Split lines with many copyright years
[alexxy/gromacs.git] / src / gromacs / nbnxm / cuda / nbnxm_cuda_kernel.cuh
1 /*
2  * This file is part of the GROMACS molecular simulation package.
3  *
4  * Copyright (c) 2012,2013,2014,2015,2016 by the GROMACS development team.
5  * Copyright (c) 2017,2018,2019,2020, 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
39  *  CUDA non-bonded kernel used through preprocessor-based code generation
40  *  of multiple kernel flavors, see nbnxn_cuda_kernels.cuh.
41  *
42  *  NOTE: No include fence as it is meant to be included multiple times.
43  *
44  *  \author Szilárd Páll <pall.szilard@gmail.com>
45  *  \author Berk Hess <hess@kth.se>
46  *  \ingroup module_nbnxm
47  */
48
49 #include "gromacs/gpu_utils/cuda_arch_utils.cuh"
50 #include "gromacs/gpu_utils/cuda_kernel_utils.cuh"
51 #include "gromacs/math/utilities.h"
52 #include "gromacs/pbcutil/ishift.h"
53 /* Note that floating-point constants in CUDA code should be suffixed
54  * with f (e.g. 0.5f), to stop the compiler producing intermediate
55  * code that is in double precision.
56  */
57
58 #if defined EL_EWALD_ANA || defined EL_EWALD_TAB
59 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
60 #    define EL_EWALD_ANY
61 #endif
62
63 #if defined EL_EWALD_ANY || defined EL_RF || defined LJ_EWALD \
64         || (defined EL_CUTOFF && defined CALC_ENERGIES)
65 /* Macro to control the calculation of exclusion forces in the kernel
66  * We do that with Ewald (elec/vdw) and RF. Cut-off only has exclusion
67  * energy terms.
68  *
69  * Note: convenience macro, needs to be undef-ed at the end of the file.
70  */
71 #    define EXCLUSION_FORCES
72 #endif
73
74 #if defined LJ_EWALD_COMB_GEOM || defined LJ_EWALD_COMB_LB
75 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
76 #    define LJ_EWALD
77 #endif
78
79 #if defined LJ_COMB_GEOM || defined LJ_COMB_LB
80 #    define LJ_COMB
81 #endif
82
83 /*
84    Kernel launch parameters:
85     - #blocks   = #pair lists, blockId = pair list Id
86     - #threads  = NTHREAD_Z * c_clSize^2
87     - shmem     = see nbnxn_cuda.cu:calc_shmem_required_nonbonded()
88
89     Each thread calculates an i force-component taking one pair of i-j atoms.
90  */
91
92 /**@{*/
93 /*! \brief Compute capability dependent definition of kernel launch configuration parameters.
94  *
95  * NTHREAD_Z controls the number of j-clusters processed concurrently on NTHREAD_Z
96  * warp-pairs per block.
97  *
98  * - On CC 3.0-3.5, and >=5.0 NTHREAD_Z == 1, translating to 64 th/block with 16
99  * blocks/multiproc, is the fastest even though this setup gives low occupancy
100  * (except on 6.0).
101  * NTHREAD_Z > 1 results in excessive register spilling unless the minimum blocks
102  * per multiprocessor is reduced proportionally to get the original number of max
103  * threads in flight (and slightly lower performance).
104  * - On CC 3.7 there are enough registers to double the number of threads; using
105  * NTHREADS_Z == 2 is fastest with 16 blocks (TODO: test with RF and other kernels
106  * with low-register use).
107  *
108  * Note that the current kernel implementation only supports NTHREAD_Z > 1 with
109  * shuffle-based reduction, hence CC >= 3.0.
110  *
111  *
112  * NOTEs on Volta / CUDA 9 extensions:
113  *
114  * - While active thread masks are required for the warp collectives
115  *   (we use any and shfl), the kernel is designed such that all conditions
116  *   (other than the inner-most distance check) including loop trip counts
117  *   are warp-synchronous. Therefore, we don't need ballot to compute the
118  *   active masks as these are all full-warp masks.
119  *
120  * - TODO: reconsider the use of __syncwarp(): its only role is currently to prevent
121  *   WAR hazard due to the cj preload; we should try to replace it with direct
122  *   loads (which may be faster given the improved L1 on Volta).
123  */
124
125 /* Kernel launch bounds for different compute capabilities. The value of NTHREAD_Z
126  * determines the number of threads per block and it is chosen such that
127  * 16 blocks/multiprocessor can be kept in flight.
128  * - CC 3.0,3.5, and >=5.0: NTHREAD_Z=1, (64, 16) bounds
129  * - CC 3.7:                NTHREAD_Z=2, (128, 16) bounds
130  *
131  * Note: convenience macros, need to be undef-ed at the end of the file.
132  */
133 #if GMX_PTX_ARCH == 370
134 #    define NTHREAD_Z (2)
135 #    define MIN_BLOCKS_PER_MP (16)
136 #else
137 #    define NTHREAD_Z (1)
138 #    define MIN_BLOCKS_PER_MP (16)
139 #endif /* GMX_PTX_ARCH == 370 */
140 #define THREADS_PER_BLOCK (c_clSize * c_clSize * NTHREAD_Z)
141
142 #if GMX_PTX_ARCH >= 350
143 /**@}*/
144 __launch_bounds__(THREADS_PER_BLOCK, MIN_BLOCKS_PER_MP)
145 #else
146 __launch_bounds__(THREADS_PER_BLOCK)
147 #endif /* GMX_PTX_ARCH >= 350 */
148 #ifdef PRUNE_NBL
149 #    ifdef CALC_ENERGIES
150         __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_prune_cuda)
151 #    else
152         __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_prune_cuda)
153 #    endif /* CALC_ENERGIES */
154 #else
155 #    ifdef CALC_ENERGIES
156         __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_cuda)
157 #    else
158         __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_cuda)
159 #    endif /* CALC_ENERGIES */
160 #endif     /* PRUNE_NBL */
161                 (const cu_atomdata_t atdat, const cu_nbparam_t nbparam, const cu_plist_t plist, bool bCalcFshift)
162 #ifdef FUNCTION_DECLARATION_ONLY
163                         ; /* Only do function declaration, omit the function body. */
164 #else
165 {
166     /* convenience variables */
167     const nbnxn_sci_t* pl_sci = plist.sci;
168 #    ifndef PRUNE_NBL
169     const
170 #    endif
171             nbnxn_cj4_t* pl_cj4      = plist.cj4;
172     const nbnxn_excl_t*  excl        = plist.excl;
173 #    ifndef LJ_COMB
174     const int*           atom_types  = atdat.atom_types;
175     int                  ntypes      = atdat.ntypes;
176 #    else
177     const float2* lj_comb = atdat.lj_comb;
178     float2        ljcp_i, ljcp_j;
179 #    endif
180     const float4*        xq          = atdat.xq;
181     float3*              f           = atdat.f;
182     const float3*        shift_vec   = atdat.shift_vec;
183     float                rcoulomb_sq = nbparam.rcoulomb_sq;
184 #    ifdef VDW_CUTOFF_CHECK
185     float                rvdw_sq     = nbparam.rvdw_sq;
186     float                vdw_in_range;
187 #    endif
188 #    ifdef LJ_EWALD
189     float                lje_coeff2, lje_coeff6_6;
190 #    endif
191 #    ifdef EL_RF
192     float                two_k_rf    = nbparam.two_k_rf;
193 #    endif
194 #    ifdef EL_EWALD_ANA
195     float                beta2       = nbparam.ewald_beta * nbparam.ewald_beta;
196     float                beta3       = nbparam.ewald_beta * nbparam.ewald_beta * nbparam.ewald_beta;
197 #    endif
198 #    ifdef PRUNE_NBL
199     float                rlist_sq    = nbparam.rlistOuter_sq;
200 #    endif
201
202 #    ifdef CALC_ENERGIES
203 #        ifdef EL_EWALD_ANY
204     float                beta        = nbparam.ewald_beta;
205     float                ewald_shift = nbparam.sh_ewald;
206 #        else
207     float c_rf = nbparam.c_rf;
208 #        endif /* EL_EWALD_ANY */
209     float*               e_lj        = atdat.e_lj;
210     float*               e_el        = atdat.e_el;
211 #    endif     /* CALC_ENERGIES */
212
213     /* thread/block/warp id-s */
214     unsigned int tidxi = threadIdx.x;
215     unsigned int tidxj = threadIdx.y;
216     unsigned int tidx  = threadIdx.y * blockDim.x + threadIdx.x;
217 #    if NTHREAD_Z == 1
218     unsigned int tidxz = 0;
219 #    else
220     unsigned int  tidxz = threadIdx.z;
221 #    endif
222     unsigned int bidx  = blockIdx.x;
223     unsigned int widx  = tidx / warp_size; /* warp index */
224
225     int          sci, ci, cj, ai, aj, cij4_start, cij4_end;
226 #    ifndef LJ_COMB
227     int          typei, typej;
228 #    endif
229     int          i, jm, j4, wexcl_idx;
230     float        qi, qj_f, r2, inv_r, inv_r2;
231 #    if !defined LJ_COMB_LB || defined CALC_ENERGIES
232     float        inv_r6, c6, c12;
233 #    endif
234 #    ifdef LJ_COMB_LB
235     float        sigma, epsilon;
236 #    endif
237     float        int_bit, F_invr;
238 #    ifdef CALC_ENERGIES
239     float        E_lj, E_el;
240 #    endif
241 #    if defined CALC_ENERGIES || defined LJ_POT_SWITCH
242     float        E_lj_p;
243 #    endif
244     unsigned int wexcl, imask, mask_ji;
245     float4       xqbuf;
246     float3       xi, xj, rv, f_ij, fcj_buf;
247     float3       fci_buf[c_numClPerSupercl]; /* i force buffer */
248     nbnxn_sci_t  nb_sci;
249
250     /*! i-cluster interaction mask for a super-cluster with all c_numClPerSupercl=8 bits set */
251     const unsigned superClInteractionMask = ((1U << c_numClPerSupercl) - 1U);
252
253     /*********************************************************************
254      * Set up shared memory pointers.
255      * sm_nextSlotPtr should always be updated to point to the "next slot",
256      * that is past the last point where data has been stored.
257      */
258     extern __shared__ char sm_dynamicShmem[];
259     char*                  sm_nextSlotPtr = sm_dynamicShmem;
260     static_assert(sizeof(char) == 1,
261                   "The shared memory offset calculation assumes that char is 1 byte");
262
263     /* shmem buffer for i x+q pre-loading */
264     float4* xqib = (float4*)sm_nextSlotPtr;
265     sm_nextSlotPtr += (c_numClPerSupercl * c_clSize * sizeof(*xqib));
266
267     /* shmem buffer for cj, for each warp separately */
268     int* cjs = (int*)(sm_nextSlotPtr);
269     /* the cjs buffer's use expects a base pointer offset for pairs of warps in the j-concurrent execution */
270     cjs += tidxz * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize;
271     sm_nextSlotPtr += (NTHREAD_Z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(*cjs));
272
273 #    ifndef LJ_COMB
274     /* shmem buffer for i atom-type pre-loading */
275     int* atib = (int*)sm_nextSlotPtr;
276     sm_nextSlotPtr += (c_numClPerSupercl * c_clSize * sizeof(*atib));
277 #    else
278     /* shmem buffer for i-atom LJ combination rule parameters */
279     float2* ljcpib = (float2*)sm_nextSlotPtr;
280     sm_nextSlotPtr += (c_numClPerSupercl * c_clSize * sizeof(*ljcpib));
281 #    endif
282     /*********************************************************************/
283
284     nb_sci     = pl_sci[bidx];         /* my i super-cluster's index = current bidx */
285     sci        = nb_sci.sci;           /* super-cluster */
286     cij4_start = nb_sci.cj4_ind_start; /* first ...*/
287     cij4_end   = nb_sci.cj4_ind_end;   /* and last index of j clusters */
288
289     if (tidxz == 0)
290     {
291         /* Pre-load i-atom x and q into shared memory */
292         ci = sci * c_numClPerSupercl + tidxj;
293         ai = ci * c_clSize + tidxi;
294
295         float* shiftptr = (float*)&shift_vec[nb_sci.shift];
296         xqbuf = xq[ai] + make_float4(LDG(shiftptr), LDG(shiftptr + 1), LDG(shiftptr + 2), 0.0f);
297         xqbuf.w *= nbparam.epsfac;
298         xqib[tidxj * c_clSize + tidxi] = xqbuf;
299
300 #    ifndef LJ_COMB
301         /* Pre-load the i-atom types into shared memory */
302         atib[tidxj * c_clSize + tidxi] = atom_types[ai];
303 #    else
304         /* Pre-load the LJ combination parameters into shared memory */
305         ljcpib[tidxj * c_clSize + tidxi] = lj_comb[ai];
306 #    endif
307     }
308     __syncthreads();
309
310     for (i = 0; i < c_numClPerSupercl; i++)
311     {
312         fci_buf[i] = make_float3(0.0f);
313     }
314
315 #    ifdef LJ_EWALD
316     /* TODO: we are trading registers with flops by keeping lje_coeff-s, try re-calculating it later */
317     lje_coeff2   = nbparam.ewaldcoeff_lj * nbparam.ewaldcoeff_lj;
318     lje_coeff6_6 = lje_coeff2 * lje_coeff2 * lje_coeff2 * c_oneSixth;
319 #    endif
320
321
322 #    ifdef CALC_ENERGIES
323     E_lj         = 0.0f;
324     E_el         = 0.0f;
325
326 #        ifdef EXCLUSION_FORCES /* Ewald or RF */
327     if (nb_sci.shift == CENTRAL && pl_cj4[cij4_start].cj[0] == sci * c_numClPerSupercl)
328     {
329         /* we have the diagonal: add the charge and LJ self interaction energy term */
330         for (i = 0; i < c_numClPerSupercl; i++)
331         {
332 #            if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
333             qi = xqib[i * c_clSize + tidxi].w;
334             E_el += qi * qi;
335 #            endif
336
337 #            ifdef LJ_EWALD
338 #                if DISABLE_CUDA_TEXTURES
339             E_lj += LDG(
340                     &nbparam.nbfp[atom_types[(sci * c_numClPerSupercl + i) * c_clSize + tidxi] * (ntypes + 1) * 2]);
341 #                else
342             E_lj += tex1Dfetch<float>(
343                     nbparam.nbfp_texobj,
344                     atom_types[(sci * c_numClPerSupercl + i) * c_clSize + tidxi] * (ntypes + 1) * 2);
345 #                endif
346 #            endif
347         }
348
349         /* divide the self term(s) equally over the j-threads, then multiply with the coefficients. */
350 #            ifdef LJ_EWALD
351         E_lj /= c_clSize * NTHREAD_Z;
352         E_lj *= 0.5f * c_oneSixth * lje_coeff6_6;
353 #            endif
354
355 #            if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
356         /* Correct for epsfac^2 due to adding qi^2 */
357         E_el /= nbparam.epsfac * c_clSize * NTHREAD_Z;
358 #                if defined EL_RF || defined EL_CUTOFF
359         E_el *= -0.5f * c_rf;
360 #                else
361         E_el *= -beta * M_FLOAT_1_SQRTPI; /* last factor 1/sqrt(pi) */
362 #                endif
363 #            endif /* EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF */
364     }
365 #        endif     /* EXCLUSION_FORCES */
366
367 #    endif /* CALC_ENERGIES */
368
369 #    ifdef EXCLUSION_FORCES
370     const int nonSelfInteraction = !(nb_sci.shift == CENTRAL & tidxj <= tidxi);
371 #    endif
372
373     /* loop over the j clusters = seen by any of the atoms in the current super-cluster;
374      * The loop stride NTHREAD_Z ensures that consecutive warps-pairs are assigned
375      * consecutive j4's entries.
376      */
377     for (j4 = cij4_start + tidxz; j4 < cij4_end; j4 += NTHREAD_Z)
378     {
379         wexcl_idx = pl_cj4[j4].imei[widx].excl_ind;
380         imask     = pl_cj4[j4].imei[widx].imask;
381         wexcl     = excl[wexcl_idx].pair[(tidx) & (warp_size - 1)];
382
383 #    ifndef PRUNE_NBL
384         if (imask)
385 #    endif
386         {
387             /* Pre-load cj into shared memory on both warps separately */
388             if ((tidxj == 0 | tidxj == 4) & (tidxi < c_nbnxnGpuJgroupSize))
389             {
390                 cjs[tidxi + tidxj * c_nbnxnGpuJgroupSize / c_splitClSize] = pl_cj4[j4].cj[tidxi];
391             }
392             __syncwarp(c_fullWarpMask);
393
394             /* Unrolling this loop
395                - with pruning leads to register spilling;
396                - on Kepler and later it is much slower;
397                Tested with up to nvcc 7.5 */
398             for (jm = 0; jm < c_nbnxnGpuJgroupSize; jm++)
399             {
400                 if (imask & (superClInteractionMask << (jm * c_numClPerSupercl)))
401                 {
402                     mask_ji = (1U << (jm * c_numClPerSupercl));
403
404                     cj = cjs[jm + (tidxj & 4) * c_nbnxnGpuJgroupSize / c_splitClSize];
405                     aj = cj * c_clSize + tidxj;
406
407                     /* load j atom data */
408                     xqbuf = xq[aj];
409                     xj    = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
410                     qj_f  = xqbuf.w;
411 #    ifndef LJ_COMB
412                     typej = atom_types[aj];
413 #    else
414                     ljcp_j = lj_comb[aj];
415 #    endif
416
417                     fcj_buf = make_float3(0.0f);
418
419 #    if !defined PRUNE_NBL
420 #        pragma unroll 8
421 #    endif
422                     for (i = 0; i < c_numClPerSupercl; i++)
423                     {
424                         if (imask & mask_ji)
425                         {
426                             ci = sci * c_numClPerSupercl + i; /* i cluster index */
427
428                             /* all threads load an atom from i cluster ci into shmem! */
429                             xqbuf = xqib[i * c_clSize + tidxi];
430                             xi    = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
431
432                             /* distance between i and j atoms */
433                             rv = xi - xj;
434                             r2 = norm2(rv);
435
436 #    ifdef PRUNE_NBL
437                             /* If _none_ of the atoms pairs are in cutoff range,
438                                the bit corresponding to the current
439                                cluster-pair in imask gets set to 0. */
440                             if (!__any_sync(c_fullWarpMask, r2 < rlist_sq))
441                             {
442                                 imask &= ~mask_ji;
443                             }
444 #    endif
445
446                             int_bit = (wexcl & mask_ji) ? 1.0f : 0.0f;
447
448                             /* cutoff & exclusion check */
449 #    ifdef EXCLUSION_FORCES
450                             if ((r2 < rcoulomb_sq) * (nonSelfInteraction | (ci != cj)))
451 #    else
452                             if ((r2 < rcoulomb_sq) * int_bit)
453 #    endif
454                             {
455                                 /* load the rest of the i-atom parameters */
456                                 qi = xqbuf.w;
457
458 #    ifndef LJ_COMB
459                                 /* LJ 6*C6 and 12*C12 */
460                                 typei = atib[i * c_clSize + tidxi];
461                                 fetch_nbfp_c6_c12(c6, c12, nbparam, ntypes * typei + typej);
462 #    else
463                                 ljcp_i       = ljcpib[i * c_clSize + tidxi];
464 #        ifdef LJ_COMB_GEOM
465                                 c6           = ljcp_i.x * ljcp_j.x;
466                                 c12          = ljcp_i.y * ljcp_j.y;
467 #        else
468                                 /* LJ 2^(1/6)*sigma and 12*epsilon */
469                                 sigma   = ljcp_i.x + ljcp_j.x;
470                                 epsilon = ljcp_i.y * ljcp_j.y;
471 #            if defined CALC_ENERGIES || defined LJ_FORCE_SWITCH || defined LJ_POT_SWITCH
472                                 convert_sigma_epsilon_to_c6_c12(sigma, epsilon, &c6, &c12);
473 #            endif
474 #        endif /* LJ_COMB_GEOM */
475 #    endif     /* LJ_COMB */
476
477                                 // Ensure distance do not become so small that r^-12 overflows
478                                 r2 = max(r2, NBNXN_MIN_RSQ);
479
480                                 inv_r  = rsqrt(r2);
481                                 inv_r2 = inv_r * inv_r;
482 #    if !defined LJ_COMB_LB || defined CALC_ENERGIES
483                                 inv_r6 = inv_r2 * inv_r2 * inv_r2;
484 #        ifdef EXCLUSION_FORCES
485                                 /* We could mask inv_r2, but with Ewald
486                                  * masking both inv_r6 and F_invr is faster */
487                                 inv_r6 *= int_bit;
488 #        endif /* EXCLUSION_FORCES */
489
490                                 F_invr = inv_r6 * (c12 * inv_r6 - c6) * inv_r2;
491 #        if defined CALC_ENERGIES || defined LJ_POT_SWITCH
492                                 E_lj_p = int_bit
493                                          * (c12 * (inv_r6 * inv_r6 + nbparam.repulsion_shift.cpot) * c_oneTwelveth
494                                             - c6 * (inv_r6 + nbparam.dispersion_shift.cpot) * c_oneSixth);
495 #        endif
496 #    else /* !LJ_COMB_LB || CALC_ENERGIES */
497                                 float sig_r  = sigma * inv_r;
498                                 float sig_r2 = sig_r * sig_r;
499                                 float sig_r6 = sig_r2 * sig_r2 * sig_r2;
500 #        ifdef EXCLUSION_FORCES
501                                 sig_r6 *= int_bit;
502 #        endif /* EXCLUSION_FORCES */
503
504                                 F_invr = epsilon * sig_r6 * (sig_r6 - 1.0f) * inv_r2;
505 #    endif     /* !LJ_COMB_LB || CALC_ENERGIES */
506
507 #    ifdef LJ_FORCE_SWITCH
508 #        ifdef CALC_ENERGIES
509                                 calculate_force_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
510 #        else
511                                 calculate_force_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr);
512 #        endif /* CALC_ENERGIES */
513 #    endif     /* LJ_FORCE_SWITCH */
514
515
516 #    ifdef LJ_EWALD
517 #        ifdef LJ_EWALD_COMB_GEOM
518 #            ifdef CALC_ENERGIES
519                                 calculate_lj_ewald_comb_geom_F_E(nbparam, typei, typej, r2, inv_r2,
520                                                                  lje_coeff2, lje_coeff6_6, int_bit,
521                                                                  &F_invr, &E_lj_p);
522 #            else
523                                 calculate_lj_ewald_comb_geom_F(nbparam, typei, typej, r2, inv_r2,
524                                                                lje_coeff2, lje_coeff6_6, &F_invr);
525 #            endif /* CALC_ENERGIES */
526 #        elif defined LJ_EWALD_COMB_LB
527                                 calculate_lj_ewald_comb_LB_F_E(nbparam, typei, typej, r2, inv_r2,
528                                                                lje_coeff2, lje_coeff6_6,
529 #            ifdef CALC_ENERGIES
530                                                                int_bit, &F_invr, &E_lj_p
531 #            else
532                                                                0, &F_invr, nullptr
533 #            endif /* CALC_ENERGIES */
534                                 );
535 #        endif     /* LJ_EWALD_COMB_GEOM */
536 #    endif         /* LJ_EWALD */
537
538 #    ifdef LJ_POT_SWITCH
539 #        ifdef CALC_ENERGIES
540                                 calculate_potential_switch_F_E(nbparam, inv_r, r2, &F_invr, &E_lj_p);
541 #        else
542                                 calculate_potential_switch_F(nbparam, inv_r, r2, &F_invr, &E_lj_p);
543 #        endif /* CALC_ENERGIES */
544 #    endif     /* LJ_POT_SWITCH */
545
546 #    ifdef VDW_CUTOFF_CHECK
547                                 /* Separate VDW cut-off check to enable twin-range cut-offs
548                                  * (rvdw < rcoulomb <= rlist)
549                                  */
550                                 vdw_in_range = (r2 < rvdw_sq) ? 1.0f : 0.0f;
551                                 F_invr *= vdw_in_range;
552 #        ifdef CALC_ENERGIES
553                                 E_lj_p *= vdw_in_range;
554 #        endif
555 #    endif /* VDW_CUTOFF_CHECK */
556
557 #    ifdef CALC_ENERGIES
558                                 E_lj += E_lj_p;
559 #    endif
560
561
562 #    ifdef EL_CUTOFF
563 #        ifdef EXCLUSION_FORCES
564                                 F_invr += qi * qj_f * int_bit * inv_r2 * inv_r;
565 #        else
566                                 F_invr += qi * qj_f * inv_r2 * inv_r;
567 #        endif
568 #    endif
569 #    ifdef EL_RF
570                                 F_invr += qi * qj_f * (int_bit * inv_r2 * inv_r - two_k_rf);
571 #    endif
572 #    if defined   EL_EWALD_ANA
573                                 F_invr += qi * qj_f
574                                           * (int_bit * inv_r2 * inv_r + pmecorrF(beta2 * r2) * beta3);
575 #    elif defined EL_EWALD_TAB
576                                 F_invr += qi * qj_f
577                                           * (int_bit * inv_r2
578                                              - interpolate_coulomb_force_r(nbparam, r2 * inv_r))
579                                           * inv_r;
580 #    endif /* EL_EWALD_ANA/TAB */
581
582 #    ifdef CALC_ENERGIES
583 #        ifdef EL_CUTOFF
584                                 E_el += qi * qj_f * (int_bit * inv_r - c_rf);
585 #        endif
586 #        ifdef EL_RF
587                                 E_el += qi * qj_f * (int_bit * inv_r + 0.5f * two_k_rf * r2 - c_rf);
588 #        endif
589 #        ifdef EL_EWALD_ANY
590                                 /* 1.0f - erff is faster than erfcf */
591                                 E_el += qi * qj_f
592                                         * (inv_r * (int_bit - erff(r2 * inv_r * beta)) - int_bit * ewald_shift);
593 #        endif /* EL_EWALD_ANY */
594 #    endif
595                                 f_ij = rv * F_invr;
596
597                                 /* accumulate j forces in registers */
598                                 fcj_buf -= f_ij;
599
600                                 /* accumulate i forces in registers */
601                                 fci_buf[i] += f_ij;
602                             }
603                         }
604
605                         /* shift the mask bit by 1 */
606                         mask_ji += mask_ji;
607                     }
608
609                     /* reduce j forces */
610                     reduce_force_j_warp_shfl(fcj_buf, f, tidxi, aj, c_fullWarpMask);
611                 }
612             }
613 #    ifdef PRUNE_NBL
614             /* Update the imask with the new one which does not contain the
615                out of range clusters anymore. */
616             pl_cj4[j4].imei[widx].imask = imask;
617 #    endif
618         }
619         // avoid shared memory WAR hazards between loop iterations
620         __syncwarp(c_fullWarpMask);
621     }
622
623     /* skip central shifts when summing shift forces */
624     if (nb_sci.shift == CENTRAL)
625     {
626         bCalcFshift = false;
627     }
628
629     float fshift_buf = 0.0f;
630
631     /* reduce i forces */
632     for (i = 0; i < c_numClPerSupercl; i++)
633     {
634         ai = (sci * c_numClPerSupercl + i) * c_clSize + tidxi;
635         reduce_force_i_warp_shfl(fci_buf[i], f, &fshift_buf, bCalcFshift, tidxj, ai, c_fullWarpMask);
636     }
637
638     /* add up local shift forces into global mem, tidxj indexes x,y,z */
639     if (bCalcFshift && (tidxj & 3) < 3)
640     {
641         atomicAdd(&(atdat.fshift[nb_sci.shift].x) + (tidxj & 3), fshift_buf);
642     }
643
644 #    ifdef CALC_ENERGIES
645     /* reduce the energies over warps and store into global memory */
646     reduce_energy_warp_shfl(E_lj, E_el, e_lj, e_el, tidx, c_fullWarpMask);
647 #    endif
648 }
649 #endif /* FUNCTION_DECLARATION_ONLY */
650
651 #undef NTHREAD_Z
652 #undef MIN_BLOCKS_PER_MP
653 #undef THREADS_PER_BLOCK
654
655 #undef EL_EWALD_ANY
656 #undef EXCLUSION_FORCES
657 #undef LJ_EWALD
658
659 #undef LJ_COMB