 float sig_r  = sigma*inv_r;
 467                                 float sig_r2 = sig_r*sig_r;
 468                                 float sig_r6 = sig_r2*sig_r2*sig_r2;
 469 #if defined EXCLUSION_FORCES
 470                                 sig_r6 *= int_bit;
 471 #endif                          /* EXCLUSION_FORCES */
 472
 473                                 F_invr  = epsilon * sig_r6 * (sig_r6 - 1.0f) * inv_r2;
 474 #endif                          /* ! LJ_COMB_LB || CALC_ENERGIES */
 475
 476
 477 #ifdef LJ_FORCE_SWITCH
 478 #ifdef CALC_ENERGIES
 479                                 calculate_force_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
 480 #else
 481                                 calculate_force_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr);
 482 #endif /* CALC_ENERGIES */
 483 #endif /* LJ_FORCE_SWITCH */
 484
 485
 486 #ifdef LJ_EWALD
 487 #ifdef LJ_EWALD_COMB_GEOM
 488 #ifdef CALC_ENERGIES
 489                                 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);
 490 #else
 491                                 calculate_lj_ewald_comb_geom_F(nbfp_comb_climg2d, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6, &F_invr);
 492 #endif                          /* CALC_ENERGIES */
 493 #elif defined LJ_EWALD_COMB_LB
 494                                 calculate_lj_ewald_comb_LB_F_E(nbfp_comb_climg2d, nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6,
 495 #ifdef CALC_ENERGIES
 496                                                                int_bit, true, &F_invr, &E_lj_p
 497 #else
 498                                                                0, false, &F_invr, 0
 499 #endif /* CALC_ENERGIES */
 500                                                                );
 501 #endif /* LJ_EWALD_COMB_GEOM */
 502 #endif /* LJ_EWALD */
 503
 504 #ifdef LJ_POT_SWITCH
 505 #ifdef CALC_ENERGIES
 506                                 calculate_potential_switch_F_E(nbparam, inv_r, r2, &F_invr, &E_lj_p);
 507 #else
 508                                 calculate_potential_switch_F(nbparam, inv_r, r2, &F_invr, &E_lj_p);
 509 #endif /* CALC_ENERGIES */
 510 #endif /* LJ_POT_SWITCH */
 511
 512 #ifdef VDW_CUTOFF_CHECK
 513                                 /* Separate VDW cut-off check to enable twin-range cut-offs
 514                                  * (rvdw < rcoulomb <= rlist)
 515                                  */
 516                                 vdw_in_range  = (r2 < rvdw_sq) ? 1.0f : 0.0f;
 517                                 F_invr       *= vdw_in_range;
 518 #ifdef CALC_ENERGIES
 519                                 E_lj_p       *= vdw_in_range;
 520 #endif
 521 #endif                          /* VDW_CUTOFF_CHECK */
 522
 523 #ifdef CALC_ENERGIES
 524                                 E_lj    += E_lj_p;
 525
 526 #endif
 527
 528
 529 #ifdef EL_CUTOFF
 530 #ifdef EXCLUSION_FORCES
 531                                 F_invr  += qi * qj_f * int_bit * inv_r2 * inv_r;
 532 #else
 533                                 F_invr  += qi * qj_f * inv_r2 * inv_r;
 534 #endif
 535 #endif
 536 #ifdef EL_RF
 537                                 F_invr  += qi * qj_f * (int_bit*inv_r2 * inv_r - two_k_rf);
 538 #endif
 539 #if defined EL_EWALD_ANA
 540                                 F_invr  += qi * qj_f * (int_bit*inv_r2*inv_r + pmecorrF(beta2*r2)*beta3);
 541 #elif defined EL_EWALD_TAB
 542                                 F_invr  += qi * qj_f * (int_bit*inv_r2 -
 543 #ifdef USE_TEXOBJ
 544                                                         interpolate_coulomb_force_r(nbparam->coulomb_tab_texobj, r2 * inv_r, coulomb_tab_scale)
 545 #else
 546                                                         interpolate_coulomb_force_r(coulomb_tab_climg2d, r2 * inv_r, coulomb_tab_scale)
 547 #endif /* USE_TEXOBJ */
 548                                                         ) * inv_r;
 549 #endif /* EL_EWALD_ANA/TAB */
 550
 551 #ifdef CALC_ENERGIES
 552 #ifdef EL_CUTOFF
 553                                 E_el    += qi * qj_f * (int_bit*inv_r - c_rf);
 554 #endif
 555 #ifdef EL_RF
 556                                 E_el    += qi * qj_f * (int_bit*inv_r + 0.5f * two_k_rf * r2 - c_rf);
 557 #endif
 558 #ifdef EL_EWALD_ANY
 559                                 /* 1.0f - erff is faster than erfcf */
 560                                 E_el    += qi * qj_f * (inv_r * (int_bit - erf(r2 * inv_r * beta)) - int_bit * ewald_shift);
 561 #endif                          /* EL_EWALD_ANY */
 562 #endif
 563                                 f_ij    = rv * F_invr;
 564
 565                                 /* accumulate j forces in registers */
 566                                 fcj_buf -= f_ij;
 567
 568                                 /* accumulate i forces in registers */
 569                                 fci_buf[ci_offset] += f_ij;
 570                             }
 571                         }
 572
 573                         /* shift the mask bit by 1 */
 574                         mask_ji += mask_ji;
 575                     }
 576
 577                     /* reduce j forces */
 578
 579                     /* store j forces in shmem */
 580                     f_buf[                  tidx] = fcj_buf.x;
 581                     f_buf[    FBUF_STRIDE + tidx] = fcj_buf.y;
 582                     f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
 583
 584                     reduce_force_j_generic(f_buf, f, tidxi, tidxj, aj);
 585                 }
 586             }
 587 #ifdef PRUNE_NBL
 588             /* Update the imask with the new one which does not contain the
 589                out of range clusters anymore. */
 590
 591             pl_cj4[j4].imei[widx].imask = imask;
 592 #endif
 593         }
 594     }
 595
 596     /* skip central shifts when summing shift forces */
 597     if (nb_sci.shift == CENTRAL)
 598     {
 599         bCalcFshift = false;
 600     }
 601
 602     fshift_buf = 0.0f;
 603
 604     /* reduce i forces */
 605     for (ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
 606     {
 607         ai  = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
 608
 609         f_buf[                  tidx] = fci_buf[ci_offset].x;
 610         f_buf[    FBUF_STRIDE + tidx] = fci_buf[ci_offset].y;
 611         f_buf[2 * FBUF_STRIDE + tidx] = fci_buf[ci_offset].z;
 612         barrier(CLK_LOCAL_MEM_FENCE);
 613         reduce_force_i(f_buf, f,
 614                        &fshift_buf, bCalcFshift,
 615                        tidxi, tidxj, ai);
 616         barrier(CLK_LOCAL_MEM_FENCE);
 617     }
 618
 619     /* add up local shift forces into global mem */
 620     if (bCalcFshift)
 621     {
 622         /* Only threads with tidxj < 3 will update fshift.
 623            The threads performing the update must be the same with the threads
 624            which stored the reduction result in reduce_force_i function
 625         */
 626         if (tidxj < 3)
 627             atomicAdd_g_f(&(fshift[3 * nb_sci.shift + tidxj]), fshift_buf);
 628     }
 629
 630 #ifdef CALC_ENERGIES
 631     /* flush the energies to shmem and reduce them */
 632     f_buf[              tidx] = E_lj;
 633     f_buf[FBUF_STRIDE + tidx] = E_el;
 634     reduce_energy_pow2(f_buf + (tidx & WARP_SIZE), e_lj, e_el, tidx & ~WARP_SIZE);
 635
 636 #endif
 637 }
 638
 639 #undef EL_EWALD_ANY
 640 #undef EXCLUSION_FORCES
 641 #undef LJ_EWALD
 642
 643 #undef LJ_COMB