| blob | 0828a943d66c0a64d19ecf4a960fc5c8bb01668f |
1 /*
2 * This file is part of the GROMACS molecular simulation package.
3 *
4 * Copyright (c) 2012,2013,2014,2016,2017,2018, 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 */
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 */
50 #include "nbnxn_ocl_kernel_utils.clh"
52 /////////////////////////////////////////////////////////////////////////////////////////////////
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
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
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
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
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)
85 Each thread calculates an i force-component taking one pair of i-j atoms.
87 TODO: implement 128 threads/wavefront by porting over the NTHREAD_Z/j4 loop
88 "horizontal splitting" over threads.
89 */
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 )
137 {
138 /* convenience variables */
139 cl_nbparam_params_t *nbparam = &nbparam_params;
141 float rcoulomb_sq = nbparam->rcoulomb_sq;
142 #ifdef LJ_COMB
143 float2 ljcp_i, ljcp_j;
144 #endif
145 #ifdef VDW_CUTOFF_CHECK
146 float rvdw_sq = nbparam_params.rvdw_sq;
147 float vdw_in_range;
148 #endif
149 #ifdef LJ_EWALD
150 float lje_coeff2, lje_coeff6_6;
151 #endif
152 #ifdef EL_RF
153 float two_k_rf = nbparam->two_k_rf;
154 #endif
155 #ifdef EL_EWALD_TAB
156 float coulomb_tab_scale = nbparam->coulomb_tab_scale;
157 #endif
158 #ifdef EL_EWALD_ANA
159 float beta2 = nbparam->ewald_beta*nbparam->ewald_beta;
160 float beta3 = nbparam->ewald_beta*nbparam->ewald_beta*nbparam->ewald_beta;
161 #endif
162 #ifdef PRUNE_NBL
163 float rlist_sq = nbparam->rlistOuter_sq;
164 #endif
166 #ifdef CALC_ENERGIES
167 #ifdef EL_EWALD_ANY
168 float beta = nbparam->ewald_beta;
169 float ewald_shift = nbparam->sh_ewald;
170 #else
171 float c_rf = nbparam->c_rf;
172 #endif /* EL_EWALD_ANY */
173 #endif /* CALC_ENERGIES */
175 /* thread/block/warp id-s */
176 unsigned int tidxi = get_local_id(0);
177 unsigned int tidxj = get_local_id(1);
178 unsigned int tidx = get_local_id(1) * get_local_size(0) + get_local_id(0);
179 unsigned int bidx = get_group_id(0);
180 unsigned int widx = tidx / WARP_SIZE; /* warp index */
181 int sci, ci, cj, ci_offset,
182 ai, aj,
183 cij4_start, cij4_end;
184 #ifndef LJ_COMB
185 int typei, typej;
186 #endif
187 int i, jm, j4, wexcl_idx;
188 float qi, qj_f,
189 r2, inv_r, inv_r2;
190 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
191 float inv_r6, c6, c12;
192 #endif
193 #if defined LJ_COMB_LB
194 float sigma, epsilon;
195 #endif
196 float int_bit,
197 F_invr;
199 #ifdef CALC_ENERGIES
200 float E_lj, E_el;
201 #endif
202 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
203 float E_lj_p;
204 #endif
205 unsigned int wexcl, imask, mask_ji;
206 float4 xqbuf;
207 float3 xi, xj, rv, f_ij, fcj_buf/*, fshift_buf*/;
208 float fshift_buf;
209 float3 fci_buf[NCL_PER_SUPERCL]; /* i force buffer */
210 nbnxn_sci_t nb_sci;
212 /*! i-cluster interaction mask for a super-cluster with all NCL_PER_SUPERCL=8 bits set */
213 const unsigned superClInteractionMask = ((1U << NCL_PER_SUPERCL) - 1U);
215 /* shmem buffer for cj, for both warps separately */
216 __local int *cjs = (__local int *)(xqib + NCL_PER_SUPERCL * CL_SIZE);
217 #define LOCAL_OFFSET cjs + 2 * NBNXN_GPU_JGROUP_SIZE
219 #ifdef IATYPE_SHMEM
220 #ifndef LJ_COMB
221 /* shmem buffer for i atom-type pre-loading */
222 __local int *atib = (__local int *)(LOCAL_OFFSET);
223 #undef LOCAL_OFFSET
224 #define LOCAL_OFFSET atib + NCL_PER_SUPERCL * CL_SIZE
225 #else
226 __local float2 *ljcpib = (__local float2 *)(LOCAL_OFFSET);
227 #undef LOCAL_OFFSET
228 #define LOCAL_OFFSET ljcpib + NCL_PER_SUPERCL * CL_SIZE
229 #endif
230 #endif
232 #ifndef REDUCE_SHUFFLE
233 /* shmem j force buffer */
234 __local float *f_buf = (__local float *)(LOCAL_OFFSET);
235 #undef LOCAL_OFFSET
236 #define LOCAL_OFFSET f_buf + CL_SIZE * CL_SIZE * 3
237 #endif
238 /* Local buffer used to implement __any warp vote function from CUDA.
239 volatile is used to avoid compiler optimizations for AMD builds. */
240 volatile __local uint *warp_any = (__local uint*)(LOCAL_OFFSET);
241 #undef LOCAL_OFFSET
243 nb_sci = pl_sci[bidx]; /* my i super-cluster's index = current bidx */
244 sci = nb_sci.sci; /* super-cluster */
245 cij4_start = nb_sci.cj4_ind_start; /* first ...*/
246 cij4_end = nb_sci.cj4_ind_end; /* and last index of j clusters */
248 /* Pre-load i-atom x and q into shared memory */
249 ci = sci * NCL_PER_SUPERCL + tidxj;
250 ai = ci * CL_SIZE + tidxi;
252 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);
253 xqbuf.w *= nbparam->epsfac;
254 xqib[tidxj * CL_SIZE + tidxi] = xqbuf;
256 #ifdef IATYPE_SHMEM
257 #ifndef LJ_COMB
258 /* Pre-load the i-atom types into shared memory */
259 atib[tidxj * CL_SIZE + tidxi] = atom_types[ai];
260 #else
261 ljcpib[tidxj * CL_SIZE + tidxi] = lj_comb[ai];
262 #endif
263 #endif
264 /* Initialise warp vote. (8x8 block) 2 warps for nvidia */
265 if(tidx==0 || tidx==32)
266 warp_any[widx] = 0;
268 barrier(CLK_LOCAL_MEM_FENCE);
270 for (ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
271 {
272 fci_buf[ci_offset] = (float3)(0.0f);
273 }
275 #ifdef LJ_EWALD
276 /* TODO: we are trading registers with flops by keeping lje_coeff-s, try re-calculating it later */
277 lje_coeff2 = nbparam->ewaldcoeff_lj*nbparam->ewaldcoeff_lj;
278 lje_coeff6_6 = lje_coeff2*lje_coeff2*lje_coeff2*ONE_SIXTH_F;
279 #endif /* LJ_EWALD */
282 #ifdef CALC_ENERGIES
283 E_lj = 0.0f;
284 E_el = 0.0f;
286 #if defined EXCLUSION_FORCES /* Ewald or RF */
287 if (nb_sci.shift == CENTRAL && pl_cj4[cij4_start].cj[0] == sci*NCL_PER_SUPERCL)
288 {
289 /* we have the diagonal: add the charge and LJ self interaction energy term */
290 for (i = 0; i < NCL_PER_SUPERCL; i++)
291 {
292 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
293 qi = xqib[i * CL_SIZE + tidxi].w;
294 E_el += qi*qi;
295 #endif
296 #if defined LJ_EWALD
297 E_lj += nbfp_climg2d[atom_types[(sci*NCL_PER_SUPERCL + i)*CL_SIZE + tidxi]*(ntypes + 1)*2];
298 #endif /* LJ_EWALD */
299 }
301 /* divide the self term(s) equally over the j-threads, then multiply with the coefficients. */
302 #ifdef LJ_EWALD
303 E_lj /= CL_SIZE;
304 E_lj *= 0.5f*ONE_SIXTH_F*lje_coeff6_6;
305 #endif /* LJ_EWALD */
307 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
308 /* Correct for epsfac^2 due to adding qi^2 */
309 E_el /= nbparam->epsfac*CL_SIZE;
310 #if defined EL_RF || defined EL_CUTOFF
311 E_el *= -0.5f*c_rf;
312 #else
313 E_el *= -beta*M_FLOAT_1_SQRTPI; /* last factor 1/sqrt(pi) */
314 #endif
315 #endif /* EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF */
316 }
317 #endif /* EXCLUSION_FORCES */
319 #endif /* CALC_ENERGIES */
321 #ifdef EXCLUSION_FORCES
322 const int nonSelfInteraction = !(nb_sci.shift == CENTRAL & tidxj <= tidxi);
323 #endif
325 /* loop over the j clusters = seen by any of the atoms in the current super-cluster */
326 for (j4 = cij4_start; j4 < cij4_end; j4++)
327 {
328 wexcl_idx = pl_cj4[j4].imei[widx].excl_ind;
329 imask = pl_cj4[j4].imei[widx].imask;
330 wexcl = excl[wexcl_idx].pair[(tidx) & (WARP_SIZE - 1)];
332 #ifndef PRUNE_NBL
333 if (imask)
334 #endif
335 {
336 /* Pre-load cj into shared memory on both warps separately */
337 if ((tidxj == 0 | tidxj == 4) & (tidxi < NBNXN_GPU_JGROUP_SIZE))
338 {
339 cjs[tidxi + tidxj * NBNXN_GPU_JGROUP_SIZE / 4] = pl_cj4[j4].cj[tidxi];
340 }
342 /* Unrolling this loop improves performance without pruning but
343 * with pruning it leads to slowdown.
344 *
345 * Tested with driver 1800.5
346 */
347 #if !defined PRUNE_NBL
348 #pragma unroll 4
349 #endif
351 for (jm = 0; jm < NBNXN_GPU_JGROUP_SIZE; jm++)
352 {
353 if (imask & (superClInteractionMask << (jm * NCL_PER_SUPERCL)))
354 {
355 mask_ji = (1U << (jm * NCL_PER_SUPERCL));
357 cj = cjs[jm + (tidxj & 4) * NBNXN_GPU_JGROUP_SIZE / 4];
358 aj = cj * CL_SIZE + tidxj;
360 /* load j atom data */
361 xqbuf = xq[aj];
362 xj = (float3)(xqbuf.xyz);
363 qj_f = xqbuf.w;
364 #ifndef LJ_COMB
365 typej = atom_types[aj];
366 #else
367 ljcp_j = lj_comb[aj];
368 #endif
370 fcj_buf = (float3)(0.0f);
372 #if !defined PRUNE_NBL
373 #pragma unroll 8
374 #endif
375 for (i = 0; i < NCL_PER_SUPERCL; i++)
376 {
377 if (imask & mask_ji)
378 {
379 ci_offset = i; /* i force buffer offset */
381 ci = sci * NCL_PER_SUPERCL + i; /* i cluster index */
382 ai = ci * CL_SIZE + tidxi; /* i atom index */
384 /* all threads load an atom from i cluster ci into shmem! */
385 xqbuf = xqib[i * CL_SIZE + tidxi];
386 xi = (float3)(xqbuf.xyz);
388 /* distance between i and j atoms */
389 rv = xi - xj;
390 r2 = norm2(rv);
392 #ifdef PRUNE_NBL
393 /* vote.. should code shmem serialisation, wonder what the hit will be */
394 if (r2 < rlist_sq)
395 warp_any[widx]=1;
397 /* If _none_ of the atoms pairs are in cutoff range,
398 the bit corresponding to the current
399 cluster-pair in imask gets set to 0. */
400 if (!warp_any[widx])
401 imask &= ~mask_ji;
403 warp_any[widx]=0;
404 #endif
406 int_bit = (wexcl & mask_ji) ? 1.0f : 0.0f;
408 /* cutoff & exclusion check */
409 #ifdef EXCLUSION_FORCES
410 if ((r2 < rcoulomb_sq) * (nonSelfInteraction | (ci != cj)))
411 #else
412 if ((r2 < rcoulomb_sq) * int_bit)
413 #endif
414 {
415 /* load the rest of the i-atom parameters */
416 qi = xqbuf.w;
417 #ifdef IATYPE_SHMEM
418 #ifndef LJ_COMB
419 typei = atib[i * CL_SIZE + tidxi];
420 #else
421 ljcp_i = ljcpib[i * CL_SIZE + tidxi];
422 #endif
423 #else /* IATYPE_SHMEM */
424 #ifndef LJ_COMB
425 typei = atom_types[ai];
426 #else
427 ljcp_i = lj_comb[ai];
428 #endif
429 #endif /* IATYPE_SHMEM */
430 /* LJ 6*C6 and 12*C12 */
431 #ifndef LJ_COMB
432 c6 = nbfp_climg2d[2 * (ntypes * typei + typej)];
433 c12 = nbfp_climg2d[2 * (ntypes * typei + typej)+1];
434 #else /* LJ_COMB */
435 #ifdef LJ_COMB_GEOM
436 c6 = ljcp_i.x * ljcp_j.x;
437 c12 = ljcp_i.y * ljcp_j.y;
438 #else
439 /* LJ 2^(1/6)*sigma and 12*epsilon */
440 sigma = ljcp_i.x + ljcp_j.x;
441 epsilon = ljcp_i.y * ljcp_j.y;
442 #if defined CALC_ENERGIES || defined LJ_FORCE_SWITCH || defined LJ_POT_SWITCH
443 convert_sigma_epsilon_to_c6_c12(sigma, epsilon, &c6, &c12);
444 #endif
445 #endif /* LJ_COMB_GEOM */
446 #endif /* LJ_COMB */
448 // Ensure distance do not become so small that r^-12 overflows
449 r2 = max(r2,NBNXN_MIN_RSQ);
451 inv_r = rsqrt(r2);
452 inv_r2 = inv_r * inv_r;
453 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
454 inv_r6 = inv_r2 * inv_r2 * inv_r2;
455 #if defined EXCLUSION_FORCES
456 /* We could mask inv_r2, but with Ewald
457 * masking both inv_r6 and F_invr is faster */
458 inv_r6 *= int_bit;
459 #endif /* EXCLUSION_FORCES */
461 F_invr = inv_r6 * (c12 * inv_r6 - c6) * inv_r2;
462 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
463 E_lj_p = int_bit * (c12 * (inv_r6 * inv_r6 + nbparam->repulsion_shift.cpot)*ONE_TWELVETH_F -
464 c6 * (inv_r6 + nbparam->dispersion_shift.cpot)*ONE_SIXTH_F);
466 #endif
467 #else /* ! LJ_COMB_LB || CALC_ENERGIES */
468 float sig_r = sigma*inv_r;
469 float sig_r2 = sig_r*sig_r;
470 float sig_r6 = sig_r2*sig_r2*sig_r2;
471 #if defined EXCLUSION_FORCES
472 sig_r6 *= int_bit;
473 #endif /* EXCLUSION_FORCES */
475 F_invr = epsilon * sig_r6 * (sig_r6 - 1.0f) * inv_r2;
476 #endif /* ! LJ_COMB_LB || CALC_ENERGIES */
479 #ifdef LJ_FORCE_SWITCH
480 #ifdef CALC_ENERGIES
481 calculate_force_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
482 #else
483 calculate_force_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr);
484 #endif /* CALC_ENERGIES */
485 #endif /* LJ_FORCE_SWITCH */
488 #ifdef LJ_EWALD
489 #ifdef LJ_EWALD_COMB_GEOM
490 #ifdef CALC_ENERGIES
491 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);
492 #else
493 calculate_lj_ewald_comb_geom_F(nbfp_comb_climg2d, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6, &F_invr);
494 #endif /* CALC_ENERGIES */
495 #elif defined LJ_EWALD_COMB_LB
496 calculate_lj_ewald_comb_LB_F_E(nbfp_comb_climg2d, nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6,
497 #ifdef CALC_ENERGIES
498 int_bit, true, &F_invr, &E_lj_p
499 #else
500 0, false, &F_invr, 0
501 #endif /* CALC_ENERGIES */
502 );
503 #endif /* LJ_EWALD_COMB_GEOM */
504 #endif /* LJ_EWALD */
506 #ifdef LJ_POT_SWITCH
507 #ifdef CALC_ENERGIES
508 calculate_potential_switch_F_E(nbparam, inv_r, r2, &F_invr, &E_lj_p);
509 #else
510 calculate_potential_switch_F(nbparam, inv_r, r2, &F_invr, &E_lj_p);
511 #endif /* CALC_ENERGIES */
512 #endif /* LJ_POT_SWITCH */
514 #ifdef VDW_CUTOFF_CHECK
515 /* Separate VDW cut-off check to enable twin-range cut-offs
516 * (rvdw < rcoulomb <= rlist)
517 */
518 vdw_in_range = (r2 < rvdw_sq) ? 1.0f : 0.0f;
519 F_invr *= vdw_in_range;
520 #ifdef CALC_ENERGIES
521 E_lj_p *= vdw_in_range;
522 #endif
523 #endif /* VDW_CUTOFF_CHECK */
525 #ifdef CALC_ENERGIES
526 E_lj += E_lj_p;
528 #endif
531 #ifdef EL_CUTOFF
532 #ifdef EXCLUSION_FORCES
533 F_invr += qi * qj_f * int_bit * inv_r2 * inv_r;
534 #else
535 F_invr += qi * qj_f * inv_r2 * inv_r;
536 #endif
537 #endif
538 #ifdef EL_RF
539 F_invr += qi * qj_f * (int_bit*inv_r2 * inv_r - two_k_rf);
540 #endif
541 #if defined EL_EWALD_ANA
542 F_invr += qi * qj_f * (int_bit*inv_r2*inv_r + pmecorrF(beta2*r2)*beta3);
543 #elif defined EL_EWALD_TAB
544 F_invr += qi * qj_f * (int_bit*inv_r2 -
545 interpolate_coulomb_force_r(coulomb_tab_climg2d, r2 * inv_r, coulomb_tab_scale)
546 ) * inv_r;
547 #endif /* EL_EWALD_ANA/TAB */
549 #ifdef CALC_ENERGIES
550 #ifdef EL_CUTOFF
551 E_el += qi * qj_f * (int_bit*inv_r - c_rf);
552 #endif
553 #ifdef EL_RF
554 E_el += qi * qj_f * (int_bit*inv_r + 0.5f * two_k_rf * r2 - c_rf);
555 #endif
556 #ifdef EL_EWALD_ANY
557 /* 1.0f - erff is faster than erfcf */
558 E_el += qi * qj_f * (inv_r * (int_bit - erf(r2 * inv_r * beta)) - int_bit * ewald_shift);
559 #endif /* EL_EWALD_ANY */
560 #endif
561 f_ij = rv * F_invr;
563 /* accumulate j forces in registers */
564 fcj_buf -= f_ij;
566 /* accumulate i forces in registers */
567 fci_buf[ci_offset] += f_ij;
568 }
569 }
571 /* shift the mask bit by 1 */
572 mask_ji += mask_ji;
573 }
575 /* reduce j forces */
577 /* store j forces in shmem */
578 f_buf[ tidx] = fcj_buf.x;
579 f_buf[ FBUF_STRIDE + tidx] = fcj_buf.y;
580 f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
582 reduce_force_j_generic(f_buf, f, tidxi, tidxj, aj);
583 }
584 }
585 #ifdef PRUNE_NBL
586 /* Update the imask with the new one which does not contain the
587 out of range clusters anymore. */
589 pl_cj4[j4].imei[widx].imask = imask;
590 #endif
591 }
592 }
594 /* skip central shifts when summing shift forces */
595 if (nb_sci.shift == CENTRAL)
596 {
597 bCalcFshift = false;
598 }
600 fshift_buf = 0.0f;
602 /* reduce i forces */
603 for (ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
604 {
605 ai = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
607 f_buf[ tidx] = fci_buf[ci_offset].x;
608 f_buf[ FBUF_STRIDE + tidx] = fci_buf[ci_offset].y;
609 f_buf[2 * FBUF_STRIDE + tidx] = fci_buf[ci_offset].z;
610 barrier(CLK_LOCAL_MEM_FENCE);
611 reduce_force_i(f_buf, f,
612 &fshift_buf, bCalcFshift,
613 tidxi, tidxj, ai);
614 barrier(CLK_LOCAL_MEM_FENCE);
615 }
617 /* add up local shift forces into global mem */
618 if (bCalcFshift)
619 {
620 /* Only threads with tidxj < 3 will update fshift.
621 The threads performing the update must be the same with the threads
622 which stored the reduction result in reduce_force_i function
623 */
624 if (tidxj < 3)
625 atomicAdd_g_f(&(fshift[3 * nb_sci.shift + tidxj]), fshift_buf);
626 }
628 #ifdef CALC_ENERGIES
629 /* flush the energies to shmem and reduce them */
630 f_buf[ tidx] = E_lj;
631 f_buf[FBUF_STRIDE + tidx] = E_el;
632 reduce_energy_pow2(f_buf + (tidx & WARP_SIZE), e_lj, e_el, tidx & ~WARP_SIZE);
634 #endif
635 }
637 #undef EL_EWALD_ANY
638 #undef EXCLUSION_FORCES
639 #undef LJ_EWALD
641 #undef LJ_COMB