Bug Summary

File:gromacs/mdlib/nbnxn_search.c
Location:line 4095, column 39
Description:Array access (from variable 'bb') results in a null pointer dereference

Annotated Source Code

1/*
2 * This file is part of the GROMACS molecular simulation package.
3 *
4 * Copyright (c) 2012,2013,2014, 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#ifdef HAVE_CONFIG_H1
37#include <config.h>
38#endif
39
40#include <math.h>
41#include <string.h>
42#include <assert.h>
43
44#include "gromacs/utility/smalloc.h"
45#include "types/commrec.h"
46#include "macros.h"
47#include "gromacs/math/utilities.h"
48#include "gromacs/math/vec.h"
49#include "pbc.h"
50#include "nbnxn_consts.h"
51/* nbnxn_internal.h included gromacs/simd/macros.h */
52#include "nbnxn_internal.h"
53#ifdef GMX_NBNXN_SIMD
54#include "gromacs/simd/vector_operations.h"
55#endif
56#include "nbnxn_atomdata.h"
57#include "nbnxn_search.h"
58#include "gmx_omp_nthreads.h"
59#include "nrnb.h"
60#include "ns.h"
61
62#include "gromacs/fileio/gmxfio.h"
63
64#ifdef NBNXN_SEARCH_BB_SIMD4
65/* Always use 4-wide SIMD for bounding box calculations */
66
67# ifndef GMX_DOUBLE
68/* Single precision BBs + coordinates, we can also load coordinates with SIMD */
69# define NBNXN_SEARCH_SIMD4_FLOAT_X_BB
70# endif
71
72# if defined NBNXN_SEARCH_SIMD4_FLOAT_X_BB && (GPU_NSUBCELL(2*2*2) == 4 || GPU_NSUBCELL(2*2*2) == 8)
73/* Store bounding boxes with x, y and z coordinates in packs of 4 */
74# define NBNXN_PBB_SIMD4
75# endif
76
77/* The packed bounding box coordinate stride is always set to 4.
78 * With AVX we could use 8, but that turns out not to be faster.
79 */
80# define STRIDE_PBB4 4
81# define STRIDE_PBB_2LOG2 2
82
83#endif /* NBNXN_SEARCH_BB_SIMD4 */
84
85#ifdef GMX_NBNXN_SIMD
86
87/* The functions below are macros as they are performance sensitive */
88
89/* 4x4 list, pack=4: no complex conversion required */
90/* i-cluster to j-cluster conversion */
91#define CI_TO_CJ_J4(ci)(ci) (ci)
92/* cluster index to coordinate array index conversion */
93#define X_IND_CI_J4(ci)((ci)*(3*4)) ((ci)*STRIDE_P4(3*4))
94#define X_IND_CJ_J4(cj)((cj)*(3*4)) ((cj)*STRIDE_P4(3*4))
95
96/* 4x2 list, pack=4: j-cluster size is half the packing width */
97/* i-cluster to j-cluster conversion */
98#define CI_TO_CJ_J2(ci)((ci)<<1) ((ci)<<1)
99/* cluster index to coordinate array index conversion */
100#define X_IND_CI_J2(ci)((ci)*(3*4)) ((ci)*STRIDE_P4(3*4))
101#define X_IND_CJ_J2(cj)(((cj)>>1)*(3*4) + ((cj) & 1)*(4>>1)) (((cj)>>1)*STRIDE_P4(3*4) + ((cj) & 1)*(PACK_X44>>1))
102
103/* 4x8 list, pack=8: i-cluster size is half the packing width */
104/* i-cluster to j-cluster conversion */
105#define CI_TO_CJ_J8(ci)((ci)>>1) ((ci)>>1)
106/* cluster index to coordinate array index conversion */
107#define X_IND_CI_J8(ci)(((ci)>>1)*(3*8) + ((ci) & 1)*(8>>1)) (((ci)>>1)*STRIDE_P8(3*8) + ((ci) & 1)*(PACK_X88>>1))
108#define X_IND_CJ_J8(cj)((cj)*(3*8)) ((cj)*STRIDE_P8(3*8))
109
110/* The j-cluster size is matched to the SIMD width */
111#if GMX_SIMD_REAL_WIDTH4 == 2
112#define CI_TO_CJ_SIMD_4XN(ci)(ci) CI_TO_CJ_J2(ci)((ci)<<1)
113#define X_IND_CI_SIMD_4XN(ci)((ci)*(3*4)) X_IND_CI_J2(ci)((ci)*(3*4))
114#define X_IND_CJ_SIMD_4XN(cj)((cj)*(3*4)) X_IND_CJ_J2(cj)(((cj)>>1)*(3*4) + ((cj) & 1)*(4>>1))
115#else
116#if GMX_SIMD_REAL_WIDTH4 == 4
117#define CI_TO_CJ_SIMD_4XN(ci)(ci) CI_TO_CJ_J4(ci)(ci)
118#define X_IND_CI_SIMD_4XN(ci)((ci)*(3*4)) X_IND_CI_J4(ci)((ci)*(3*4))
119#define X_IND_CJ_SIMD_4XN(cj)((cj)*(3*4)) X_IND_CJ_J4(cj)((cj)*(3*4))
120#else
121#if GMX_SIMD_REAL_WIDTH4 == 8
122#define CI_TO_CJ_SIMD_4XN(ci)(ci) CI_TO_CJ_J8(ci)((ci)>>1)
123#define X_IND_CI_SIMD_4XN(ci)((ci)*(3*4)) X_IND_CI_J8(ci)(((ci)>>1)*(3*8) + ((ci) & 1)*(8>>1))
124#define X_IND_CJ_SIMD_4XN(cj)((cj)*(3*4)) X_IND_CJ_J8(cj)((cj)*(3*8))
125/* Half SIMD with j-cluster size */
126#define CI_TO_CJ_SIMD_2XNN(ci) CI_TO_CJ_J4(ci)(ci)
127#define X_IND_CI_SIMD_2XNN(ci) X_IND_CI_J4(ci)((ci)*(3*4))
128#define X_IND_CJ_SIMD_2XNN(cj) X_IND_CJ_J4(cj)((cj)*(3*4))
129#else
130#if GMX_SIMD_REAL_WIDTH4 == 16
131#define CI_TO_CJ_SIMD_2XNN(ci) CI_TO_CJ_J8(ci)((ci)>>1)
132#define X_IND_CI_SIMD_2XNN(ci) X_IND_CI_J8(ci)(((ci)>>1)*(3*8) + ((ci) & 1)*(8>>1))
133#define X_IND_CJ_SIMD_2XNN(cj) X_IND_CJ_J8(cj)((cj)*(3*8))
134#else
135#error "unsupported GMX_SIMD_REAL_WIDTH"
136#endif
137#endif
138#endif
139#endif
140
141#endif /* GMX_NBNXN_SIMD */
142
143
144#ifdef NBNXN_SEARCH_BB_SIMD4
145/* Store bounding boxes corners as quadruplets: xxxxyyyyzzzz */
146#define NBNXN_BBXXXX
147/* Size of bounding box corners quadruplet */
148#define NNBSBB_XXXX(2*3*4) (NNBSBB_D2*DIM3*STRIDE_PBB4)
149#endif
150
151/* We shift the i-particles backward for PBC.
152 * This leads to more conditionals than shifting forward.
153 * We do this to get more balanced pair lists.
154 */
155#define NBNXN_SHIFT_BACKWARD
156
157
158/* This define is a lazy way to avoid interdependence of the grid
159 * and searching data structures.
160 */
161#define NBNXN_NA_SC_MAX((2*2*2)*8) (GPU_NSUBCELL(2*2*2)*NBNXN_GPU_CLUSTER_SIZE8)
162
163
164static void nbs_cycle_clear(nbnxn_cycle_t *cc)
165{
166 int i;
167
168 for (i = 0; i < enbsCCnr; i++)
169 {
170 cc[i].count = 0;
171 cc[i].c = 0;
172 }
173}
174
175static double Mcyc_av(const nbnxn_cycle_t *cc)
176{
177 return (double)cc->c*1e-6/cc->count;
178}
179
180static void nbs_cycle_print(FILE *fp, const nbnxn_search_t nbs)
181{
182 int n;
183 int t;
184
185 fprintf(fp, "\n");
186 fprintf(fp, "ns %4d grid %4.1f search %4.1f red.f %5.3f",
187 nbs->cc[enbsCCgrid].count,
188 Mcyc_av(&nbs->cc[enbsCCgrid]),
189 Mcyc_av(&nbs->cc[enbsCCsearch]),
190 Mcyc_av(&nbs->cc[enbsCCreducef]));
191
192 if (nbs->nthread_max > 1)
193 {
194 if (nbs->cc[enbsCCcombine].count > 0)
195 {
196 fprintf(fp, " comb %5.2f",
197 Mcyc_av(&nbs->cc[enbsCCcombine]));
198 }
199 fprintf(fp, " s. th");
200 for (t = 0; t < nbs->nthread_max; t++)
201 {
202 fprintf(fp, " %4.1f",
203 Mcyc_av(&nbs->work[t].cc[enbsCCsearch]));
204 }
205 }
206 fprintf(fp, "\n");
207}
208
209static void nbnxn_grid_init(nbnxn_grid_t * grid)
210{
211 grid->cxy_na = NULL((void*)0);
212 grid->cxy_ind = NULL((void*)0);
213 grid->cxy_nalloc = 0;
214 grid->bb = NULL((void*)0);
215 grid->bbj = NULL((void*)0);
216 grid->nc_nalloc = 0;
217}
218
219static int get_2log(int n)
220{
221 int log2;
222
223 log2 = 0;
224 while ((1<<log2) < n)
225 {
226 log2++;
227 }
228 if ((1<<log2) != n)
229 {
230 gmx_fatal(FARGS0, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 230, "nbnxn na_c (%d) is not a power of 2", n);
231 }
232
233 return log2;
234}
235
236static int nbnxn_kernel_to_ci_size(int nb_kernel_type)
237{
238 switch (nb_kernel_type)
239 {
240 case nbnxnk4x4_PlainC:
241 case nbnxnk4xN_SIMD_4xN:
242 case nbnxnk4xN_SIMD_2xNN:
243 return NBNXN_CPU_CLUSTER_I_SIZE4;
244 case nbnxnk8x8x8_CUDA:
245 case nbnxnk8x8x8_PlainC:
246 /* The cluster size for super/sub lists is only set here.
247 * Any value should work for the pair-search and atomdata code.
248 * The kernels, of course, might require a particular value.
249 */
250 return NBNXN_GPU_CLUSTER_SIZE8;
251 default:
252 gmx_incons("unknown kernel type")_gmx_error("incons", "unknown kernel type", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 252);
253 }
254
255 return 0;
256}
257
258int nbnxn_kernel_to_cj_size(int nb_kernel_type)
259{
260 int nbnxn_simd_width = 0;
261 int cj_size = 0;
262
263#ifdef GMX_NBNXN_SIMD
264 nbnxn_simd_width = GMX_SIMD_REAL_WIDTH4;
265#endif
266
267 switch (nb_kernel_type)
268 {
269 case nbnxnk4x4_PlainC:
270 cj_size = NBNXN_CPU_CLUSTER_I_SIZE4;
271 break;
272 case nbnxnk4xN_SIMD_4xN:
273 cj_size = nbnxn_simd_width;
274 break;
275 case nbnxnk4xN_SIMD_2xNN:
276 cj_size = nbnxn_simd_width/2;
277 break;
278 case nbnxnk8x8x8_CUDA:
279 case nbnxnk8x8x8_PlainC:
280 cj_size = nbnxn_kernel_to_ci_size(nb_kernel_type);
281 break;
282 default:
283 gmx_incons("unknown kernel type")_gmx_error("incons", "unknown kernel type", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 283);
284 }
285
286 return cj_size;
287}
288
289static int ci_to_cj(int na_cj_2log, int ci)
290{
291 switch (na_cj_2log)
292 {
293 case 2: return ci; break;
294 case 1: return (ci<<1); break;
295 case 3: return (ci>>1); break;
296 }
297
298 return 0;
299}
300
301gmx_bool nbnxn_kernel_pairlist_simple(int nb_kernel_type)
302{
303 if (nb_kernel_type == nbnxnkNotSet)
304 {
305 gmx_fatal(FARGS0, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 305, "Non-bonded kernel type not set for Verlet-style pair-list.");
306 }
307
308 switch (nb_kernel_type)
309 {
310 case nbnxnk8x8x8_CUDA:
311 case nbnxnk8x8x8_PlainC:
312 return FALSE0;
313
314 case nbnxnk4x4_PlainC:
315 case nbnxnk4xN_SIMD_4xN:
316 case nbnxnk4xN_SIMD_2xNN:
317 return TRUE1;
318
319 default:
320 gmx_incons("Invalid nonbonded kernel type passed!")_gmx_error("incons", "Invalid nonbonded kernel type passed!",
"/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 320);
321 return FALSE0;
322 }
323}
324
325/* Initializes a single nbnxn_pairlist_t data structure */
326static void nbnxn_init_pairlist_fep(t_nblist *nl)
327{
328 nl->type = GMX_NBLIST_INTERACTION_FREE_ENERGY;
329 nl->igeometry = GMX_NBLIST_GEOMETRY_PARTICLE_PARTICLE;
330 /* The interaction functions are set in the free energy kernel fuction */
331 nl->ivdw = -1;
332 nl->ivdwmod = -1;
333 nl->ielec = -1;
334 nl->ielecmod = -1;
335
336 nl->maxnri = 0;
337 nl->maxnrj = 0;
338 nl->nri = 0;
339 nl->nrj = 0;
340 nl->iinr = NULL((void*)0);
341 nl->gid = NULL((void*)0);
342 nl->shift = NULL((void*)0);
343 nl->jindex = NULL((void*)0);
344 nl->jjnr = NULL((void*)0);
345 nl->excl_fep = NULL((void*)0);
346
347}
348
349void nbnxn_init_search(nbnxn_search_t * nbs_ptr,
350 ivec *n_dd_cells,
351 gmx_domdec_zones_t *zones,
352 gmx_bool bFEP,
353 int nthread_max)
354{
355 nbnxn_search_t nbs;
356 int d, g, t;
357
358 snew(nbs, 1)(nbs) = save_calloc("nbs", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 358, (1), sizeof(*(nbs)));
359 *nbs_ptr = nbs;
360
361 nbs->bFEP = bFEP;
362
363 nbs->DomDec = (n_dd_cells != NULL((void*)0));
364
365 clear_ivec(nbs->dd_dim);
366 nbs->ngrid = 1;
367 if (nbs->DomDec)
368 {
369 nbs->zones = zones;
370
371 for (d = 0; d < DIM3; d++)
372 {
373 if ((*n_dd_cells)[d] > 1)
374 {
375 nbs->dd_dim[d] = 1;
376 /* Each grid matches a DD zone */
377 nbs->ngrid *= 2;
378 }
379 }
380 }
381
382 snew(nbs->grid, nbs->ngrid)(nbs->grid) = save_calloc("nbs->grid", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 382, (nbs->ngrid), sizeof(*(nbs->grid)));
383 for (g = 0; g < nbs->ngrid; g++)
384 {
385 nbnxn_grid_init(&nbs->grid[g]);
386 }
387 nbs->cell = NULL((void*)0);
388 nbs->cell_nalloc = 0;
389 nbs->a = NULL((void*)0);
390 nbs->a_nalloc = 0;
391
392 nbs->nthread_max = nthread_max;
393
394 /* Initialize the work data structures for each thread */
395 snew(nbs->work, nbs->nthread_max)(nbs->work) = save_calloc("nbs->work", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 395, (nbs->nthread_max), sizeof(*(nbs->work)));
396 for (t = 0; t < nbs->nthread_max; t++)
397 {
398 nbs->work[t].cxy_na = NULL((void*)0);
399 nbs->work[t].cxy_na_nalloc = 0;
400 nbs->work[t].sort_work = NULL((void*)0);
401 nbs->work[t].sort_work_nalloc = 0;
402
403 snew(nbs->work[t].nbl_fep, 1)(nbs->work[t].nbl_fep) = save_calloc("nbs->work[t].nbl_fep"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 403, (1), sizeof(*(nbs->work[t].nbl_fep)));
404 nbnxn_init_pairlist_fep(nbs->work[t].nbl_fep);
405 }
406
407 /* Initialize detailed nbsearch cycle counting */
408 nbs->print_cycles = (getenv("GMX_NBNXN_CYCLE") != 0);
409 nbs->search_count = 0;
410 nbs_cycle_clear(nbs->cc);
411 for (t = 0; t < nbs->nthread_max; t++)
412 {
413 nbs_cycle_clear(nbs->work[t].cc);
414 }
415}
416
417static real grid_atom_density(int n, rvec corner0, rvec corner1)
418{
419 rvec size;
420
421 rvec_sub(corner1, corner0, size);
422
423 return n/(size[XX0]*size[YY1]*size[ZZ2]);
424}
425
426static int set_grid_size_xy(const nbnxn_search_t nbs,
427 nbnxn_grid_t *grid,
428 int dd_zone,
429 int n, rvec corner0, rvec corner1,
430 real atom_density)
431{
432 rvec size;
433 int na_c;
434 real adens, tlen, tlen_x, tlen_y, nc_max;
435 int t;
436
437 rvec_sub(corner1, corner0, size);
438
439 if (n > grid->na_sc)
440 {
441 /* target cell length */
442 if (grid->bSimple)
443 {
444 /* To minimize the zero interactions, we should make
445 * the largest of the i/j cell cubic.
446 */
447 na_c = max(grid->na_c, grid->na_cj)(((grid->na_c) > (grid->na_cj)) ? (grid->na_c) : (
grid->na_cj) );
448
449 /* Approximately cubic cells */
450 tlen = pow(na_c/atom_density, 1.0/3.0);
451 tlen_x = tlen;
452 tlen_y = tlen;
453 }
454 else
455 {
456 /* Approximately cubic sub cells */
457 tlen = pow(grid->na_c/atom_density, 1.0/3.0);
458 tlen_x = tlen*GPU_NSUBCELL_X2;
459 tlen_y = tlen*GPU_NSUBCELL_Y2;
460 }
461 /* We round ncx and ncy down, because we get less cell pairs
462 * in the nbsist when the fixed cell dimensions (x,y) are
463 * larger than the variable one (z) than the other way around.
464 */
465 grid->ncx = max(1, (int)(size[XX]/tlen_x))(((1) > ((int)(size[0]/tlen_x))) ? (1) : ((int)(size[0]/tlen_x
)) );
466 grid->ncy = max(1, (int)(size[YY]/tlen_y))(((1) > ((int)(size[1]/tlen_y))) ? (1) : ((int)(size[1]/tlen_y
)) );
467 }
468 else
469 {
470 grid->ncx = 1;
471 grid->ncy = 1;
472 }
473
474 grid->sx = size[XX0]/grid->ncx;
475 grid->sy = size[YY1]/grid->ncy;
476 grid->inv_sx = 1/grid->sx;
477 grid->inv_sy = 1/grid->sy;
478
479 if (dd_zone > 0)
480 {
481 /* This is a non-home zone, add an extra row of cells
482 * for particles communicated for bonded interactions.
483 * These can be beyond the cut-off. It doesn't matter where
484 * they end up on the grid, but for performance it's better
485 * if they don't end up in cells that can be within cut-off range.
486 */
487 grid->ncx++;
488 grid->ncy++;
489 }
490
491 /* We need one additional cell entry for particles moved by DD */
492 if (grid->ncx*grid->ncy+1 > grid->cxy_nalloc)
493 {
494 grid->cxy_nalloc = over_alloc_large(grid->ncx*grid->ncy+1)(int)(1.19*(grid->ncx*grid->ncy+1) + 1000);
495 srenew(grid->cxy_na, grid->cxy_nalloc)(grid->cxy_na) = save_realloc("grid->cxy_na", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 495, (grid->cxy_na), (grid->cxy_nalloc), sizeof(*(grid
->cxy_na)));
496 srenew(grid->cxy_ind, grid->cxy_nalloc+1)(grid->cxy_ind) = save_realloc("grid->cxy_ind", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 496, (grid->cxy_ind), (grid->cxy_nalloc+1), sizeof(*(
grid->cxy_ind)));
497 }
498 for (t = 0; t < nbs->nthread_max; t++)
499 {
500 if (grid->ncx*grid->ncy+1 > nbs->work[t].cxy_na_nalloc)
501 {
502 nbs->work[t].cxy_na_nalloc = over_alloc_large(grid->ncx*grid->ncy+1)(int)(1.19*(grid->ncx*grid->ncy+1) + 1000);
503 srenew(nbs->work[t].cxy_na, nbs->work[t].cxy_na_nalloc)(nbs->work[t].cxy_na) = save_realloc("nbs->work[t].cxy_na"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 503, (nbs->work[t].cxy_na), (nbs->work[t].cxy_na_nalloc
), sizeof(*(nbs->work[t].cxy_na)));
504 }
505 }
506
507 /* Worst case scenario of 1 atom in each last cell */
508 if (grid->na_cj <= grid->na_c)
509 {
510 nc_max = n/grid->na_sc + grid->ncx*grid->ncy;
511 }
512 else
513 {
514 nc_max = n/grid->na_sc + grid->ncx*grid->ncy*grid->na_cj/grid->na_c;
515 }
516
517 if (nc_max > grid->nc_nalloc)
518 {
519 grid->nc_nalloc = over_alloc_large(nc_max)(int)(1.19*(nc_max) + 1000);
520 srenew(grid->nsubc, grid->nc_nalloc)(grid->nsubc) = save_realloc("grid->nsubc", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 520, (grid->nsubc), (grid->nc_nalloc), sizeof(*(grid->
nsubc)));
521 srenew(grid->bbcz, grid->nc_nalloc*NNBSBB_D)(grid->bbcz) = save_realloc("grid->bbcz", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 521, (grid->bbcz), (grid->nc_nalloc*2), sizeof(*(grid
->bbcz)));
522
523 sfree_aligned(grid->bb)save_free_aligned("grid->bb", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 523, (grid->bb));
524 /* This snew also zeros the contents, this avoid possible
525 * floating exceptions in SIMD with the unused bb elements.
526 */
527 if (grid->bSimple)
528 {
529 snew_aligned(grid->bb, grid->nc_nalloc, 16)(grid->bb) = save_calloc_aligned("grid->bb", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 529, (grid->nc_nalloc), sizeof(*(grid->bb)), 16);
530 }
531 else
532 {
533#ifdef NBNXN_BBXXXX
534 int pbb_nalloc;
535
536 pbb_nalloc = grid->nc_nalloc*GPU_NSUBCELL(2*2*2)/STRIDE_PBB4*NNBSBB_XXXX(2*3*4);
537 snew_aligned(grid->pbb, pbb_nalloc, 16)(grid->pbb) = save_calloc_aligned("grid->pbb", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 537, (pbb_nalloc), sizeof(*(grid->pbb)), 16);
538#else
539 snew_aligned(grid->bb, grid->nc_nalloc*GPU_NSUBCELL, 16)(grid->bb) = save_calloc_aligned("grid->bb", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 539, (grid->nc_nalloc*(2*2*2)), sizeof(*(grid->bb)), 16
);
540#endif
541 }
542
543 if (grid->bSimple)
544 {
545 if (grid->na_cj == grid->na_c)
546 {
547 grid->bbj = grid->bb;
548 }
549 else
550 {
551 sfree_aligned(grid->bbj)save_free_aligned("grid->bbj", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 551, (grid->bbj));
552 snew_aligned(grid->bbj, grid->nc_nalloc*grid->na_c/grid->na_cj, 16)(grid->bbj) = save_calloc_aligned("grid->bbj", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 552, (grid->nc_nalloc*grid->na_c/grid->na_cj), sizeof
(*(grid->bbj)), 16);
553 }
554 }
555
556 srenew(grid->flags, grid->nc_nalloc)(grid->flags) = save_realloc("grid->flags", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 556, (grid->flags), (grid->nc_nalloc), sizeof(*(grid->
flags)));
557 if (nbs->bFEP)
558 {
559 srenew(grid->fep, grid->nc_nalloc*grid->na_sc/grid->na_c)(grid->fep) = save_realloc("grid->fep", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 559, (grid->fep), (grid->nc_nalloc*grid->na_sc/grid
->na_c), sizeof(*(grid->fep)));
560 }
561 }
562
563 copy_rvec(corner0, grid->c0);
564 copy_rvec(corner1, grid->c1);
565
566 return nc_max;
567}
568
569/* We need to sort paricles in grid columns on z-coordinate.
570 * As particle are very often distributed homogeneously, we a sorting
571 * algorithm similar to pigeonhole sort. We multiply the z-coordinate
572 * by a factor, cast to an int and try to store in that hole. If the hole
573 * is full, we move this or another particle. A second pass is needed to make
574 * contiguous elements. SORT_GRID_OVERSIZE is the ratio of holes to particles.
575 * 4 is the optimal value for homogeneous particle distribution and allows
576 * for an O(#particles) sort up till distributions were all particles are
577 * concentrated in 1/4 of the space. No NlogN fallback is implemented,
578 * as it can be expensive to detect imhomogeneous particle distributions.
579 * SGSF is the maximum ratio of holes used, in the worst case all particles
580 * end up in the last hole and we need #particles extra holes at the end.
581 */
582#define SORT_GRID_OVERSIZE4 4
583#define SGSF(4 + 1) (SORT_GRID_OVERSIZE4 + 1)
584
585/* Sort particle index a on coordinates x along dim.
586 * Backwards tells if we want decreasing iso increasing coordinates.
587 * h0 is the minimum of the coordinate range.
588 * invh is the 1/length of the sorting range.
589 * n_per_h (>=n) is the expected average number of particles per 1/invh
590 * sort is the sorting work array.
591 * sort should have a size of at least n_per_h*SORT_GRID_OVERSIZE + n,
592 * or easier, allocate at least n*SGSF elements.
593 */
594static void sort_atoms(int dim, gmx_bool Backwards,
595 int gmx_unused__attribute__ ((unused)) dd_zone,
596 int *a, int n, rvec *x,
597 real h0, real invh, int n_per_h,
598 int *sort)
599{
600 int nsort, i, c;
601 int zi, zim, zi_min, zi_max;
602 int cp, tmp;
603
604 if (n <= 1)
605 {
606 /* Nothing to do */
607 return;
608 }
609
610#ifndef NDEBUG1
611 if (n > n_per_h)
612 {
613 gmx_incons("n > n_per_h")_gmx_error("incons", "n > n_per_h", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 613);
614 }
615#endif
616
617 /* Transform the inverse range height into the inverse hole height */
618 invh *= n_per_h*SORT_GRID_OVERSIZE4;
619
620 /* Set nsort to the maximum possible number of holes used.
621 * In worst case all n elements end up in the last bin.
622 */
623 nsort = n_per_h*SORT_GRID_OVERSIZE4 + n;
624
625 /* Determine the index range used, so we can limit it for the second pass */
626 zi_min = INT_MAX2147483647;
627 zi_max = -1;
628
629 /* Sort the particles using a simple index sort */
630 for (i = 0; i < n; i++)
631 {
632 /* The cast takes care of float-point rounding effects below zero.
633 * This code assumes particles are less than 1/SORT_GRID_OVERSIZE
634 * times the box height out of the box.
635 */
636 zi = (int)((x[a[i]][dim] - h0)*invh);
637
638#ifndef NDEBUG1
639 /* As we can have rounding effect, we use > iso >= here */
640 if (zi < 0 || (dd_zone == 0 && zi > n_per_h*SORT_GRID_OVERSIZE4))
641 {
642 gmx_fatal(FARGS0, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 642, "(int)((x[%d][%c]=%f - %f)*%f) = %d, not in 0 - %d*%d\n",
643 a[i], 'x'+dim, x[a[i]][dim], h0, invh, zi,
644 n_per_h, SORT_GRID_OVERSIZE4);
645 }
646#endif
647
648 /* In a non-local domain, particles communcated for bonded interactions
649 * can be far beyond the grid size, which is set by the non-bonded
650 * cut-off distance. We sort such particles into the last cell.
651 */
652 if (zi > n_per_h*SORT_GRID_OVERSIZE4)
653 {
654 zi = n_per_h*SORT_GRID_OVERSIZE4;
655 }
656
657 /* Ideally this particle should go in sort cell zi,
658 * but that might already be in use,
659 * in that case find the first empty cell higher up
660 */
661 if (sort[zi] < 0)
662 {
663 sort[zi] = a[i];
664 zi_min = min(zi_min, zi)(((zi_min) < (zi)) ? (zi_min) : (zi) );
665 zi_max = max(zi_max, zi)(((zi_max) > (zi)) ? (zi_max) : (zi) );
666 }
667 else
668 {
669 /* We have multiple atoms in the same sorting slot.
670 * Sort on real z for minimal bounding box size.
671 * There is an extra check for identical z to ensure
672 * well-defined output order, independent of input order
673 * to ensure binary reproducibility after restarts.
674 */
675 while (sort[zi] >= 0 && ( x[a[i]][dim] > x[sort[zi]][dim] ||
676 (x[a[i]][dim] == x[sort[zi]][dim] &&
677 a[i] > sort[zi])))
678 {
679 zi++;
680 }
681
682 if (sort[zi] >= 0)
683 {
684 /* Shift all elements by one slot until we find an empty slot */
685 cp = sort[zi];
686 zim = zi + 1;
687 while (sort[zim] >= 0)
688 {
689 tmp = sort[zim];
690 sort[zim] = cp;
691 cp = tmp;
692 zim++;
693 }
694 sort[zim] = cp;
695 zi_max = max(zi_max, zim)(((zi_max) > (zim)) ? (zi_max) : (zim) );
696 }
697 sort[zi] = a[i];
698 zi_max = max(zi_max, zi)(((zi_max) > (zi)) ? (zi_max) : (zi) );
699 }
700 }
701
702 c = 0;
703 if (!Backwards)
704 {
705 for (zi = 0; zi < nsort; zi++)
706 {
707 if (sort[zi] >= 0)
708 {
709 a[c++] = sort[zi];
710 sort[zi] = -1;
711 }
712 }
713 }
714 else
715 {
716 for (zi = zi_max; zi >= zi_min; zi--)
717 {
718 if (sort[zi] >= 0)
719 {
720 a[c++] = sort[zi];
721 sort[zi] = -1;
722 }
723 }
724 }
725 if (c < n)
726 {
727 gmx_incons("Lost particles while sorting")_gmx_error("incons", "Lost particles while sorting", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 727);
728 }
729}
730
731#ifdef GMX_DOUBLE
732#define R2F_D(x)(x) ((float)((x) >= 0 ? ((1-GMX_FLOAT_EPS5.96046448E-08)*(x)) : ((1+GMX_FLOAT_EPS5.96046448E-08)*(x))))
733#define R2F_U(x)(x) ((float)((x) >= 0 ? ((1+GMX_FLOAT_EPS5.96046448E-08)*(x)) : ((1-GMX_FLOAT_EPS5.96046448E-08)*(x))))
734#else
735#define R2F_D(x)(x) (x)
736#define R2F_U(x)(x) (x)
737#endif
738
739/* Coordinate order x,y,z, bb order xyz0 */
740static void calc_bounding_box(int na, int stride, const real *x, nbnxn_bb_t *bb)
741{
742 int i, j;
743 real xl, xh, yl, yh, zl, zh;
744
745 i = 0;
746 xl = x[i+XX0];
747 xh = x[i+XX0];
748 yl = x[i+YY1];
749 yh = x[i+YY1];
750 zl = x[i+ZZ2];
751 zh = x[i+ZZ2];
752 i += stride;
753 for (j = 1; j < na; j++)
754 {
755 xl = min(xl, x[i+XX])(((xl) < (x[i+0])) ? (xl) : (x[i+0]) );
756 xh = max(xh, x[i+XX])(((xh) > (x[i+0])) ? (xh) : (x[i+0]) );
757 yl = min(yl, x[i+YY])(((yl) < (x[i+1])) ? (yl) : (x[i+1]) );
758 yh = max(yh, x[i+YY])(((yh) > (x[i+1])) ? (yh) : (x[i+1]) );
759 zl = min(zl, x[i+ZZ])(((zl) < (x[i+2])) ? (zl) : (x[i+2]) );
760 zh = max(zh, x[i+ZZ])(((zh) > (x[i+2])) ? (zh) : (x[i+2]) );
761 i += stride;
762 }
763 /* Note: possible double to float conversion here */
764 bb->lower[BB_X0] = R2F_D(xl)(xl);
765 bb->lower[BB_Y1] = R2F_D(yl)(yl);
766 bb->lower[BB_Z2] = R2F_D(zl)(zl);
767 bb->upper[BB_X0] = R2F_U(xh)(xh);
768 bb->upper[BB_Y1] = R2F_U(yh)(yh);
769 bb->upper[BB_Z2] = R2F_U(zh)(zh);
770}
771
772/* Packed coordinates, bb order xyz0 */
773static void calc_bounding_box_x_x4(int na, const real *x, nbnxn_bb_t *bb)
774{
775 int j;
776 real xl, xh, yl, yh, zl, zh;
777
778 xl = x[XX0*PACK_X44];
779 xh = x[XX0*PACK_X44];
780 yl = x[YY1*PACK_X44];
781 yh = x[YY1*PACK_X44];
782 zl = x[ZZ2*PACK_X44];
783 zh = x[ZZ2*PACK_X44];
784 for (j = 1; j < na; j++)
785 {
786 xl = min(xl, x[j+XX*PACK_X4])(((xl) < (x[j+0*4])) ? (xl) : (x[j+0*4]) );
787 xh = max(xh, x[j+XX*PACK_X4])(((xh) > (x[j+0*4])) ? (xh) : (x[j+0*4]) );
788 yl = min(yl, x[j+YY*PACK_X4])(((yl) < (x[j+1*4])) ? (yl) : (x[j+1*4]) );
789 yh = max(yh, x[j+YY*PACK_X4])(((yh) > (x[j+1*4])) ? (yh) : (x[j+1*4]) );
790 zl = min(zl, x[j+ZZ*PACK_X4])(((zl) < (x[j+2*4])) ? (zl) : (x[j+2*4]) );
791 zh = max(zh, x[j+ZZ*PACK_X4])(((zh) > (x[j+2*4])) ? (zh) : (x[j+2*4]) );
792 }
793 /* Note: possible double to float conversion here */
794 bb->lower[BB_X0] = R2F_D(xl)(xl);
795 bb->lower[BB_Y1] = R2F_D(yl)(yl);
796 bb->lower[BB_Z2] = R2F_D(zl)(zl);
797 bb->upper[BB_X0] = R2F_U(xh)(xh);
798 bb->upper[BB_Y1] = R2F_U(yh)(yh);
799 bb->upper[BB_Z2] = R2F_U(zh)(zh);
800}
801
802/* Packed coordinates, bb order xyz0 */
803static void calc_bounding_box_x_x8(int na, const real *x, nbnxn_bb_t *bb)
804{
805 int j;
806 real xl, xh, yl, yh, zl, zh;
807
808 xl = x[XX0*PACK_X88];
809 xh = x[XX0*PACK_X88];
810 yl = x[YY1*PACK_X88];
811 yh = x[YY1*PACK_X88];
812 zl = x[ZZ2*PACK_X88];
813 zh = x[ZZ2*PACK_X88];
814 for (j = 1; j < na; j++)
815 {
816 xl = min(xl, x[j+XX*PACK_X8])(((xl) < (x[j+0*8])) ? (xl) : (x[j+0*8]) );
817 xh = max(xh, x[j+XX*PACK_X8])(((xh) > (x[j+0*8])) ? (xh) : (x[j+0*8]) );
818 yl = min(yl, x[j+YY*PACK_X8])(((yl) < (x[j+1*8])) ? (yl) : (x[j+1*8]) );
819 yh = max(yh, x[j+YY*PACK_X8])(((yh) > (x[j+1*8])) ? (yh) : (x[j+1*8]) );
820 zl = min(zl, x[j+ZZ*PACK_X8])(((zl) < (x[j+2*8])) ? (zl) : (x[j+2*8]) );
821 zh = max(zh, x[j+ZZ*PACK_X8])(((zh) > (x[j+2*8])) ? (zh) : (x[j+2*8]) );
822 }
823 /* Note: possible double to float conversion here */
824 bb->lower[BB_X0] = R2F_D(xl)(xl);
825 bb->lower[BB_Y1] = R2F_D(yl)(yl);
826 bb->lower[BB_Z2] = R2F_D(zl)(zl);
827 bb->upper[BB_X0] = R2F_U(xh)(xh);
828 bb->upper[BB_Y1] = R2F_U(yh)(yh);
829 bb->upper[BB_Z2] = R2F_U(zh)(zh);
830}
831
832/* Packed coordinates, bb order xyz0 */
833static void calc_bounding_box_x_x4_halves(int na, const real *x,
834 nbnxn_bb_t *bb, nbnxn_bb_t *bbj)
835{
836 calc_bounding_box_x_x4(min(na, 2)(((na) < (2)) ? (na) : (2) ), x, bbj);
837
838 if (na > 2)
839 {
840 calc_bounding_box_x_x4(min(na-2, 2)(((na-2) < (2)) ? (na-2) : (2) ), x+(PACK_X44>>1), bbj+1);
841 }
842 else
843 {
844 /* Set the "empty" bounding box to the same as the first one,
845 * so we don't need to treat special cases in the rest of the code.
846 */
847#ifdef NBNXN_SEARCH_BB_SIMD4
848 gmx_simd4_store_f_mm_store_ps(&bbj[1].lower[0], gmx_simd4_load_f_mm_load_ps(&bbj[0].lower[0]));
849 gmx_simd4_store_f_mm_store_ps(&bbj[1].upper[0], gmx_simd4_load_f_mm_load_ps(&bbj[0].upper[0]));
850#else
851 bbj[1] = bbj[0];
852#endif
853 }
854
855#ifdef NBNXN_SEARCH_BB_SIMD4
856 gmx_simd4_store_f_mm_store_ps(&bb->lower[0],
857 gmx_simd4_min_f_mm_min_ps(gmx_simd4_load_f_mm_load_ps(&bbj[0].lower[0]),
858 gmx_simd4_load_f_mm_load_ps(&bbj[1].lower[0])));
859 gmx_simd4_store_f_mm_store_ps(&bb->upper[0],
860 gmx_simd4_max_f_mm_max_ps(gmx_simd4_load_f_mm_load_ps(&bbj[0].upper[0]),
861 gmx_simd4_load_f_mm_load_ps(&bbj[1].upper[0])));
862#else
863 {
864 int i;
865
866 for (i = 0; i < NNBSBB_C4; i++)
867 {
868 bb->lower[i] = min(bbj[0].lower[i], bbj[1].lower[i])(((bbj[0].lower[i]) < (bbj[1].lower[i])) ? (bbj[0].lower[i
]) : (bbj[1].lower[i]) );
869 bb->upper[i] = max(bbj[0].upper[i], bbj[1].upper[i])(((bbj[0].upper[i]) > (bbj[1].upper[i])) ? (bbj[0].upper[i
]) : (bbj[1].upper[i]) );
870 }
871 }
872#endif
873}
874
875#ifdef NBNXN_SEARCH_BB_SIMD4
876
877/* Coordinate order xyz, bb order xxxxyyyyzzzz */
878static void calc_bounding_box_xxxx(int na, int stride, const real *x, float *bb)
879{
880 int i, j;
881 real xl, xh, yl, yh, zl, zh;
882
883 i = 0;
884 xl = x[i+XX0];
885 xh = x[i+XX0];
886 yl = x[i+YY1];
887 yh = x[i+YY1];
888 zl = x[i+ZZ2];
889 zh = x[i+ZZ2];
890 i += stride;
891 for (j = 1; j < na; j++)
892 {
893 xl = min(xl, x[i+XX])(((xl) < (x[i+0])) ? (xl) : (x[i+0]) );
894 xh = max(xh, x[i+XX])(((xh) > (x[i+0])) ? (xh) : (x[i+0]) );
895 yl = min(yl, x[i+YY])(((yl) < (x[i+1])) ? (yl) : (x[i+1]) );
896 yh = max(yh, x[i+YY])(((yh) > (x[i+1])) ? (yh) : (x[i+1]) );
897 zl = min(zl, x[i+ZZ])(((zl) < (x[i+2])) ? (zl) : (x[i+2]) );
898 zh = max(zh, x[i+ZZ])(((zh) > (x[i+2])) ? (zh) : (x[i+2]) );
899 i += stride;
900 }
901 /* Note: possible double to float conversion here */
902 bb[0*STRIDE_PBB4] = R2F_D(xl)(xl);
903 bb[1*STRIDE_PBB4] = R2F_D(yl)(yl);
904 bb[2*STRIDE_PBB4] = R2F_D(zl)(zl);
905 bb[3*STRIDE_PBB4] = R2F_U(xh)(xh);
906 bb[4*STRIDE_PBB4] = R2F_U(yh)(yh);
907 bb[5*STRIDE_PBB4] = R2F_U(zh)(zh);
908}
909
910#endif /* NBNXN_SEARCH_BB_SIMD4 */
911
912#ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
913
914/* Coordinate order xyz?, bb order xyz0 */
915static void calc_bounding_box_simd4(int na, const float *x, nbnxn_bb_t *bb)
916{
917 gmx_simd4_float_t__m128 bb_0_S, bb_1_S;
918 gmx_simd4_float_t__m128 x_S;
919
920 int i;
921
922 bb_0_S = gmx_simd4_load_f_mm_load_ps(x);
923 bb_1_S = bb_0_S;
924
925 for (i = 1; i < na; i++)
926 {
927 x_S = gmx_simd4_load_f_mm_load_ps(x+i*NNBSBB_C4);
928 bb_0_S = gmx_simd4_min_f_mm_min_ps(bb_0_S, x_S);
929 bb_1_S = gmx_simd4_max_f_mm_max_ps(bb_1_S, x_S);
930 }
931
932 gmx_simd4_store_f_mm_store_ps(&bb->lower[0], bb_0_S);
933 gmx_simd4_store_f_mm_store_ps(&bb->upper[0], bb_1_S);
934}
935
936/* Coordinate order xyz?, bb order xxxxyyyyzzzz */
937static void calc_bounding_box_xxxx_simd4(int na, const float *x,
938 nbnxn_bb_t *bb_work_aligned,
939 real *bb)
940{
941 calc_bounding_box_simd4(na, x, bb_work_aligned);
942
943 bb[0*STRIDE_PBB4] = bb_work_aligned->lower[BB_X0];
944 bb[1*STRIDE_PBB4] = bb_work_aligned->lower[BB_Y1];
945 bb[2*STRIDE_PBB4] = bb_work_aligned->lower[BB_Z2];
946 bb[3*STRIDE_PBB4] = bb_work_aligned->upper[BB_X0];
947 bb[4*STRIDE_PBB4] = bb_work_aligned->upper[BB_Y1];
948 bb[5*STRIDE_PBB4] = bb_work_aligned->upper[BB_Z2];
949}
950
951#endif /* NBNXN_SEARCH_SIMD4_FLOAT_X_BB */
952
953
954/* Combines pairs of consecutive bounding boxes */
955static void combine_bounding_box_pairs(nbnxn_grid_t *grid, const nbnxn_bb_t *bb)
956{
957 int i, j, sc2, nc2, c2;
958
959 for (i = 0; i < grid->ncx*grid->ncy; i++)
960 {
961 /* Starting bb in a column is expected to be 2-aligned */
962 sc2 = grid->cxy_ind[i]>>1;
963 /* For odd numbers skip the last bb here */
964 nc2 = (grid->cxy_na[i]+3)>>(2+1);
965 for (c2 = sc2; c2 < sc2+nc2; c2++)
966 {
967#ifdef NBNXN_SEARCH_BB_SIMD4
968 gmx_simd4_float_t__m128 min_S, max_S;
969
970 min_S = gmx_simd4_min_f_mm_min_ps(gmx_simd4_load_f_mm_load_ps(&bb[c2*2+0].lower[0]),
971 gmx_simd4_load_f_mm_load_ps(&bb[c2*2+1].lower[0]));
972 max_S = gmx_simd4_max_f_mm_max_ps(gmx_simd4_load_f_mm_load_ps(&bb[c2*2+0].upper[0]),
973 gmx_simd4_load_f_mm_load_ps(&bb[c2*2+1].upper[0]));
974 gmx_simd4_store_f_mm_store_ps(&grid->bbj[c2].lower[0], min_S);
975 gmx_simd4_store_f_mm_store_ps(&grid->bbj[c2].upper[0], max_S);
976#else
977 for (j = 0; j < NNBSBB_C4; j++)
978 {
979 grid->bbj[c2].lower[j] = min(bb[c2*2+0].lower[j],(((bb[c2*2+0].lower[j]) < (bb[c2*2+1].lower[j])) ? (bb[c2*
2+0].lower[j]) : (bb[c2*2+1].lower[j]) )
980 bb[c2*2+1].lower[j])(((bb[c2*2+0].lower[j]) < (bb[c2*2+1].lower[j])) ? (bb[c2*
2+0].lower[j]) : (bb[c2*2+1].lower[j]) );
981 grid->bbj[c2].upper[j] = max(bb[c2*2+0].upper[j],(((bb[c2*2+0].upper[j]) > (bb[c2*2+1].upper[j])) ? (bb[c2*
2+0].upper[j]) : (bb[c2*2+1].upper[j]) )
982 bb[c2*2+1].upper[j])(((bb[c2*2+0].upper[j]) > (bb[c2*2+1].upper[j])) ? (bb[c2*
2+0].upper[j]) : (bb[c2*2+1].upper[j]) );
983 }
984#endif
985 }
986 if (((grid->cxy_na[i]+3)>>2) & 1)
987 {
988 /* The bb count in this column is odd: duplicate the last bb */
989 for (j = 0; j < NNBSBB_C4; j++)
990 {
991 grid->bbj[c2].lower[j] = bb[c2*2].lower[j];
992 grid->bbj[c2].upper[j] = bb[c2*2].upper[j];
993 }
994 }
995 }
996}
997
998
999/* Prints the average bb size, used for debug output */
1000static void print_bbsizes_simple(FILE *fp,
1001 const nbnxn_search_t nbs,
1002 const nbnxn_grid_t *grid)
1003{
1004 int c, d;
1005 dvec ba;
1006
1007 clear_dvec(ba);
1008 for (c = 0; c < grid->nc; c++)
1009 {
1010 for (d = 0; d < DIM3; d++)
1011 {
1012 ba[d] += grid->bb[c].upper[d] - grid->bb[c].lower[d];
1013 }
1014 }
1015 dsvmul(1.0/grid->nc, ba, ba);
1016
1017 fprintf(fp, "ns bb: %4.2f %4.2f %4.2f %4.2f %4.2f %4.2f rel %4.2f %4.2f %4.2f\n",
1018 nbs->box[XX0][XX0]/grid->ncx,
1019 nbs->box[YY1][YY1]/grid->ncy,
1020 nbs->box[ZZ2][ZZ2]*grid->ncx*grid->ncy/grid->nc,
1021 ba[XX0], ba[YY1], ba[ZZ2],
1022 ba[XX0]*grid->ncx/nbs->box[XX0][XX0],
1023 ba[YY1]*grid->ncy/nbs->box[YY1][YY1],
1024 ba[ZZ2]*grid->nc/(grid->ncx*grid->ncy*nbs->box[ZZ2][ZZ2]));
1025}
1026
1027/* Prints the average bb size, used for debug output */
1028static void print_bbsizes_supersub(FILE *fp,
1029 const nbnxn_search_t nbs,
1030 const nbnxn_grid_t *grid)
1031{
1032 int ns, c, s;
1033 dvec ba;
1034
1035 clear_dvec(ba);
1036 ns = 0;
1037 for (c = 0; c < grid->nc; c++)
1038 {
1039#ifdef NBNXN_BBXXXX
1040 for (s = 0; s < grid->nsubc[c]; s += STRIDE_PBB4)
1041 {
1042 int cs_w, i, d;
1043
1044 cs_w = (c*GPU_NSUBCELL(2*2*2) + s)/STRIDE_PBB4;
1045 for (i = 0; i < STRIDE_PBB4; i++)
1046 {
1047 for (d = 0; d < DIM3; d++)
1048 {
1049 ba[d] +=
1050 grid->pbb[cs_w*NNBSBB_XXXX(2*3*4)+(DIM3+d)*STRIDE_PBB4+i] -
1051 grid->pbb[cs_w*NNBSBB_XXXX(2*3*4)+ d *STRIDE_PBB4+i];
1052 }
1053 }
1054 }
1055#else
1056 for (s = 0; s < grid->nsubc[c]; s++)
1057 {
1058 int cs, d;
1059
1060 cs = c*GPU_NSUBCELL(2*2*2) + s;
1061 for (d = 0; d < DIM3; d++)
1062 {
1063 ba[d] += grid->bb[cs].upper[d] - grid->bb[cs].lower[d];
1064 }
1065 }
1066#endif
1067 ns += grid->nsubc[c];
1068 }
1069 dsvmul(1.0/ns, ba, ba);
1070
1071 fprintf(fp, "ns bb: %4.2f %4.2f %4.2f %4.2f %4.2f %4.2f rel %4.2f %4.2f %4.2f\n",
1072 nbs->box[XX0][XX0]/(grid->ncx*GPU_NSUBCELL_X2),
1073 nbs->box[YY1][YY1]/(grid->ncy*GPU_NSUBCELL_Y2),
1074 nbs->box[ZZ2][ZZ2]*grid->ncx*grid->ncy/(grid->nc*GPU_NSUBCELL_Z2),
1075 ba[XX0], ba[YY1], ba[ZZ2],
1076 ba[XX0]*grid->ncx*GPU_NSUBCELL_X2/nbs->box[XX0][XX0],
1077 ba[YY1]*grid->ncy*GPU_NSUBCELL_Y2/nbs->box[YY1][YY1],
1078 ba[ZZ2]*grid->nc*GPU_NSUBCELL_Z2/(grid->ncx*grid->ncy*nbs->box[ZZ2][ZZ2]));
1079}
1080
1081/* Potentially sorts atoms on LJ coefficients !=0 and ==0.
1082 * Also sets interaction flags.
1083 */
1084void sort_on_lj(int na_c,
1085 int a0, int a1, const int *atinfo,
1086 int *order,
1087 int *flags)
1088{
1089 int subc, s, a, n1, n2, a_lj_max, i, j;
1090 int sort1[NBNXN_NA_SC_MAX((2*2*2)*8)/GPU_NSUBCELL(2*2*2)];
1091 int sort2[NBNXN_NA_SC_MAX((2*2*2)*8)/GPU_NSUBCELL(2*2*2)];
1092 gmx_bool haveQ, bFEP;
1093
1094 *flags = 0;
1095
1096 subc = 0;
1097 for (s = a0; s < a1; s += na_c)
1098 {
1099 /* Make lists for this (sub-)cell on atoms with and without LJ */
1100 n1 = 0;
1101 n2 = 0;
1102 haveQ = FALSE0;
1103 a_lj_max = -1;
1104 for (a = s; a < min(s+na_c, a1)(((s+na_c) < (a1)) ? (s+na_c) : (a1) ); a++)
1105 {
1106 haveQ = haveQ || GET_CGINFO_HAS_Q(atinfo[order[a]])( (atinfo[order[a]]) & (1<<24));
1107
1108 if (GET_CGINFO_HAS_VDW(atinfo[order[a]])( (atinfo[order[a]]) & (1<<23)))
1109 {
1110 sort1[n1++] = order[a];
1111 a_lj_max = a;
1112 }
1113 else
1114 {
1115 sort2[n2++] = order[a];
1116 }
1117 }
1118
1119 /* If we don't have atoms with LJ, there's nothing to sort */
1120 if (n1 > 0)
1121 {
1122 *flags |= NBNXN_CI_DO_LJ(subc)(1<<(7+3*(subc)));
1123
1124 if (2*n1 <= na_c)
1125 {
1126 /* Only sort when strictly necessary. Ordering particles
1127 * Ordering particles can lead to less accurate summation
1128 * due to rounding, both for LJ and Coulomb interactions.
1129 */
1130 if (2*(a_lj_max - s) >= na_c)
1131 {
1132 for (i = 0; i < n1; i++)
1133 {
1134 order[a0+i] = sort1[i];
1135 }
1136 for (j = 0; j < n2; j++)
1137 {
1138 order[a0+n1+j] = sort2[j];
1139 }
1140 }
1141
1142 *flags |= NBNXN_CI_HALF_LJ(subc)(1<<(8+3*(subc)));
1143 }
1144 }
1145 if (haveQ)
1146 {
1147 *flags |= NBNXN_CI_DO_COUL(subc)(1<<(9+3*(subc)));
1148 }
1149 subc++;
1150 }
1151}
1152
1153/* Fill a pair search cell with atoms.
1154 * Potentially sorts atoms and sets the interaction flags.
1155 */
1156void fill_cell(const nbnxn_search_t nbs,
1157 nbnxn_grid_t *grid,
1158 nbnxn_atomdata_t *nbat,
1159 int a0, int a1,
1160 const int *atinfo,
1161 rvec *x,
1162 int sx, int sy, int sz,
1163 nbnxn_bb_t gmx_unused__attribute__ ((unused)) *bb_work_aligned)
1164{
1165 int na, a;
1166 size_t offset;
1167 nbnxn_bb_t *bb_ptr;
1168#ifdef NBNXN_BBXXXX
1169 float *pbb_ptr;
1170#endif
1171
1172 na = a1 - a0;
1173
1174 if (grid->bSimple)
1175 {
1176 sort_on_lj(grid->na_c, a0, a1, atinfo, nbs->a,
1177 grid->flags+(a0>>grid->na_c_2log)-grid->cell0);
1178 }
1179
1180 if (nbs->bFEP)
1181 {
1182 /* Set the fep flag for perturbed atoms in this (sub-)cell */
1183 int c, at;
1184
1185 /* The grid-local cluster/(sub-)cell index */
1186 c = (a0 >> grid->na_c_2log) - grid->cell0*(grid->bSimple ? 1 : GPU_NSUBCELL(2*2*2));
1187 grid->fep[c] = 0;
1188 for (at = a0; at < a1; at++)
1189 {
1190 if (nbs->a[at] >= 0 && GET_CGINFO_FEP(atinfo[nbs->a[at]])( (atinfo[nbs->a[at]]) & (1<<15)))
1191 {
1192 grid->fep[c] |= (1 << (at - a0));
1193 }
1194 }
1195 }
1196
1197 /* Now we have sorted the atoms, set the cell indices */
1198 for (a = a0; a < a1; a++)
1199 {
1200 nbs->cell[nbs->a[a]] = a;
1201 }
1202
1203 copy_rvec_to_nbat_real(nbs->a+a0, a1-a0, grid->na_c, x,
1204 nbat->XFormat, nbat->x, a0,
1205 sx, sy, sz);
1206
1207 if (nbat->XFormat == nbatX4)
1208 {
1209 /* Store the bounding boxes as xyz.xyz. */
1210 offset = (a0 - grid->cell0*grid->na_sc) >> grid->na_c_2log;
1211 bb_ptr = grid->bb + offset;
1212
1213#if defined GMX_NBNXN_SIMD && GMX_SIMD_REAL_WIDTH4 == 2
1214 if (2*grid->na_cj == grid->na_c)
1215 {
1216 calc_bounding_box_x_x4_halves(na, nbat->x+X4_IND_A(a0)((3*4)*((a0) >> 2) + ((a0) & (4 - 1))), bb_ptr,
1217 grid->bbj+offset*2);
1218 }
1219 else
1220#endif
1221 {
1222 calc_bounding_box_x_x4(na, nbat->x+X4_IND_A(a0)((3*4)*((a0) >> 2) + ((a0) & (4 - 1))), bb_ptr);
1223 }
1224 }
1225 else if (nbat->XFormat == nbatX8)
1226 {
1227 /* Store the bounding boxes as xyz.xyz. */
1228 offset = (a0 - grid->cell0*grid->na_sc) >> grid->na_c_2log;
1229 bb_ptr = grid->bb + offset;
1230
1231 calc_bounding_box_x_x8(na, nbat->x+X8_IND_A(a0)((3*8)*((a0) >> 3) + ((a0) & (8 - 1))), bb_ptr);
1232 }
1233#ifdef NBNXN_BBXXXX
1234 else if (!grid->bSimple)
1235 {
1236 /* Store the bounding boxes in a format convenient
1237 * for SIMD4 calculations: xxxxyyyyzzzz...
1238 */
1239 pbb_ptr =
1240 grid->pbb +
1241 ((a0-grid->cell0*grid->na_sc)>>(grid->na_c_2log+STRIDE_PBB_2LOG2))*NNBSBB_XXXX(2*3*4) +
1242 (((a0-grid->cell0*grid->na_sc)>>grid->na_c_2log) & (STRIDE_PBB4-1));
1243
1244#ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
1245 if (nbat->XFormat == nbatXYZQ)
1246 {
1247 calc_bounding_box_xxxx_simd4(na, nbat->x+a0*nbat->xstride,
1248 bb_work_aligned, pbb_ptr);
1249 }
1250 else
1251#endif
1252 {
1253 calc_bounding_box_xxxx(na, nbat->xstride, nbat->x+a0*nbat->xstride,
1254 pbb_ptr);
1255 }
1256 if (gmx_debug_at)
1257 {
1258 fprintf(debug, "%2d %2d %2d bb %5.2f %5.2f %5.2f %5.2f %5.2f %5.2f\n",
1259 sx, sy, sz,
1260 pbb_ptr[0*STRIDE_PBB4], pbb_ptr[3*STRIDE_PBB4],
1261 pbb_ptr[1*STRIDE_PBB4], pbb_ptr[4*STRIDE_PBB4],
1262 pbb_ptr[2*STRIDE_PBB4], pbb_ptr[5*STRIDE_PBB4]);
1263 }
1264 }
1265#endif
1266 else
1267 {
1268 /* Store the bounding boxes as xyz.xyz. */
1269 bb_ptr = grid->bb+((a0-grid->cell0*grid->na_sc)>>grid->na_c_2log);
1270
1271 calc_bounding_box(na, nbat->xstride, nbat->x+a0*nbat->xstride,
1272 bb_ptr);
1273
1274 if (gmx_debug_at)
1275 {
1276 int bbo;
1277 bbo = (a0 - grid->cell0*grid->na_sc)/grid->na_c;
1278 fprintf(debug, "%2d %2d %2d bb %5.2f %5.2f %5.2f %5.2f %5.2f %5.2f\n",
1279 sx, sy, sz,
1280 grid->bb[bbo].lower[BB_X0],
1281 grid->bb[bbo].lower[BB_Y1],
1282 grid->bb[bbo].lower[BB_Z2],
1283 grid->bb[bbo].upper[BB_X0],
1284 grid->bb[bbo].upper[BB_Y1],
1285 grid->bb[bbo].upper[BB_Z2]);
1286 }
1287 }
1288}
1289
1290/* Spatially sort the atoms within one grid column */
1291static void sort_columns_simple(const nbnxn_search_t nbs,
1292 int dd_zone,
1293 nbnxn_grid_t *grid,
1294 int a0, int a1,
1295 const int *atinfo,
1296 rvec *x,
1297 nbnxn_atomdata_t *nbat,
1298 int cxy_start, int cxy_end,
1299 int *sort_work)
1300{
1301 int cxy;
1302 int cx, cy, cz, ncz, cfilled, c;
1303 int na, ash, ind, a;
1304 int na_c, ash_c;
1305
1306 if (debug)
1307 {
1308 fprintf(debug, "cell0 %d sorting columns %d - %d, atoms %d - %d\n",
1309 grid->cell0, cxy_start, cxy_end, a0, a1);
1310 }
1311
1312 /* Sort the atoms within each x,y column in 3 dimensions */
1313 for (cxy = cxy_start; cxy < cxy_end; cxy++)
1314 {
1315 cx = cxy/grid->ncy;
1316 cy = cxy - cx*grid->ncy;
1317
1318 na = grid->cxy_na[cxy];
1319 ncz = grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy];
1320 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1321
1322 /* Sort the atoms within each x,y column on z coordinate */
1323 sort_atoms(ZZ2, FALSE0, dd_zone,
1324 nbs->a+ash, na, x,
1325 grid->c0[ZZ2],
1326 1.0/nbs->box[ZZ2][ZZ2], ncz*grid->na_sc,
1327 sort_work);
1328
1329 /* Fill the ncz cells in this column */
1330 cfilled = grid->cxy_ind[cxy];
1331 for (cz = 0; cz < ncz; cz++)
1332 {
1333 c = grid->cxy_ind[cxy] + cz;
1334
1335 ash_c = ash + cz*grid->na_sc;
1336 na_c = min(grid->na_sc, na-(ash_c-ash))(((grid->na_sc) < (na-(ash_c-ash))) ? (grid->na_sc) :
(na-(ash_c-ash)) );
1337
1338 fill_cell(nbs, grid, nbat,
1339 ash_c, ash_c+na_c, atinfo, x,
1340 grid->na_sc*cx + (dd_zone >> 2),
1341 grid->na_sc*cy + (dd_zone & 3),
1342 grid->na_sc*cz,
1343 NULL((void*)0));
1344
1345 /* This copy to bbcz is not really necessary.
1346 * But it allows to use the same grid search code
1347 * for the simple and supersub cell setups.
1348 */
1349 if (na_c > 0)
1350 {
1351 cfilled = c;
1352 }
1353 grid->bbcz[c*NNBSBB_D2 ] = grid->bb[cfilled].lower[BB_Z2];
1354 grid->bbcz[c*NNBSBB_D2+1] = grid->bb[cfilled].upper[BB_Z2];
1355 }
1356
1357 /* Set the unused atom indices to -1 */
1358 for (ind = na; ind < ncz*grid->na_sc; ind++)
1359 {
1360 nbs->a[ash+ind] = -1;
1361 }
1362 }
1363}
1364
1365/* Spatially sort the atoms within one grid column */
1366static void sort_columns_supersub(const nbnxn_search_t nbs,
1367 int dd_zone,
1368 nbnxn_grid_t *grid,
1369 int a0, int a1,
1370 const int *atinfo,
1371 rvec *x,
1372 nbnxn_atomdata_t *nbat,
1373 int cxy_start, int cxy_end,
1374 int *sort_work)
1375{
1376 int cxy;
1377 int cx, cy, cz = -1, c = -1, ncz;
1378 int na, ash, na_c, ind, a;
1379 int subdiv_z, sub_z, na_z, ash_z;
1380 int subdiv_y, sub_y, na_y, ash_y;
1381 int subdiv_x, sub_x, na_x, ash_x;
1382
1383 /* cppcheck-suppress unassignedVariable */
1384 nbnxn_bb_t bb_work_array[2], *bb_work_aligned;
1385
1386 bb_work_aligned = (nbnxn_bb_t *)(((size_t)(bb_work_array+1)) & (~((size_t)15)));
1387
1388 if (debug)
1389 {
1390 fprintf(debug, "cell0 %d sorting columns %d - %d, atoms %d - %d\n",
1391 grid->cell0, cxy_start, cxy_end, a0, a1);
1392 }
1393
1394 subdiv_x = grid->na_c;
1395 subdiv_y = GPU_NSUBCELL_X2*subdiv_x;
1396 subdiv_z = GPU_NSUBCELL_Y2*subdiv_y;
1397
1398 /* Sort the atoms within each x,y column in 3 dimensions */
1399 for (cxy = cxy_start; cxy < cxy_end; cxy++)
1400 {
1401 cx = cxy/grid->ncy;
1402 cy = cxy - cx*grid->ncy;
1403
1404 na = grid->cxy_na[cxy];
1405 ncz = grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy];
1406 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1407
1408 /* Sort the atoms within each x,y column on z coordinate */
1409 sort_atoms(ZZ2, FALSE0, dd_zone,
1410 nbs->a+ash, na, x,
1411 grid->c0[ZZ2],
1412 1.0/nbs->box[ZZ2][ZZ2], ncz*grid->na_sc,
1413 sort_work);
1414
1415 /* This loop goes over the supercells and subcells along z at once */
1416 for (sub_z = 0; sub_z < ncz*GPU_NSUBCELL_Z2; sub_z++)
1417 {
1418 ash_z = ash + sub_z*subdiv_z;
1419 na_z = min(subdiv_z, na-(ash_z-ash))(((subdiv_z) < (na-(ash_z-ash))) ? (subdiv_z) : (na-(ash_z
-ash)) );
1420
1421 /* We have already sorted on z */
1422
1423 if (sub_z % GPU_NSUBCELL_Z2 == 0)
1424 {
1425 cz = sub_z/GPU_NSUBCELL_Z2;
1426 c = grid->cxy_ind[cxy] + cz;
1427
1428 /* The number of atoms in this supercell */
1429 na_c = min(grid->na_sc, na-(ash_z-ash))(((grid->na_sc) < (na-(ash_z-ash))) ? (grid->na_sc) :
(na-(ash_z-ash)) );
1430
1431 grid->nsubc[c] = min(GPU_NSUBCELL, (na_c+grid->na_c-1)/grid->na_c)((((2*2*2)) < ((na_c+grid->na_c-1)/grid->na_c)) ? ((
2*2*2)) : ((na_c+grid->na_c-1)/grid->na_c) );
1432
1433 /* Store the z-boundaries of the super cell */
1434 grid->bbcz[c*NNBSBB_D2 ] = x[nbs->a[ash_z]][ZZ2];
1435 grid->bbcz[c*NNBSBB_D2+1] = x[nbs->a[ash_z+na_c-1]][ZZ2];
1436 }
1437
1438#if GPU_NSUBCELL_Y2 > 1
1439 /* Sort the atoms along y */
1440 sort_atoms(YY1, (sub_z & 1), dd_zone,
1441 nbs->a+ash_z, na_z, x,
1442 grid->c0[YY1]+cy*grid->sy,
1443 grid->inv_sy, subdiv_z,
1444 sort_work);
1445#endif
1446
1447 for (sub_y = 0; sub_y < GPU_NSUBCELL_Y2; sub_y++)
1448 {
1449 ash_y = ash_z + sub_y*subdiv_y;
1450 na_y = min(subdiv_y, na-(ash_y-ash))(((subdiv_y) < (na-(ash_y-ash))) ? (subdiv_y) : (na-(ash_y
-ash)) );
1451
1452#if GPU_NSUBCELL_X2 > 1
1453 /* Sort the atoms along x */
1454 sort_atoms(XX0, ((cz*GPU_NSUBCELL_Y2 + sub_y) & 1), dd_zone,
1455 nbs->a+ash_y, na_y, x,
1456 grid->c0[XX0]+cx*grid->sx,
1457 grid->inv_sx, subdiv_y,
1458 sort_work);
1459#endif
1460
1461 for (sub_x = 0; sub_x < GPU_NSUBCELL_X2; sub_x++)
1462 {
1463 ash_x = ash_y + sub_x*subdiv_x;
1464 na_x = min(subdiv_x, na-(ash_x-ash))(((subdiv_x) < (na-(ash_x-ash))) ? (subdiv_x) : (na-(ash_x
-ash)) );
1465
1466 fill_cell(nbs, grid, nbat,
1467 ash_x, ash_x+na_x, atinfo, x,
1468 grid->na_c*(cx*GPU_NSUBCELL_X2+sub_x) + (dd_zone >> 2),
1469 grid->na_c*(cy*GPU_NSUBCELL_Y2+sub_y) + (dd_zone & 3),
1470 grid->na_c*sub_z,
1471 bb_work_aligned);
1472 }
1473 }
1474 }
1475
1476 /* Set the unused atom indices to -1 */
1477 for (ind = na; ind < ncz*grid->na_sc; ind++)
1478 {
1479 nbs->a[ash+ind] = -1;
1480 }
1481 }
1482}
1483
1484/* Determine in which grid column atoms should go */
1485static void calc_column_indices(nbnxn_grid_t *grid,
1486 int a0, int a1,
1487 rvec *x,
1488 int dd_zone, const int *move,
1489 int thread, int nthread,
1490 int *cell,
1491 int *cxy_na)
1492{
1493 int n0, n1, i;
1494 int cx, cy;
1495
1496 /* We add one extra cell for particles which moved during DD */
1497 for (i = 0; i < grid->ncx*grid->ncy+1; i++)
1498 {
1499 cxy_na[i] = 0;
1500 }
1501
1502 n0 = a0 + (int)((thread+0)*(a1 - a0))/nthread;
1503 n1 = a0 + (int)((thread+1)*(a1 - a0))/nthread;
1504 if (dd_zone == 0)
1505 {
1506 /* Home zone */
1507 for (i = n0; i < n1; i++)
1508 {
1509 if (move == NULL((void*)0) || move[i] >= 0)
1510 {
1511 /* We need to be careful with rounding,
1512 * particles might be a few bits outside the local zone.
1513 * The int cast takes care of the lower bound,
1514 * we will explicitly take care of the upper bound.
1515 */
1516 cx = (int)((x[i][XX0] - grid->c0[XX0])*grid->inv_sx);
1517 cy = (int)((x[i][YY1] - grid->c0[YY1])*grid->inv_sy);
1518
1519#ifndef NDEBUG1
1520 if (cx < 0 || cx > grid->ncx ||
1521 cy < 0 || cy > grid->ncy)
1522 {
1523 gmx_fatal(FARGS0, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 1523,
1524 "grid cell cx %d cy %d out of range (max %d %d)\n"
1525 "atom %f %f %f, grid->c0 %f %f",
1526 cx, cy, grid->ncx, grid->ncy,
1527 x[i][XX0], x[i][YY1], x[i][ZZ2], grid->c0[XX0], grid->c0[YY1]);
1528 }
1529#endif
1530 /* Take care of potential rouding issues */
1531 cx = min(cx, grid->ncx - 1)(((cx) < (grid->ncx - 1)) ? (cx) : (grid->ncx - 1) );
1532 cy = min(cy, grid->ncy - 1)(((cy) < (grid->ncy - 1)) ? (cy) : (grid->ncy - 1) );
1533
1534 /* For the moment cell will contain only the, grid local,
1535 * x and y indices, not z.
1536 */
1537 cell[i] = cx*grid->ncy + cy;
1538 }
1539 else
1540 {
1541 /* Put this moved particle after the end of the grid,
1542 * so we can process it later without using conditionals.
1543 */
1544 cell[i] = grid->ncx*grid->ncy;
1545 }
1546
1547 cxy_na[cell[i]]++;
1548 }
1549 }
1550 else
1551 {
1552 /* Non-home zone */
1553 for (i = n0; i < n1; i++)
1554 {
1555 cx = (int)((x[i][XX0] - grid->c0[XX0])*grid->inv_sx);
1556 cy = (int)((x[i][YY1] - grid->c0[YY1])*grid->inv_sy);
1557
1558 /* For non-home zones there could be particles outside
1559 * the non-bonded cut-off range, which have been communicated
1560 * for bonded interactions only. For the result it doesn't
1561 * matter where these end up on the grid. For performance
1562 * we put them in an extra row at the border.
1563 */
1564 cx = max(cx, 0)(((cx) > (0)) ? (cx) : (0) );
1565 cx = min(cx, grid->ncx - 1)(((cx) < (grid->ncx - 1)) ? (cx) : (grid->ncx - 1) );
1566 cy = max(cy, 0)(((cy) > (0)) ? (cy) : (0) );
1567 cy = min(cy, grid->ncy - 1)(((cy) < (grid->ncy - 1)) ? (cy) : (grid->ncy - 1) );
1568
1569 /* For the moment cell will contain only the, grid local,
1570 * x and y indices, not z.
1571 */
1572 cell[i] = cx*grid->ncy + cy;
1573
1574 cxy_na[cell[i]]++;
1575 }
1576 }
1577}
1578
1579/* Determine in which grid cells the atoms should go */
1580static void calc_cell_indices(const nbnxn_search_t nbs,
1581 int dd_zone,
1582 nbnxn_grid_t *grid,
1583 int a0, int a1,
1584 const int *atinfo,
1585 rvec *x,
1586 const int *move,
1587 nbnxn_atomdata_t *nbat)
1588{
1589 int n0, n1, i;
1590 int cx, cy, cxy, ncz_max, ncz;
1591 int nthread, thread;
1592 int *cxy_na, cxy_na_i;
1593
1594 nthread = gmx_omp_nthreads_get(emntPairsearch);
1595
1596#pragma omp parallel for num_threads(nthread) schedule(static)
1597 for (thread = 0; thread < nthread; thread++)
1598 {
1599 calc_column_indices(grid, a0, a1, x, dd_zone, move, thread, nthread,
1600 nbs->cell, nbs->work[thread].cxy_na);
1601 }
1602
1603 /* Make the cell index as a function of x and y */
1604 ncz_max = 0;
1605 ncz = 0;
1606 grid->cxy_ind[0] = 0;
1607 for (i = 0; i < grid->ncx*grid->ncy+1; i++)
1608 {
1609 /* We set ncz_max at the beginning of the loop iso at the end
1610 * to skip i=grid->ncx*grid->ncy which are moved particles
1611 * that do not need to be ordered on the grid.
1612 */
1613 if (ncz > ncz_max)
1614 {
1615 ncz_max = ncz;
1616 }
1617 cxy_na_i = nbs->work[0].cxy_na[i];
1618 for (thread = 1; thread < nthread; thread++)
1619 {
1620 cxy_na_i += nbs->work[thread].cxy_na[i];
1621 }
1622 ncz = (cxy_na_i + grid->na_sc - 1)/grid->na_sc;
1623 if (nbat->XFormat == nbatX8)
1624 {
1625 /* Make the number of cell a multiple of 2 */
1626 ncz = (ncz + 1) & ~1;
1627 }
1628 grid->cxy_ind[i+1] = grid->cxy_ind[i] + ncz;
1629 /* Clear cxy_na, so we can reuse the array below */
1630 grid->cxy_na[i] = 0;
1631 }
1632 grid->nc = grid->cxy_ind[grid->ncx*grid->ncy] - grid->cxy_ind[0];
1633
1634 nbat->natoms = (grid->cell0 + grid->nc)*grid->na_sc;
1635
1636 if (debug)
1637 {
1638 fprintf(debug, "ns na_sc %d na_c %d super-cells: %d x %d y %d z %.1f maxz %d\n",
1639 grid->na_sc, grid->na_c, grid->nc,
1640 grid->ncx, grid->ncy, grid->nc/((double)(grid->ncx*grid->ncy)),
1641 ncz_max);
1642 if (gmx_debug_at)
1643 {
1644 i = 0;
1645 for (cy = 0; cy < grid->ncy; cy++)
1646 {
1647 for (cx = 0; cx < grid->ncx; cx++)
1648 {
1649 fprintf(debug, " %2d", grid->cxy_ind[i+1]-grid->cxy_ind[i]);
1650 i++;
1651 }
1652 fprintf(debug, "\n");
1653 }
1654 }
1655 }
1656
1657 /* Make sure the work array for sorting is large enough */
1658 if (ncz_max*grid->na_sc*SGSF(4 + 1) > nbs->work[0].sort_work_nalloc)
1659 {
1660 for (thread = 0; thread < nbs->nthread_max; thread++)
1661 {
1662 nbs->work[thread].sort_work_nalloc =
1663 over_alloc_large(ncz_max*grid->na_sc*SGSF)(int)(1.19*(ncz_max*grid->na_sc*(4 + 1)) + 1000);
1664 srenew(nbs->work[thread].sort_work,(nbs->work[thread].sort_work) = save_realloc("nbs->work[thread].sort_work"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 1665, (nbs->work[thread].sort_work), (nbs->work[thread
].sort_work_nalloc), sizeof(*(nbs->work[thread].sort_work)
))
1665 nbs->work[thread].sort_work_nalloc)(nbs->work[thread].sort_work) = save_realloc("nbs->work[thread].sort_work"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 1665, (nbs->work[thread].sort_work), (nbs->work[thread
].sort_work_nalloc), sizeof(*(nbs->work[thread].sort_work)
));
1666 /* When not in use, all elements should be -1 */
1667 for (i = 0; i < nbs->work[thread].sort_work_nalloc; i++)
1668 {
1669 nbs->work[thread].sort_work[i] = -1;
1670 }
1671 }
1672 }
1673
1674 /* Now we know the dimensions we can fill the grid.
1675 * This is the first, unsorted fill. We sort the columns after this.
1676 */
1677 for (i = a0; i < a1; i++)
1678 {
1679 /* At this point nbs->cell contains the local grid x,y indices */
1680 cxy = nbs->cell[i];
1681 nbs->a[(grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc + grid->cxy_na[cxy]++] = i;
1682 }
1683
1684 if (dd_zone == 0)
1685 {
1686 /* Set the cell indices for the moved particles */
1687 n0 = grid->nc*grid->na_sc;
1688 n1 = grid->nc*grid->na_sc+grid->cxy_na[grid->ncx*grid->ncy];
1689 if (dd_zone == 0)
1690 {
1691 for (i = n0; i < n1; i++)
1692 {
1693 nbs->cell[nbs->a[i]] = i;
1694 }
1695 }
1696 }
1697
1698 /* Sort the super-cell columns along z into the sub-cells. */
1699#pragma omp parallel for num_threads(nbs->nthread_max) schedule(static)
1700 for (thread = 0; thread < nbs->nthread_max; thread++)
1701 {
1702 if (grid->bSimple)
1703 {
1704 sort_columns_simple(nbs, dd_zone, grid, a0, a1, atinfo, x, nbat,
1705 ((thread+0)*grid->ncx*grid->ncy)/nthread,
1706 ((thread+1)*grid->ncx*grid->ncy)/nthread,
1707 nbs->work[thread].sort_work);
1708 }
1709 else
1710 {
1711 sort_columns_supersub(nbs, dd_zone, grid, a0, a1, atinfo, x, nbat,
1712 ((thread+0)*grid->ncx*grid->ncy)/nthread,
1713 ((thread+1)*grid->ncx*grid->ncy)/nthread,
1714 nbs->work[thread].sort_work);
1715 }
1716 }
1717
1718 if (grid->bSimple && nbat->XFormat == nbatX8)
1719 {
1720 combine_bounding_box_pairs(grid, grid->bb);
1721 }
1722
1723 if (!grid->bSimple)
1724 {
1725 grid->nsubc_tot = 0;
1726 for (i = 0; i < grid->nc; i++)
1727 {
1728 grid->nsubc_tot += grid->nsubc[i];
1729 }
1730 }
1731
1732 if (debug)
1733 {
1734 if (grid->bSimple)
1735 {
1736 print_bbsizes_simple(debug, nbs, grid);
1737 }
1738 else
1739 {
1740 fprintf(debug, "ns non-zero sub-cells: %d average atoms %.2f\n",
1741 grid->nsubc_tot, (a1-a0)/(double)grid->nsubc_tot);
1742
1743 print_bbsizes_supersub(debug, nbs, grid);
1744 }
1745 }
1746}
1747
1748static void init_buffer_flags(nbnxn_buffer_flags_t *flags,
1749 int natoms)
1750{
1751 int b;
1752
1753 flags->nflag = (natoms + NBNXN_BUFFERFLAG_SIZE16 - 1)/NBNXN_BUFFERFLAG_SIZE16;
1754 if (flags->nflag > flags->flag_nalloc)
1755 {
1756 flags->flag_nalloc = over_alloc_large(flags->nflag)(int)(1.19*(flags->nflag) + 1000);
1757 srenew(flags->flag, flags->flag_nalloc)(flags->flag) = save_realloc("flags->flag", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 1757, (flags->flag), (flags->flag_nalloc), sizeof(*(flags
->flag)));
1758 }
1759 for (b = 0; b < flags->nflag; b++)
1760 {
1761 flags->flag[b] = 0;
1762 }
1763}
1764
1765/* Sets up a grid and puts the atoms on the grid.
1766 * This function only operates on one domain of the domain decompostion.
1767 * Note that without domain decomposition there is only one domain.
1768 */
1769void nbnxn_put_on_grid(nbnxn_search_t nbs,
1770 int ePBC, matrix box,
1771 int dd_zone,
1772 rvec corner0, rvec corner1,
1773 int a0, int a1,
1774 real atom_density,
1775 const int *atinfo,
1776 rvec *x,
1777 int nmoved, int *move,
1778 int nb_kernel_type,
1779 nbnxn_atomdata_t *nbat)
1780{
1781 nbnxn_grid_t *grid;
1782 int n;
1783 int nc_max_grid, nc_max;
1784
1785 grid = &nbs->grid[dd_zone];
1786
1787 nbs_cycle_start(&nbs->cc[enbsCCgrid]);
1788
1789 grid->bSimple = nbnxn_kernel_pairlist_simple(nb_kernel_type);
1790
1791 grid->na_c = nbnxn_kernel_to_ci_size(nb_kernel_type);
1792 grid->na_cj = nbnxn_kernel_to_cj_size(nb_kernel_type);
1793 grid->na_sc = (grid->bSimple ? 1 : GPU_NSUBCELL(2*2*2))*grid->na_c;
1794 grid->na_c_2log = get_2log(grid->na_c);
1795
1796 nbat->na_c = grid->na_c;
1797
1798 if (dd_zone == 0)
1799 {
1800 grid->cell0 = 0;
1801 }
1802 else
1803 {
1804 grid->cell0 =
1805 (nbs->grid[dd_zone-1].cell0 + nbs->grid[dd_zone-1].nc)*
1806 nbs->grid[dd_zone-1].na_sc/grid->na_sc;
1807 }
1808
1809 n = a1 - a0;
1810
1811 if (dd_zone == 0)
1812 {
1813 nbs->ePBC = ePBC;
1814 copy_mat(box, nbs->box);
1815
1816 if (atom_density >= 0)
1817 {
1818 grid->atom_density = atom_density;
1819 }
1820 else
1821 {
1822 grid->atom_density = grid_atom_density(n-nmoved, corner0, corner1);
1823 }
1824
1825 grid->cell0 = 0;
1826
1827 nbs->natoms_local = a1 - nmoved;
1828 /* We assume that nbnxn_put_on_grid is called first
1829 * for the local atoms (dd_zone=0).
1830 */
1831 nbs->natoms_nonlocal = a1 - nmoved;
1832 }
1833 else
1834 {
1835 nbs->natoms_nonlocal = max(nbs->natoms_nonlocal, a1)(((nbs->natoms_nonlocal) > (a1)) ? (nbs->natoms_nonlocal
) : (a1) );
1836 }
1837
1838 nc_max_grid = set_grid_size_xy(nbs, grid,
1839 dd_zone, n-nmoved, corner0, corner1,
1840 nbs->grid[0].atom_density);
1841
1842 nc_max = grid->cell0 + nc_max_grid;
1843
1844 if (a1 > nbs->cell_nalloc)
1845 {
1846 nbs->cell_nalloc = over_alloc_large(a1)(int)(1.19*(a1) + 1000);
1847 srenew(nbs->cell, nbs->cell_nalloc)(nbs->cell) = save_realloc("nbs->cell", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 1847, (nbs->cell), (nbs->cell_nalloc), sizeof(*(nbs->
cell)));
1848 }
1849
1850 /* To avoid conditionals we store the moved particles at the end of a,
1851 * make sure we have enough space.
1852 */
1853 if (nc_max*grid->na_sc + nmoved > nbs->a_nalloc)
1854 {
1855 nbs->a_nalloc = over_alloc_large(nc_max*grid->na_sc + nmoved)(int)(1.19*(nc_max*grid->na_sc + nmoved) + 1000);
1856 srenew(nbs->a, nbs->a_nalloc)(nbs->a) = save_realloc("nbs->a", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 1856, (nbs->a), (nbs->a_nalloc), sizeof(*(nbs->a))
);
1857 }
1858
1859 /* We need padding up to a multiple of the buffer flag size: simply add */
1860 if (nc_max*grid->na_sc + NBNXN_BUFFERFLAG_SIZE16 > nbat->nalloc)
1861 {
1862 nbnxn_atomdata_realloc(nbat, nc_max*grid->na_sc+NBNXN_BUFFERFLAG_SIZE16);
1863 }
1864
1865 calc_cell_indices(nbs, dd_zone, grid, a0, a1, atinfo, x, move, nbat);
1866
1867 if (dd_zone == 0)
1868 {
1869 nbat->natoms_local = nbat->natoms;
1870 }
1871
1872 nbs_cycle_stop(&nbs->cc[enbsCCgrid]);
1873}
1874
1875/* Calls nbnxn_put_on_grid for all non-local domains */
1876void nbnxn_put_on_grid_nonlocal(nbnxn_search_t nbs,
1877 const gmx_domdec_zones_t *zones,
1878 const int *atinfo,
1879 rvec *x,
1880 int nb_kernel_type,
1881 nbnxn_atomdata_t *nbat)
1882{
1883 int zone, d;
1884 rvec c0, c1;
1885
1886 for (zone = 1; zone < zones->n; zone++)
1887 {
1888 for (d = 0; d < DIM3; d++)
1889 {
1890 c0[d] = zones->size[zone].bb_x0[d];
1891 c1[d] = zones->size[zone].bb_x1[d];
1892 }
1893
1894 nbnxn_put_on_grid(nbs, nbs->ePBC, NULL((void*)0),
1895 zone, c0, c1,
1896 zones->cg_range[zone],
1897 zones->cg_range[zone+1],
1898 -1,
1899 atinfo,
1900 x,
1901 0, NULL((void*)0),
1902 nb_kernel_type,
1903 nbat);
1904 }
1905}
1906
1907/* Add simple grid type information to the local super/sub grid */
1908void nbnxn_grid_add_simple(nbnxn_search_t nbs,
1909 nbnxn_atomdata_t *nbat)
1910{
1911 nbnxn_grid_t *grid;
1912 float *bbcz;
1913 nbnxn_bb_t *bb;
1914 int ncd, sc;
1915
1916 grid = &nbs->grid[0];
1917
1918 if (grid->bSimple)
1919 {
1920 gmx_incons("nbnxn_grid_simple called with a simple grid")_gmx_error("incons", "nbnxn_grid_simple called with a simple grid"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 1920);
1921 }
1922
1923 ncd = grid->na_sc/NBNXN_CPU_CLUSTER_I_SIZE4;
1924
1925 if (grid->nc*ncd > grid->nc_nalloc_simple)
1926 {
1927 grid->nc_nalloc_simple = over_alloc_large(grid->nc*ncd)(int)(1.19*(grid->nc*ncd) + 1000);
1928 srenew(grid->bbcz_simple, grid->nc_nalloc_simple*NNBSBB_D)(grid->bbcz_simple) = save_realloc("grid->bbcz_simple",
"/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 1928, (grid->bbcz_simple), (grid->nc_nalloc_simple*2)
, sizeof(*(grid->bbcz_simple)));
1929 srenew(grid->bb_simple, grid->nc_nalloc_simple)(grid->bb_simple) = save_realloc("grid->bb_simple", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 1929, (grid->bb_simple), (grid->nc_nalloc_simple), sizeof
(*(grid->bb_simple)));
1930 srenew(grid->flags_simple, grid->nc_nalloc_simple)(grid->flags_simple) = save_realloc("grid->flags_simple"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 1930, (grid->flags_simple), (grid->nc_nalloc_simple),
sizeof(*(grid->flags_simple)));
1931 if (nbat->XFormat)
1932 {
1933 sfree_aligned(grid->bbj)save_free_aligned("grid->bbj", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 1933, (grid->bbj));
1934 snew_aligned(grid->bbj, grid->nc_nalloc_simple/2, 16)(grid->bbj) = save_calloc_aligned("grid->bbj", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 1934, (grid->nc_nalloc_simple/2), sizeof(*(grid->bbj)
), 16);
1935 }
1936 }
1937
1938 bbcz = grid->bbcz_simple;
1939 bb = grid->bb_simple;
1940
1941#pragma omp parallel for num_threads(gmx_omp_nthreads_get(emntPairsearch)) schedule(static)
1942 for (sc = 0; sc < grid->nc; sc++)
1943 {
1944 int c, tx, na;
1945
1946 for (c = 0; c < ncd; c++)
1947 {
1948 tx = sc*ncd + c;
1949
1950 na = NBNXN_CPU_CLUSTER_I_SIZE4;
1951 while (na > 0 &&
1952 nbat->type[tx*NBNXN_CPU_CLUSTER_I_SIZE4+na-1] == nbat->ntype-1)
1953 {
1954 na--;
1955 }
1956
1957 if (na > 0)
1958 {
1959 switch (nbat->XFormat)
1960 {
1961 case nbatX4:
1962 /* PACK_X4==NBNXN_CPU_CLUSTER_I_SIZE, so this is simple */
1963 calc_bounding_box_x_x4(na, nbat->x+tx*STRIDE_P4(3*4),
1964 bb+tx);
1965 break;
1966 case nbatX8:
1967 /* PACK_X8>NBNXN_CPU_CLUSTER_I_SIZE, more complicated */
1968 calc_bounding_box_x_x8(na, nbat->x+X8_IND_A(tx*NBNXN_CPU_CLUSTER_I_SIZE)((3*8)*((tx*4) >> 3) + ((tx*4) & (8 - 1))),
1969 bb+tx);
1970 break;
1971 default:
1972 calc_bounding_box(na, nbat->xstride,
1973 nbat->x+tx*NBNXN_CPU_CLUSTER_I_SIZE4*nbat->xstride,
1974 bb+tx);
1975 break;
1976 }
1977 bbcz[tx*NNBSBB_D2+0] = bb[tx].lower[BB_Z2];
1978 bbcz[tx*NNBSBB_D2+1] = bb[tx].upper[BB_Z2];
1979
1980 /* No interaction optimization yet here */
1981 grid->flags_simple[tx] = NBNXN_CI_DO_LJ(0)(1<<(7+3*(0))) | NBNXN_CI_DO_COUL(0)(1<<(9+3*(0)));
1982 }
1983 else
1984 {
1985 grid->flags_simple[tx] = 0;
1986 }
1987 }
1988 }
1989
1990 if (grid->bSimple && nbat->XFormat == nbatX8)
1991 {
1992 combine_bounding_box_pairs(grid, grid->bb_simple);
1993 }
1994}
1995
1996void nbnxn_get_ncells(nbnxn_search_t nbs, int *ncx, int *ncy)
1997{
1998 *ncx = nbs->grid[0].ncx;
1999 *ncy = nbs->grid[0].ncy;
2000}
2001
2002void nbnxn_get_atomorder(nbnxn_search_t nbs, int **a, int *n)
2003{
2004 const nbnxn_grid_t *grid;
2005
2006 grid = &nbs->grid[0];
2007
2008 /* Return the atom order for the home cell (index 0) */
2009 *a = nbs->a;
2010
2011 *n = grid->cxy_ind[grid->ncx*grid->ncy]*grid->na_sc;
2012}
2013
2014void nbnxn_set_atomorder(nbnxn_search_t nbs)
2015{
2016 nbnxn_grid_t *grid;
2017 int ao, cx, cy, cxy, cz, j;
2018
2019 /* Set the atom order for the home cell (index 0) */
2020 grid = &nbs->grid[0];
2021
2022 ao = 0;
2023 for (cx = 0; cx < grid->ncx; cx++)
2024 {
2025 for (cy = 0; cy < grid->ncy; cy++)
2026 {
2027 cxy = cx*grid->ncy + cy;
2028 j = grid->cxy_ind[cxy]*grid->na_sc;
2029 for (cz = 0; cz < grid->cxy_na[cxy]; cz++)
2030 {
2031 nbs->a[j] = ao;
2032 nbs->cell[ao] = j;
2033 ao++;
2034 j++;
2035 }
2036 }
2037 }
2038}
2039
2040/* Determines the cell range along one dimension that
2041 * the bounding box b0 - b1 sees.
2042 */
2043static void get_cell_range(real b0, real b1,
2044 int nc, real c0, real s, real invs,
2045 real d2, real r2, int *cf, int *cl)
2046{
2047 *cf = max((int)((b0 - c0)*invs), 0)((((int)((b0 - c0)*invs)) > (0)) ? ((int)((b0 - c0)*invs))
: (0) );
2048
2049 while (*cf > 0 && d2 + sqr((b0 - c0) - (*cf-1+1)*s) < r2)
2050 {
2051 (*cf)--;
2052 }
2053
2054 *cl = min((int)((b1 - c0)*invs), nc-1)((((int)((b1 - c0)*invs)) < (nc-1)) ? ((int)((b1 - c0)*invs
)) : (nc-1) );
2055 while (*cl < nc-1 && d2 + sqr((*cl+1)*s - (b1 - c0)) < r2)
2056 {
2057 (*cl)++;
2058 }
2059}
2060
2061/* Reference code calculating the distance^2 between two bounding boxes */
2062static float box_dist2(float bx0, float bx1, float by0,
2063 float by1, float bz0, float bz1,
2064 const nbnxn_bb_t *bb)
2065{
2066 float d2;
2067 float dl, dh, dm, dm0;
2068
2069 d2 = 0;
2070
2071 dl = bx0 - bb->upper[BB_X0];
2072 dh = bb->lower[BB_X0] - bx1;
2073 dm = max(dl, dh)(((dl) > (dh)) ? (dl) : (dh) );
2074 dm0 = max(dm, 0)(((dm) > (0)) ? (dm) : (0) );
2075 d2 += dm0*dm0;
2076
2077 dl = by0 - bb->upper[BB_Y1];
2078 dh = bb->lower[BB_Y1] - by1;
2079 dm = max(dl, dh)(((dl) > (dh)) ? (dl) : (dh) );
2080 dm0 = max(dm, 0)(((dm) > (0)) ? (dm) : (0) );
2081 d2 += dm0*dm0;
2082
2083 dl = bz0 - bb->upper[BB_Z2];
2084 dh = bb->lower[BB_Z2] - bz1;
2085 dm = max(dl, dh)(((dl) > (dh)) ? (dl) : (dh) );
2086 dm0 = max(dm, 0)(((dm) > (0)) ? (dm) : (0) );
2087 d2 += dm0*dm0;
2088
2089 return d2;
2090}
2091
2092/* Plain C code calculating the distance^2 between two bounding boxes */
2093static float subc_bb_dist2(int si, const nbnxn_bb_t *bb_i_ci,
2094 int csj, const nbnxn_bb_t *bb_j_all)
2095{
2096 const nbnxn_bb_t *bb_i, *bb_j;
2097 float d2;
2098 float dl, dh, dm, dm0;
2099
2100 bb_i = bb_i_ci + si;
2101 bb_j = bb_j_all + csj;
2102
2103 d2 = 0;
2104
2105 dl = bb_i->lower[BB_X0] - bb_j->upper[BB_X0];
2106 dh = bb_j->lower[BB_X0] - bb_i->upper[BB_X0];
2107 dm = max(dl, dh)(((dl) > (dh)) ? (dl) : (dh) );
2108 dm0 = max(dm, 0)(((dm) > (0)) ? (dm) : (0) );
2109 d2 += dm0*dm0;
2110
2111 dl = bb_i->lower[BB_Y1] - bb_j->upper[BB_Y1];
2112 dh = bb_j->lower[BB_Y1] - bb_i->upper[BB_Y1];
2113 dm = max(dl, dh)(((dl) > (dh)) ? (dl) : (dh) );
2114 dm0 = max(dm, 0)(((dm) > (0)) ? (dm) : (0) );
2115 d2 += dm0*dm0;
2116
2117 dl = bb_i->lower[BB_Z2] - bb_j->upper[BB_Z2];
2118 dh = bb_j->lower[BB_Z2] - bb_i->upper[BB_Z2];
2119 dm = max(dl, dh)(((dl) > (dh)) ? (dl) : (dh) );
2120 dm0 = max(dm, 0)(((dm) > (0)) ? (dm) : (0) );
2121 d2 += dm0*dm0;
2122
2123 return d2;
2124}
2125
2126#ifdef NBNXN_SEARCH_BB_SIMD4
2127
2128/* 4-wide SIMD code for bb distance for bb format xyz0 */
2129static float subc_bb_dist2_simd4(int si, const nbnxn_bb_t *bb_i_ci,
2130 int csj, const nbnxn_bb_t *bb_j_all)
2131{
2132 gmx_simd4_float_t__m128 bb_i_S0, bb_i_S1;
2133 gmx_simd4_float_t__m128 bb_j_S0, bb_j_S1;
2134 gmx_simd4_float_t__m128 dl_S;
2135 gmx_simd4_float_t__m128 dh_S;
2136 gmx_simd4_float_t__m128 dm_S;
2137 gmx_simd4_float_t__m128 dm0_S;
2138
2139 bb_i_S0 = gmx_simd4_load_f_mm_load_ps(&bb_i_ci[si].lower[0]);
2140 bb_i_S1 = gmx_simd4_load_f_mm_load_ps(&bb_i_ci[si].upper[0]);
2141 bb_j_S0 = gmx_simd4_load_f_mm_load_ps(&bb_j_all[csj].lower[0]);
2142 bb_j_S1 = gmx_simd4_load_f_mm_load_ps(&bb_j_all[csj].upper[0]);
2143
2144 dl_S = gmx_simd4_sub_f_mm_sub_ps(bb_i_S0, bb_j_S1);
2145 dh_S = gmx_simd4_sub_f_mm_sub_ps(bb_j_S0, bb_i_S1);
2146
2147 dm_S = gmx_simd4_max_f_mm_max_ps(dl_S, dh_S);
2148 dm0_S = gmx_simd4_max_f_mm_max_ps(dm_S, gmx_simd4_setzero_f_mm_setzero_ps());
2149
2150 return gmx_simd4_dotproduct3_fgmx_simd4_dotproduct3_f_sse4_1(dm0_S, dm0_S);
2151}
2152
2153/* Calculate bb bounding distances of bb_i[si,...,si+3] and store them in d2 */
2154#define SUBC_BB_DIST2_SIMD4_XXXX_INNER(si, bb_i, d2){ int shi; __m128 dx_0, dy_0, dz_0; __m128 dx_1, dy_1, dz_1; __m128
mx, my, mz; __m128 m0x, m0y, m0z; __m128 d2x, d2y, d2z; __m128
d2s, d2t; shi = si*2*3; xi_l = _mm_load_ps(bb_i+shi+0*4); yi_l
= _mm_load_ps(bb_i+shi+1*4); zi_l = _mm_load_ps(bb_i+shi+2*4
); xi_h = _mm_load_ps(bb_i+shi+3*4); yi_h = _mm_load_ps(bb_i+
shi+4*4); zi_h = _mm_load_ps(bb_i+shi+5*4); dx_0 = _mm_sub_ps
(xi_l, xj_h); dy_0 = _mm_sub_ps(yi_l, yj_h); dz_0 = _mm_sub_ps
(zi_l, zj_h); dx_1 = _mm_sub_ps(xj_l, xi_h); dy_1 = _mm_sub_ps
(yj_l, yi_h); dz_1 = _mm_sub_ps(zj_l, zi_h); mx = _mm_max_ps(
dx_0, dx_1); my = _mm_max_ps(dy_0, dy_1); mz = _mm_max_ps(dz_0
, dz_1); m0x = _mm_max_ps(mx, zero); m0y = _mm_max_ps(my, zero
); m0z = _mm_max_ps(mz, zero); d2x = _mm_mul_ps(m0x, m0x); d2y
= _mm_mul_ps(m0y, m0y); d2z = _mm_mul_ps(m0z, m0z); d2s = _mm_add_ps
(d2x, d2y); d2t = _mm_add_ps(d2s, d2z); _mm_store_ps(d2+si, d2t
); } \
2155 { \
2156 int shi; \
2157 \
2158 gmx_simd4_float_t__m128 dx_0, dy_0, dz_0; \
2159 gmx_simd4_float_t__m128 dx_1, dy_1, dz_1; \
2160 \
2161 gmx_simd4_float_t__m128 mx, my, mz; \
2162 gmx_simd4_float_t__m128 m0x, m0y, m0z; \
2163 \
2164 gmx_simd4_float_t__m128 d2x, d2y, d2z; \
2165 gmx_simd4_float_t__m128 d2s, d2t; \
2166 \
2167 shi = si*NNBSBB_D2*DIM3; \
2168 \
2169 xi_l = gmx_simd4_load_f_mm_load_ps(bb_i+shi+0*STRIDE_PBB4); \
2170 yi_l = gmx_simd4_load_f_mm_load_ps(bb_i+shi+1*STRIDE_PBB4); \
2171 zi_l = gmx_simd4_load_f_mm_load_ps(bb_i+shi+2*STRIDE_PBB4); \
2172 xi_h = gmx_simd4_load_f_mm_load_ps(bb_i+shi+3*STRIDE_PBB4); \
2173 yi_h = gmx_simd4_load_f_mm_load_ps(bb_i+shi+4*STRIDE_PBB4); \
2174 zi_h = gmx_simd4_load_f_mm_load_ps(bb_i+shi+5*STRIDE_PBB4); \
2175 \
2176 dx_0 = gmx_simd4_sub_f_mm_sub_ps(xi_l, xj_h); \
2177 dy_0 = gmx_simd4_sub_f_mm_sub_ps(yi_l, yj_h); \
2178 dz_0 = gmx_simd4_sub_f_mm_sub_ps(zi_l, zj_h); \
2179 \
2180 dx_1 = gmx_simd4_sub_f_mm_sub_ps(xj_l, xi_h); \
2181 dy_1 = gmx_simd4_sub_f_mm_sub_ps(yj_l, yi_h); \
2182 dz_1 = gmx_simd4_sub_f_mm_sub_ps(zj_l, zi_h); \
2183 \
2184 mx = gmx_simd4_max_f_mm_max_ps(dx_0, dx_1); \
2185 my = gmx_simd4_max_f_mm_max_ps(dy_0, dy_1); \
2186 mz = gmx_simd4_max_f_mm_max_ps(dz_0, dz_1); \
2187 \
2188 m0x = gmx_simd4_max_f_mm_max_ps(mx, zero); \
2189 m0y = gmx_simd4_max_f_mm_max_ps(my, zero); \
2190 m0z = gmx_simd4_max_f_mm_max_ps(mz, zero); \
2191 \
2192 d2x = gmx_simd4_mul_f_mm_mul_ps(m0x, m0x); \
2193 d2y = gmx_simd4_mul_f_mm_mul_ps(m0y, m0y); \
2194 d2z = gmx_simd4_mul_f_mm_mul_ps(m0z, m0z); \
2195 \
2196 d2s = gmx_simd4_add_f_mm_add_ps(d2x, d2y); \
2197 d2t = gmx_simd4_add_f_mm_add_ps(d2s, d2z); \
2198 \
2199 gmx_simd4_store_f_mm_store_ps(d2+si, d2t); \
2200 }
2201
2202/* 4-wide SIMD code for nsi bb distances for bb format xxxxyyyyzzzz */
2203static void subc_bb_dist2_simd4_xxxx(const float *bb_j,
2204 int nsi, const float *bb_i,
2205 float *d2)
2206{
2207 gmx_simd4_float_t__m128 xj_l, yj_l, zj_l;
2208 gmx_simd4_float_t__m128 xj_h, yj_h, zj_h;
2209 gmx_simd4_float_t__m128 xi_l, yi_l, zi_l;
2210 gmx_simd4_float_t__m128 xi_h, yi_h, zi_h;
2211
2212 gmx_simd4_float_t__m128 zero;
2213
2214 zero = gmx_simd4_setzero_f_mm_setzero_ps();
2215
2216 xj_l = gmx_simd4_set1_f_mm_set1_ps(bb_j[0*STRIDE_PBB4]);
2217 yj_l = gmx_simd4_set1_f_mm_set1_ps(bb_j[1*STRIDE_PBB4]);
2218 zj_l = gmx_simd4_set1_f_mm_set1_ps(bb_j[2*STRIDE_PBB4]);
2219 xj_h = gmx_simd4_set1_f_mm_set1_ps(bb_j[3*STRIDE_PBB4]);
2220 yj_h = gmx_simd4_set1_f_mm_set1_ps(bb_j[4*STRIDE_PBB4]);
2221 zj_h = gmx_simd4_set1_f_mm_set1_ps(bb_j[5*STRIDE_PBB4]);
2222
2223 /* Here we "loop" over si (0,STRIDE_PBB) from 0 to nsi with step STRIDE_PBB.
2224 * But as we know the number of iterations is 1 or 2, we unroll manually.
2225 */
2226 SUBC_BB_DIST2_SIMD4_XXXX_INNER(0, bb_i, d2){ int shi; __m128 dx_0, dy_0, dz_0; __m128 dx_1, dy_1, dz_1; __m128
mx, my, mz; __m128 m0x, m0y, m0z; __m128 d2x, d2y, d2z; __m128
d2s, d2t; shi = 0*2*3; xi_l = _mm_load_ps(bb_i+shi+0*4); yi_l
= _mm_load_ps(bb_i+shi+1*4); zi_l = _mm_load_ps(bb_i+shi+2*4
); xi_h = _mm_load_ps(bb_i+shi+3*4); yi_h = _mm_load_ps(bb_i+
shi+4*4); zi_h = _mm_load_ps(bb_i+shi+5*4); dx_0 = _mm_sub_ps
(xi_l, xj_h); dy_0 = _mm_sub_ps(yi_l, yj_h); dz_0 = _mm_sub_ps
(zi_l, zj_h); dx_1 = _mm_sub_ps(xj_l, xi_h); dy_1 = _mm_sub_ps
(yj_l, yi_h); dz_1 = _mm_sub_ps(zj_l, zi_h); mx = _mm_max_ps(
dx_0, dx_1); my = _mm_max_ps(dy_0, dy_1); mz = _mm_max_ps(dz_0
, dz_1); m0x = _mm_max_ps(mx, zero); m0y = _mm_max_ps(my, zero
); m0z = _mm_max_ps(mz, zero); d2x = _mm_mul_ps(m0x, m0x); d2y
= _mm_mul_ps(m0y, m0y); d2z = _mm_mul_ps(m0z, m0z); d2s = _mm_add_ps
(d2x, d2y); d2t = _mm_add_ps(d2s, d2z); _mm_store_ps(d2+0, d2t
); };
2227 if (STRIDE_PBB4 < nsi)
2228 {
2229 SUBC_BB_DIST2_SIMD4_XXXX_INNER(STRIDE_PBB, bb_i, d2){ int shi; __m128 dx_0, dy_0, dz_0; __m128 dx_1, dy_1, dz_1; __m128
mx, my, mz; __m128 m0x, m0y, m0z; __m128 d2x, d2y, d2z; __m128
d2s, d2t; shi = 4*2*3; xi_l = _mm_load_ps(bb_i+shi+0*4); yi_l
= _mm_load_ps(bb_i+shi+1*4); zi_l = _mm_load_ps(bb_i+shi+2*4
); xi_h = _mm_load_ps(bb_i+shi+3*4); yi_h = _mm_load_ps(bb_i+
shi+4*4); zi_h = _mm_load_ps(bb_i+shi+5*4); dx_0 = _mm_sub_ps
(xi_l, xj_h); dy_0 = _mm_sub_ps(yi_l, yj_h); dz_0 = _mm_sub_ps
(zi_l, zj_h); dx_1 = _mm_sub_ps(xj_l, xi_h); dy_1 = _mm_sub_ps
(yj_l, yi_h); dz_1 = _mm_sub_ps(zj_l, zi_h); mx = _mm_max_ps(
dx_0, dx_1); my = _mm_max_ps(dy_0, dy_1); mz = _mm_max_ps(dz_0
, dz_1); m0x = _mm_max_ps(mx, zero); m0y = _mm_max_ps(my, zero
); m0z = _mm_max_ps(mz, zero); d2x = _mm_mul_ps(m0x, m0x); d2y
= _mm_mul_ps(m0y, m0y); d2z = _mm_mul_ps(m0z, m0z); d2s = _mm_add_ps
(d2x, d2y); d2t = _mm_add_ps(d2s, d2z); _mm_store_ps(d2+4, d2t
); };
2230 }
2231}
2232
2233#endif /* NBNXN_SEARCH_BB_SIMD4 */
2234
2235/* Plain C function which determines if any atom pair between two cells
2236 * is within distance sqrt(rl2).
2237 */
2238static gmx_bool subc_in_range_x(int na_c,
2239 int si, const real *x_i,
2240 int csj, int stride, const real *x_j,
2241 real rl2)
2242{
2243 int i, j, i0, j0;
2244 real d2;
2245
2246 for (i = 0; i < na_c; i++)
2247 {
2248 i0 = (si*na_c + i)*DIM3;
2249 for (j = 0; j < na_c; j++)
2250 {
2251 j0 = (csj*na_c + j)*stride;
2252
2253 d2 = sqr(x_i[i0 ] - x_j[j0 ]) +
2254 sqr(x_i[i0+1] - x_j[j0+1]) +
2255 sqr(x_i[i0+2] - x_j[j0+2]);
2256
2257 if (d2 < rl2)
2258 {
2259 return TRUE1;
2260 }
2261 }
2262 }
2263
2264 return FALSE0;
2265}
2266
2267#ifdef NBNXN_SEARCH_SIMD4_FLOAT_X_BB
2268
2269/* 4-wide SIMD function which determines if any atom pair between two cells,
2270 * both with 8 atoms, is within distance sqrt(rl2).
2271 * Using 8-wide AVX is not faster on Intel Sandy Bridge.
2272 */
2273static gmx_bool subc_in_range_simd4(int na_c,
2274 int si, const real *x_i,
2275 int csj, int stride, const real *x_j,
2276 real rl2)
2277{
2278 gmx_simd4_real_t__m128 ix_S0, iy_S0, iz_S0;
2279 gmx_simd4_real_t__m128 ix_S1, iy_S1, iz_S1;
2280
2281 gmx_simd4_real_t__m128 rc2_S;
2282
2283 int dim_stride;
2284 int j0, j1;
2285
2286 rc2_S = gmx_simd4_set1_r_mm_set1_ps(rl2);
2287
2288 dim_stride = NBNXN_GPU_CLUSTER_SIZE8/STRIDE_PBB4*DIM3;
2289 ix_S0 = gmx_simd4_load_r_mm_load_ps(x_i+(si*dim_stride+0)*STRIDE_PBB4);
2290 iy_S0 = gmx_simd4_load_r_mm_load_ps(x_i+(si*dim_stride+1)*STRIDE_PBB4);
2291 iz_S0 = gmx_simd4_load_r_mm_load_ps(x_i+(si*dim_stride+2)*STRIDE_PBB4);
2292 ix_S1 = gmx_simd4_load_r_mm_load_ps(x_i+(si*dim_stride+3)*STRIDE_PBB4);
2293 iy_S1 = gmx_simd4_load_r_mm_load_ps(x_i+(si*dim_stride+4)*STRIDE_PBB4);
2294 iz_S1 = gmx_simd4_load_r_mm_load_ps(x_i+(si*dim_stride+5)*STRIDE_PBB4);
2295
2296 /* We loop from the outer to the inner particles to maximize
2297 * the chance that we find a pair in range quickly and return.
2298 */
2299 j0 = csj*na_c;
2300 j1 = j0 + na_c - 1;
2301 while (j0 < j1)
2302 {
2303 gmx_simd4_real_t__m128 jx0_S, jy0_S, jz0_S;
2304 gmx_simd4_real_t__m128 jx1_S, jy1_S, jz1_S;
2305
2306 gmx_simd4_real_t__m128 dx_S0, dy_S0, dz_S0;
2307 gmx_simd4_real_t__m128 dx_S1, dy_S1, dz_S1;
2308 gmx_simd4_real_t__m128 dx_S2, dy_S2, dz_S2;
2309 gmx_simd4_real_t__m128 dx_S3, dy_S3, dz_S3;
2310
2311 gmx_simd4_real_t__m128 rsq_S0;
2312 gmx_simd4_real_t__m128 rsq_S1;
2313 gmx_simd4_real_t__m128 rsq_S2;
2314 gmx_simd4_real_t__m128 rsq_S3;
2315
2316 gmx_simd4_bool_t__m128 wco_S0;
2317 gmx_simd4_bool_t__m128 wco_S1;
2318 gmx_simd4_bool_t__m128 wco_S2;
2319 gmx_simd4_bool_t__m128 wco_S3;
2320 gmx_simd4_bool_t__m128 wco_any_S01, wco_any_S23, wco_any_S;
2321
2322 jx0_S = gmx_simd4_set1_r_mm_set1_ps(x_j[j0*stride+0]);
2323 jy0_S = gmx_simd4_set1_r_mm_set1_ps(x_j[j0*stride+1]);
2324 jz0_S = gmx_simd4_set1_r_mm_set1_ps(x_j[j0*stride+2]);
2325
2326 jx1_S = gmx_simd4_set1_r_mm_set1_ps(x_j[j1*stride+0]);
2327 jy1_S = gmx_simd4_set1_r_mm_set1_ps(x_j[j1*stride+1]);
2328 jz1_S = gmx_simd4_set1_r_mm_set1_ps(x_j[j1*stride+2]);
2329
2330 /* Calculate distance */
2331 dx_S0 = gmx_simd4_sub_r_mm_sub_ps(ix_S0, jx0_S);
2332 dy_S0 = gmx_simd4_sub_r_mm_sub_ps(iy_S0, jy0_S);
2333 dz_S0 = gmx_simd4_sub_r_mm_sub_ps(iz_S0, jz0_S);
2334 dx_S1 = gmx_simd4_sub_r_mm_sub_ps(ix_S1, jx0_S);
2335 dy_S1 = gmx_simd4_sub_r_mm_sub_ps(iy_S1, jy0_S);
2336 dz_S1 = gmx_simd4_sub_r_mm_sub_ps(iz_S1, jz0_S);
2337 dx_S2 = gmx_simd4_sub_r_mm_sub_ps(ix_S0, jx1_S);
2338 dy_S2 = gmx_simd4_sub_r_mm_sub_ps(iy_S0, jy1_S);
2339 dz_S2 = gmx_simd4_sub_r_mm_sub_ps(iz_S0, jz1_S);
2340 dx_S3 = gmx_simd4_sub_r_mm_sub_ps(ix_S1, jx1_S);
2341 dy_S3 = gmx_simd4_sub_r_mm_sub_ps(iy_S1, jy1_S);
2342 dz_S3 = gmx_simd4_sub_r_mm_sub_ps(iz_S1, jz1_S);
2343
2344 /* rsq = dx*dx+dy*dy+dz*dz */
2345 rsq_S0 = gmx_simd4_calc_rsq_rgmx_simd4_norm2_f(dx_S0, dy_S0, dz_S0);
2346 rsq_S1 = gmx_simd4_calc_rsq_rgmx_simd4_norm2_f(dx_S1, dy_S1, dz_S1);
2347 rsq_S2 = gmx_simd4_calc_rsq_rgmx_simd4_norm2_f(dx_S2, dy_S2, dz_S2);
2348 rsq_S3 = gmx_simd4_calc_rsq_rgmx_simd4_norm2_f(dx_S3, dy_S3, dz_S3);
2349
2350 wco_S0 = gmx_simd4_cmplt_r_mm_cmplt_ps(rsq_S0, rc2_S);
2351 wco_S1 = gmx_simd4_cmplt_r_mm_cmplt_ps(rsq_S1, rc2_S);
2352 wco_S2 = gmx_simd4_cmplt_r_mm_cmplt_ps(rsq_S2, rc2_S);
2353 wco_S3 = gmx_simd4_cmplt_r_mm_cmplt_ps(rsq_S3, rc2_S);
2354
2355 wco_any_S01 = gmx_simd4_or_b_mm_or_ps(wco_S0, wco_S1);
2356 wco_any_S23 = gmx_simd4_or_b_mm_or_ps(wco_S2, wco_S3);
2357 wco_any_S = gmx_simd4_or_b_mm_or_ps(wco_any_S01, wco_any_S23);
2358
2359 if (gmx_simd4_anytrue_b_mm_movemask_ps(wco_any_S))
2360 {
2361 return TRUE1;
2362 }
2363
2364 j0++;
2365 j1--;
2366 }
2367 return FALSE0;
2368
2369}
2370#endif
2371
2372
2373/* Returns the j sub-cell for index cj_ind */
2374static int nbl_cj(const nbnxn_pairlist_t *nbl, int cj_ind)
2375{
2376 return nbl->cj4[cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG2].cj[cj_ind & (NBNXN_GPU_JGROUP_SIZE4 - 1)];
2377}
2378
2379/* Returns the i-interaction mask of the j sub-cell for index cj_ind */
2380static unsigned int nbl_imask0(const nbnxn_pairlist_t *nbl, int cj_ind)
2381{
2382 return nbl->cj4[cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG2].imei[0].imask;
2383}
2384
2385/* Ensures there is enough space for extra extra exclusion masks */
2386static void check_excl_space(nbnxn_pairlist_t *nbl, int extra)
2387{
2388 if (nbl->nexcl+extra > nbl->excl_nalloc)
2389 {
2390 nbl->excl_nalloc = over_alloc_small(nbl->nexcl+extra)(int)(1.19*(nbl->nexcl+extra) + 8000);
2391 nbnxn_realloc_void((void **)&nbl->excl,
2392 nbl->nexcl*sizeof(*nbl->excl),
2393 nbl->excl_nalloc*sizeof(*nbl->excl),
2394 nbl->alloc, nbl->free);
2395 }
2396}
2397
2398/* Ensures there is enough space for ncell extra j-cells in the list */
2399static void check_subcell_list_space_simple(nbnxn_pairlist_t *nbl,
2400 int ncell)
2401{
2402 int cj_max;
2403
2404 cj_max = nbl->ncj + ncell;
2405
2406 if (cj_max > nbl->cj_nalloc)
2407 {
2408 nbl->cj_nalloc = over_alloc_small(cj_max)(int)(1.19*(cj_max) + 8000);
2409 nbnxn_realloc_void((void **)&nbl->cj,
2410 nbl->ncj*sizeof(*nbl->cj),
2411 nbl->cj_nalloc*sizeof(*nbl->cj),
2412 nbl->alloc, nbl->free);
2413 }
2414}
2415
2416/* Ensures there is enough space for ncell extra j-subcells in the list */
2417static void check_subcell_list_space_supersub(nbnxn_pairlist_t *nbl,
2418 int nsupercell)
2419{
2420 int ncj4_max, j4, j, w, t;
2421
2422#define NWARP2 2
2423#define WARP_SIZE32 32
2424
2425 /* We can have maximally nsupercell*GPU_NSUBCELL sj lists */
2426 /* We can store 4 j-subcell - i-supercell pairs in one struct.
2427 * since we round down, we need one extra entry.
2428 */
2429 ncj4_max = ((nbl->work->cj_ind + nsupercell*GPU_NSUBCELL(2*2*2) + NBNXN_GPU_JGROUP_SIZE4 - 1) >> NBNXN_GPU_JGROUP_SIZE_2LOG2);
2430
2431 if (ncj4_max > nbl->cj4_nalloc)
2432 {
2433 nbl->cj4_nalloc = over_alloc_small(ncj4_max)(int)(1.19*(ncj4_max) + 8000);
2434 nbnxn_realloc_void((void **)&nbl->cj4,
2435 nbl->work->cj4_init*sizeof(*nbl->cj4),
2436 nbl->cj4_nalloc*sizeof(*nbl->cj4),
2437 nbl->alloc, nbl->free);
2438 }
2439
2440 if (ncj4_max > nbl->work->cj4_init)
2441 {
2442 for (j4 = nbl->work->cj4_init; j4 < ncj4_max; j4++)
2443 {
2444 /* No i-subcells and no excl's in the list initially */
2445 for (w = 0; w < NWARP2; w++)
2446 {
2447 nbl->cj4[j4].imei[w].imask = 0U;
2448 nbl->cj4[j4].imei[w].excl_ind = 0;
2449
2450 }
2451 }
2452 nbl->work->cj4_init = ncj4_max;
2453 }
2454}
2455
2456/* Set all excl masks for one GPU warp no exclusions */
2457static void set_no_excls(nbnxn_excl_t *excl)
2458{
2459 int t;
2460
2461 for (t = 0; t < WARP_SIZE32; t++)
2462 {
2463 /* Turn all interaction bits on */
2464 excl->pair[t] = NBNXN_INTERACTION_MASK_ALL;
2465 }
2466}
2467
2468/* Initializes a single nbnxn_pairlist_t data structure */
2469static void nbnxn_init_pairlist(nbnxn_pairlist_t *nbl,
2470 gmx_bool bSimple,
2471 nbnxn_alloc_t *alloc,
2472 nbnxn_free_t *free)
2473{
2474 if (alloc == NULL((void*)0))
2475 {
2476 nbl->alloc = nbnxn_alloc_aligned;
2477 }
2478 else
2479 {
2480 nbl->alloc = alloc;
2481 }
2482 if (free == NULL((void*)0))
2483 {
2484 nbl->free = nbnxn_free_aligned;
2485 }
2486 else
2487 {
2488 nbl->free = free;
2489 }
2490
2491 nbl->bSimple = bSimple;
2492 nbl->na_sc = 0;
2493 nbl->na_ci = 0;
2494 nbl->na_cj = 0;
2495 nbl->nci = 0;
2496 nbl->ci = NULL((void*)0);
2497 nbl->ci_nalloc = 0;
2498 nbl->ncj = 0;
2499 nbl->cj = NULL((void*)0);
2500 nbl->cj_nalloc = 0;
2501 nbl->ncj4 = 0;
2502 /* We need one element extra in sj, so alloc initially with 1 */
2503 nbl->cj4_nalloc = 0;
2504 nbl->cj4 = NULL((void*)0);
2505 nbl->nci_tot = 0;
2506
2507 if (!nbl->bSimple)
2508 {
2509 nbl->excl = NULL((void*)0);
2510 nbl->excl_nalloc = 0;
2511 nbl->nexcl = 0;
2512 check_excl_space(nbl, 1);
2513 nbl->nexcl = 1;
2514 set_no_excls(&nbl->excl[0]);
2515 }
2516
2517 snew(nbl->work, 1)(nbl->work) = save_calloc("nbl->work", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2517, (1), sizeof(*(nbl->work)));
2518 if (nbl->bSimple)
2519 {
2520 snew_aligned(nbl->work->bb_ci, 1, NBNXN_SEARCH_BB_MEM_ALIGN)(nbl->work->bb_ci) = save_calloc_aligned("nbl->work->bb_ci"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2520, (1), sizeof(*(nbl->work->bb_ci)), (4*sizeof(float
)));
2521 }
2522 else
2523 {
2524#ifdef NBNXN_BBXXXX
2525 snew_aligned(nbl->work->pbb_ci, GPU_NSUBCELL/STRIDE_PBB*NNBSBB_XXXX, NBNXN_SEARCH_BB_MEM_ALIGN)(nbl->work->pbb_ci) = save_calloc_aligned("nbl->work->pbb_ci"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2525, ((2*2*2)/4*(2*3*4)), sizeof(*(nbl->work->pbb_ci
)), (4*sizeof(float)));
2526#else
2527 snew_aligned(nbl->work->bb_ci, GPU_NSUBCELL, NBNXN_SEARCH_BB_MEM_ALIGN)(nbl->work->bb_ci) = save_calloc_aligned("nbl->work->bb_ci"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2527, ((2*2*2)), sizeof(*(nbl->work->bb_ci)), (4*sizeof
(float)));
2528#endif
2529 }
2530 snew_aligned(nbl->work->x_ci, NBNXN_NA_SC_MAX*DIM, NBNXN_SEARCH_BB_MEM_ALIGN)(nbl->work->x_ci) = save_calloc_aligned("nbl->work->x_ci"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2530, (((2*2*2)*8)*3), sizeof(*(nbl->work->x_ci)), (4
*sizeof(float)));
2531#ifdef GMX_NBNXN_SIMD
2532 snew_aligned(nbl->work->x_ci_simd_4xn, 1, NBNXN_MEM_ALIGN)(nbl->work->x_ci_simd_4xn) = save_calloc_aligned("nbl->work->x_ci_simd_4xn"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2532, (1), sizeof(*(nbl->work->x_ci_simd_4xn)), (4*sizeof
(real)));
2533 snew_aligned(nbl->work->x_ci_simd_2xnn, 1, NBNXN_MEM_ALIGN)(nbl->work->x_ci_simd_2xnn) = save_calloc_aligned("nbl->work->x_ci_simd_2xnn"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2533, (1), sizeof(*(nbl->work->x_ci_simd_2xnn)), (4*sizeof
(real)));
2534#endif
2535 snew_aligned(nbl->work->d2, GPU_NSUBCELL, NBNXN_SEARCH_BB_MEM_ALIGN)(nbl->work->d2) = save_calloc_aligned("nbl->work->d2"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2535, ((2*2*2)), sizeof(*(nbl->work->d2)), (4*sizeof(
float)));
2536
2537 nbl->work->sort = NULL((void*)0);
2538 nbl->work->sort_nalloc = 0;
2539 nbl->work->sci_sort = NULL((void*)0);
2540 nbl->work->sci_sort_nalloc = 0;
2541}
2542
2543void nbnxn_init_pairlist_set(nbnxn_pairlist_set_t *nbl_list,
2544 gmx_bool bSimple, gmx_bool bCombined,
2545 nbnxn_alloc_t *alloc,
2546 nbnxn_free_t *free)
2547{
2548 int i;
2549
2550 nbl_list->bSimple = bSimple;
2551 nbl_list->bCombined = bCombined;
2552
2553 nbl_list->nnbl = gmx_omp_nthreads_get(emntNonbonded);
2554
2555 if (!nbl_list->bCombined &&
2556 nbl_list->nnbl > NBNXN_BUFFERFLAG_MAX_THREADS32)
2557 {
2558 gmx_fatal(FARGS0, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2558, "%d OpenMP threads were requested. Since the non-bonded force buffer reduction is prohibitively slow with more than %d threads, we do not allow this. Use %d or less OpenMP threads.",
2559 nbl_list->nnbl, NBNXN_BUFFERFLAG_MAX_THREADS32, NBNXN_BUFFERFLAG_MAX_THREADS32);
2560 }
2561
2562 snew(nbl_list->nbl, nbl_list->nnbl)(nbl_list->nbl) = save_calloc("nbl_list->nbl", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2562, (nbl_list->nnbl), sizeof(*(nbl_list->nbl)));
2563 snew(nbl_list->nbl_fep, nbl_list->nnbl)(nbl_list->nbl_fep) = save_calloc("nbl_list->nbl_fep", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2563, (nbl_list->nnbl), sizeof(*(nbl_list->nbl_fep)));
2564 /* Execute in order to avoid memory interleaving between threads */
2565#pragma omp parallel for num_threads(nbl_list->nnbl) schedule(static)
2566 for (i = 0; i < nbl_list->nnbl; i++)
2567 {
2568 /* Allocate the nblist data structure locally on each thread
2569 * to optimize memory access for NUMA architectures.
2570 */
2571 snew(nbl_list->nbl[i], 1)(nbl_list->nbl[i]) = save_calloc("nbl_list->nbl[i]", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2571, (1), sizeof(*(nbl_list->nbl[i])));
2572
2573 /* Only list 0 is used on the GPU, use normal allocation for i>0 */
2574 if (i == 0)
2575 {
2576 nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, alloc, free);
2577 }
2578 else
2579 {
2580 nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, NULL((void*)0), NULL((void*)0));
2581 }
2582
2583 snew(nbl_list->nbl_fep[i], 1)(nbl_list->nbl_fep[i]) = save_calloc("nbl_list->nbl_fep[i]"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 2583, (1), sizeof(*(nbl_list->nbl_fep[i])));
2584 nbnxn_init_pairlist_fep(nbl_list->nbl_fep[i]);
2585 }
2586}
2587
2588/* Print statistics of a pair list, used for debug output */
2589static void print_nblist_statistics_simple(FILE *fp, const nbnxn_pairlist_t *nbl,
2590 const nbnxn_search_t nbs, real rl)
2591{
2592 const nbnxn_grid_t *grid;
2593 int cs[SHIFTS((2*1 +1)*(2*1 +1)*(2*2 +1))];
2594 int s, i, j;
2595 int npexcl;
2596
2597 /* This code only produces correct statistics with domain decomposition */
2598 grid = &nbs->grid[0];
2599
2600 fprintf(fp, "nbl nci %d ncj %d\n",
2601 nbl->nci, nbl->ncj);
2602 fprintf(fp, "nbl na_sc %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
2603 nbl->na_sc, rl, nbl->ncj, nbl->ncj/(double)grid->nc,
2604 nbl->ncj/(double)grid->nc*grid->na_sc,
2605 nbl->ncj/(double)grid->nc*grid->na_sc/(0.5*4.0/3.0*M_PI3.14159265358979323846*rl*rl*rl*grid->nc*grid->na_sc/det(nbs->box)));
2606
2607 fprintf(fp, "nbl average j cell list length %.1f\n",
2608 0.25*nbl->ncj/(double)nbl->nci);
2609
2610 for (s = 0; s < SHIFTS((2*1 +1)*(2*1 +1)*(2*2 +1)); s++)
2611 {
2612 cs[s] = 0;
2613 }
2614 npexcl = 0;
2615 for (i = 0; i < nbl->nci; i++)
2616 {
2617 cs[nbl->ci[i].shift & NBNXN_CI_SHIFT127] +=
2618 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start;
2619
2620 j = nbl->ci[i].cj_ind_start;
2621 while (j < nbl->ci[i].cj_ind_end &&
2622 nbl->cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
2623 {
2624 npexcl++;
2625 j++;
2626 }
2627 }
2628 fprintf(fp, "nbl cell pairs, total: %d excl: %d %.1f%%\n",
2629 nbl->ncj, npexcl, 100*npexcl/(double)nbl->ncj);
2630 for (s = 0; s < SHIFTS((2*1 +1)*(2*1 +1)*(2*2 +1)); s++)
2631 {
2632 if (cs[s] > 0)
2633 {
2634 fprintf(fp, "nbl shift %2d ncj %3d\n", s, cs[s]);
2635 }
2636 }
2637}
2638
2639/* Print statistics of a pair lists, used for debug output */
2640static void print_nblist_statistics_supersub(FILE *fp, const nbnxn_pairlist_t *nbl,
2641 const nbnxn_search_t nbs, real rl)
2642{
2643 const nbnxn_grid_t *grid;
2644 int i, j4, j, si, b;
2645 int c[GPU_NSUBCELL(2*2*2)+1];
2646
2647 /* This code only produces correct statistics with domain decomposition */
2648 grid = &nbs->grid[0];
2649
2650 fprintf(fp, "nbl nsci %d ncj4 %d nsi %d excl4 %d\n",
2651 nbl->nsci, nbl->ncj4, nbl->nci_tot, nbl->nexcl);
2652 fprintf(fp, "nbl na_c %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
2653 nbl->na_ci, rl, nbl->nci_tot, nbl->nci_tot/(double)grid->nsubc_tot,
2654 nbl->nci_tot/(double)grid->nsubc_tot*grid->na_c,
2655 nbl->nci_tot/(double)grid->nsubc_tot*grid->na_c/(0.5*4.0/3.0*M_PI3.14159265358979323846*rl*rl*rl*grid->nsubc_tot*grid->na_c/det(nbs->box)));
2656
2657 fprintf(fp, "nbl average j super cell list length %.1f\n",
2658 0.25*nbl->ncj4/(double)nbl->nsci);
2659 fprintf(fp, "nbl average i sub cell list length %.1f\n",
2660 nbl->nci_tot/((double)nbl->ncj4));
2661
2662 for (si = 0; si <= GPU_NSUBCELL(2*2*2); si++)
2663 {
2664 c[si] = 0;
2665 }
2666 for (i = 0; i < nbl->nsci; i++)
2667 {
2668 for (j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
2669 {
2670 for (j = 0; j < NBNXN_GPU_JGROUP_SIZE4; j++)
2671 {
2672 b = 0;
2673 for (si = 0; si < GPU_NSUBCELL(2*2*2); si++)
2674 {
2675 if (nbl->cj4[j4].imei[0].imask & (1U << (j*GPU_NSUBCELL(2*2*2) + si)))
2676 {
2677 b++;
2678 }
2679 }
2680 c[b]++;
2681 }
2682 }
2683 }
2684 for (b = 0; b <= GPU_NSUBCELL(2*2*2); b++)
2685 {
2686 fprintf(fp, "nbl j-list #i-subcell %d %7d %4.1f\n",
2687 b, c[b], 100.0*c[b]/(double)(nbl->ncj4*NBNXN_GPU_JGROUP_SIZE4));
2688 }
2689}
2690
2691/* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp */
2692static void low_get_nbl_exclusions(nbnxn_pairlist_t *nbl, int cj4,
2693 int warp, nbnxn_excl_t **excl)
2694{
2695 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
2696 {
2697 /* No exclusions set, make a new list entry */
2698 nbl->cj4[cj4].imei[warp].excl_ind = nbl->nexcl;
2699 nbl->nexcl++;
2700 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
2701 set_no_excls(*excl);
2702 }
2703 else
2704 {
2705 /* We already have some exclusions, new ones can be added to the list */
2706 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
2707 }
2708}
2709
2710/* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp,
2711 * generates a new element and allocates extra memory, if necessary.
2712 */
2713static void get_nbl_exclusions_1(nbnxn_pairlist_t *nbl, int cj4,
2714 int warp, nbnxn_excl_t **excl)
2715{
2716 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
2717 {
2718 /* We need to make a new list entry, check if we have space */
2719 check_excl_space(nbl, 1);
2720 }
2721 low_get_nbl_exclusions(nbl, cj4, warp, excl);
2722}
2723
2724/* Returns pointers to the exclusion mask for cj4-unit cj4 for both warps,
2725 * generates a new element and allocates extra memory, if necessary.
2726 */
2727static void get_nbl_exclusions_2(nbnxn_pairlist_t *nbl, int cj4,
2728 nbnxn_excl_t **excl_w0,
2729 nbnxn_excl_t **excl_w1)
2730{
2731 /* Check for space we might need */
2732 check_excl_space(nbl, 2);
2733
2734 low_get_nbl_exclusions(nbl, cj4, 0, excl_w0);
2735 low_get_nbl_exclusions(nbl, cj4, 1, excl_w1);
2736}
2737
2738/* Sets the self exclusions i=j and pair exclusions i>j */
2739static void set_self_and_newton_excls_supersub(nbnxn_pairlist_t *nbl,
2740 int cj4_ind, int sj_offset,
2741 int si)
2742{
2743 nbnxn_excl_t *excl[2];
2744 int ei, ej, w;
2745
2746 /* Here we only set the set self and double pair exclusions */
2747
2748 get_nbl_exclusions_2(nbl, cj4_ind, &excl[0], &excl[1]);
2749
2750 /* Only minor < major bits set */
2751 for (ej = 0; ej < nbl->na_ci; ej++)
2752 {
2753 w = (ej>>2);
2754 for (ei = ej; ei < nbl->na_ci; ei++)
2755 {
2756 excl[w]->pair[(ej & (NBNXN_GPU_JGROUP_SIZE4-1))*nbl->na_ci + ei] &=
2757 ~(1U << (sj_offset*GPU_NSUBCELL(2*2*2) + si));
2758 }
2759 }
2760}
2761
2762/* Returns a diagonal or off-diagonal interaction mask for plain C lists */
2763static unsigned int get_imask(gmx_bool rdiag, int ci, int cj)
2764{
2765 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
2766}
2767
2768/* Returns a diagonal or off-diagonal interaction mask for cj-size=2 */
2769static unsigned int get_imask_simd_j2(gmx_bool rdiag, int ci, int cj)
2770{
2771 return (rdiag && ci*2 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_0 :
2772 (rdiag && ci*2+1 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_1 :
2773 NBNXN_INTERACTION_MASK_ALL));
2774}
2775
2776/* Returns a diagonal or off-diagonal interaction mask for cj-size=4 */
2777static unsigned int get_imask_simd_j4(gmx_bool rdiag, int ci, int cj)
2778{
2779 return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
2780}
2781
2782/* Returns a diagonal or off-diagonal interaction mask for cj-size=8 */
2783static unsigned int get_imask_simd_j8(gmx_bool rdiag, int ci, int cj)
2784{
2785 return (rdiag && ci == cj*2 ? NBNXN_INTERACTION_MASK_DIAG_J8_0 :
2786 (rdiag && ci == cj*2+1 ? NBNXN_INTERACTION_MASK_DIAG_J8_1 :
2787 NBNXN_INTERACTION_MASK_ALL));
2788}
2789
2790#ifdef GMX_NBNXN_SIMD
2791#if GMX_SIMD_REAL_WIDTH4 == 2
2792#define get_imask_simd_4xnget_imask_simd_j4 get_imask_simd_j2
2793#endif
2794#if GMX_SIMD_REAL_WIDTH4 == 4
2795#define get_imask_simd_4xnget_imask_simd_j4 get_imask_simd_j4
2796#endif
2797#if GMX_SIMD_REAL_WIDTH4 == 8
2798#define get_imask_simd_4xnget_imask_simd_j4 get_imask_simd_j8
2799#define get_imask_simd_2xnn get_imask_simd_j4
2800#endif
2801#if GMX_SIMD_REAL_WIDTH4 == 16
2802#define get_imask_simd_2xnn get_imask_simd_j8
2803#endif
2804#endif
2805
2806/* Plain C code for making a pair list of cell ci vs cell cjf-cjl.
2807 * Checks bounding box distances and possibly atom pair distances.
2808 */
2809static void make_cluster_list_simple(const nbnxn_grid_t *gridj,
2810 nbnxn_pairlist_t *nbl,
2811 int ci, int cjf, int cjl,
2812 gmx_bool remove_sub_diag,
2813 const real *x_j,
2814 real rl2, float rbb2,
2815 int *ndistc)
2816{
2817 const nbnxn_list_work_t *work;
2818
2819 const nbnxn_bb_t *bb_ci;
2820 const real *x_ci;
2821
2822 gmx_bool InRange;
2823 real d2;
2824 int cjf_gl, cjl_gl, cj;
2825
2826 work = nbl->work;
2827
2828 bb_ci = nbl->work->bb_ci;
2829 x_ci = nbl->work->x_ci;
2830
2831 InRange = FALSE0;
2832 while (!InRange && cjf <= cjl)
2833 {
2834 d2 = subc_bb_dist2(0, bb_ci, cjf, gridj->bb);
2835 *ndistc += 2;
2836
2837 /* Check if the distance is within the distance where
2838 * we use only the bounding box distance rbb,
2839 * or within the cut-off and there is at least one atom pair
2840 * within the cut-off.
2841 */
2842 if (d2 < rbb2)
2843 {
2844 InRange = TRUE1;
2845 }
2846 else if (d2 < rl2)
2847 {
2848 int i, j;
2849
2850 cjf_gl = gridj->cell0 + cjf;
2851 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE4 && !InRange; i++)
2852 {
2853 for (j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE4; j++)
2854 {
2855 InRange = InRange ||
2856 (sqr(x_ci[i*STRIDE_XYZ3+XX0] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE4+j)*STRIDE_XYZ3+XX0]) +
2857 sqr(x_ci[i*STRIDE_XYZ3+YY1] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE4+j)*STRIDE_XYZ3+YY1]) +
2858 sqr(x_ci[i*STRIDE_XYZ3+ZZ2] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE4+j)*STRIDE_XYZ3+ZZ2]) < rl2);
2859 }
2860 }
2861 *ndistc += NBNXN_CPU_CLUSTER_I_SIZE4*NBNXN_CPU_CLUSTER_I_SIZE4;
2862 }
2863 if (!InRange)
2864 {
2865 cjf++;
2866 }
2867 }
2868 if (!InRange)
2869 {
2870 return;
2871 }
2872
2873 InRange = FALSE0;
2874 while (!InRange && cjl > cjf)
2875 {
2876 d2 = subc_bb_dist2(0, bb_ci, cjl, gridj->bb);
2877 *ndistc += 2;
2878
2879 /* Check if the distance is within the distance where
2880 * we use only the bounding box distance rbb,
2881 * or within the cut-off and there is at least one atom pair
2882 * within the cut-off.
2883 */
2884 if (d2 < rbb2)
2885 {
2886 InRange = TRUE1;
2887 }
2888 else if (d2 < rl2)
2889 {
2890 int i, j;
2891
2892 cjl_gl = gridj->cell0 + cjl;
2893 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE4 && !InRange; i++)
2894 {
2895 for (j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE4; j++)
2896 {
2897 InRange = InRange ||
2898 (sqr(x_ci[i*STRIDE_XYZ3+XX0] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE4+j)*STRIDE_XYZ3+XX0]) +
2899 sqr(x_ci[i*STRIDE_XYZ3+YY1] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE4+j)*STRIDE_XYZ3+YY1]) +
2900 sqr(x_ci[i*STRIDE_XYZ3+ZZ2] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE4+j)*STRIDE_XYZ3+ZZ2]) < rl2);
2901 }
2902 }
2903 *ndistc += NBNXN_CPU_CLUSTER_I_SIZE4*NBNXN_CPU_CLUSTER_I_SIZE4;
2904 }
2905 if (!InRange)
2906 {
2907 cjl--;
2908 }
2909 }
2910
2911 if (cjf <= cjl)
2912 {
2913 for (cj = cjf; cj <= cjl; cj++)
2914 {
2915 /* Store cj and the interaction mask */
2916 nbl->cj[nbl->ncj].cj = gridj->cell0 + cj;
2917 nbl->cj[nbl->ncj].excl = get_imask(remove_sub_diag, ci, cj);
2918 nbl->ncj++;
2919 }
2920 /* Increase the closing index in i super-cell list */
2921 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
2922 }
2923}
2924
2925#ifdef GMX_NBNXN_SIMD_4XN
2926#include "nbnxn_search_simd_4xn.h"
2927#endif
2928#ifdef GMX_NBNXN_SIMD_2XNN
2929#include "nbnxn_search_simd_2xnn.h"
2930#endif
2931
2932/* Plain C or SIMD4 code for making a pair list of super-cell sci vs scj.
2933 * Checks bounding box distances and possibly atom pair distances.
2934 */
2935static void make_cluster_list_supersub(const nbnxn_grid_t *gridi,
2936 const nbnxn_grid_t *gridj,
2937 nbnxn_pairlist_t *nbl,
2938 int sci, int scj,
2939 gmx_bool sci_equals_scj,
2940 int stride, const real *x,
2941 real rl2, float rbb2,
2942 int *ndistc)
2943{
2944 int na_c;
2945 int npair;
2946 int cjo, ci1, ci, cj, cj_gl;
2947 int cj4_ind, cj_offset;
2948 unsigned int imask;
2949 nbnxn_cj4_t *cj4;
2950#ifdef NBNXN_BBXXXX
2951 const float *pbb_ci;
2952#else
2953 const nbnxn_bb_t *bb_ci;
2954#endif
2955 const real *x_ci;
2956 float *d2l, d2;
2957 int w;
2958#define PRUNE_LIST_CPU_ONE
2959#ifdef PRUNE_LIST_CPU_ONE
2960 int ci_last = -1;
2961#endif
2962
2963 d2l = nbl->work->d2;
2964
2965#ifdef NBNXN_BBXXXX
2966 pbb_ci = nbl->work->pbb_ci;
2967#else
2968 bb_ci = nbl->work->bb_ci;
2969#endif
2970 x_ci = nbl->work->x_ci;
2971
2972 na_c = gridj->na_c;
2973
2974 for (cjo = 0; cjo < gridj->nsubc[scj]; cjo++)
2975 {
2976 cj4_ind = (nbl->work->cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG2);
2977 cj_offset = nbl->work->cj_ind - cj4_ind*NBNXN_GPU_JGROUP_SIZE4;
2978 cj4 = &nbl->cj4[cj4_ind];
2979
2980 cj = scj*GPU_NSUBCELL(2*2*2) + cjo;
2981
2982 cj_gl = gridj->cell0*GPU_NSUBCELL(2*2*2) + cj;
2983
2984 /* Initialize this j-subcell i-subcell list */
2985 cj4->cj[cj_offset] = cj_gl;
2986 imask = 0;
2987
2988 if (sci_equals_scj)
2989 {
2990 ci1 = cjo + 1;
2991 }
2992 else
2993 {
2994 ci1 = gridi->nsubc[sci];
2995 }
2996
2997#ifdef NBNXN_BBXXXX
2998 /* Determine all ci1 bb distances in one call with SIMD4 */
2999 subc_bb_dist2_simd4_xxxx(gridj->pbb+(cj>>STRIDE_PBB_2LOG2)*NNBSBB_XXXX(2*3*4)+(cj & (STRIDE_PBB4-1)),
3000 ci1, pbb_ci, d2l);
3001 *ndistc += na_c*2;
3002#endif
3003
3004 npair = 0;
3005 /* We use a fixed upper-bound instead of ci1 to help optimization */
3006 for (ci = 0; ci < GPU_NSUBCELL(2*2*2); ci++)
3007 {
3008 if (ci == ci1)
3009 {
3010 break;
3011 }
3012
3013#ifndef NBNXN_BBXXXX
3014 /* Determine the bb distance between ci and cj */
3015 d2l[ci] = subc_bb_dist2(ci, bb_ci, cj, gridj->bb);
3016 *ndistc += 2;
3017#endif
3018 d2 = d2l[ci];
3019
3020#ifdef PRUNE_LIST_CPU_ALL
3021 /* Check if the distance is within the distance where
3022 * we use only the bounding box distance rbb,
3023 * or within the cut-off and there is at least one atom pair
3024 * within the cut-off. This check is very costly.
3025 */
3026 *ndistc += na_c*na_c;
3027 if (d2 < rbb2 ||
3028 (d2 < rl2 &&
3029#ifdef NBNXN_PBB_SIMD4
3030 subc_in_range_simd4
3031#else
3032 subc_in_range_x
3033#endif
3034 (na_c, ci, x_ci, cj_gl, stride, x, rl2)))
3035#else
3036 /* Check if the distance between the two bounding boxes
3037 * in within the pair-list cut-off.
3038 */
3039 if (d2 < rl2)
3040#endif
3041 {
3042 /* Flag this i-subcell to be taken into account */
3043 imask |= (1U << (cj_offset*GPU_NSUBCELL(2*2*2)+ci));
3044
3045#ifdef PRUNE_LIST_CPU_ONE
3046 ci_last = ci;
3047#endif
3048
3049 npair++;
3050 }
3051 }
3052
3053#ifdef PRUNE_LIST_CPU_ONE
3054 /* If we only found 1 pair, check if any atoms are actually
3055 * within the cut-off, so we could get rid of it.
3056 */
3057 if (npair == 1 && d2l[ci_last] >= rbb2)
3058 {
3059 /* Avoid using function pointers here, as it's slower */
3060 if (
3061#ifdef NBNXN_PBB_SIMD4
3062 !subc_in_range_simd4
3063#else
3064 !subc_in_range_x
3065#endif
3066 (na_c, ci_last, x_ci, cj_gl, stride, x, rl2))
3067 {
3068 imask &= ~(1U << (cj_offset*GPU_NSUBCELL(2*2*2)+ci_last));
3069 npair--;
3070 }
3071 }
3072#endif
3073
3074 if (npair > 0)
3075 {
3076 /* We have a useful sj entry, close it now */
3077
3078 /* Set the exclucions for the ci== sj entry.
3079 * Here we don't bother to check if this entry is actually flagged,
3080 * as it will nearly always be in the list.
3081 */
3082 if (sci_equals_scj)
3083 {
3084 set_self_and_newton_excls_supersub(nbl, cj4_ind, cj_offset, cjo);
3085 }
3086
3087 /* Copy the cluster interaction mask to the list */
3088 for (w = 0; w < NWARP2; w++)
3089 {
3090 cj4->imei[w].imask |= imask;
3091 }
3092
3093 nbl->work->cj_ind++;
3094
3095 /* Keep the count */
3096 nbl->nci_tot += npair;
3097
3098 /* Increase the closing index in i super-cell list */
3099 nbl->sci[nbl->nsci].cj4_ind_end =
3100 ((nbl->work->cj_ind+NBNXN_GPU_JGROUP_SIZE4-1) >> NBNXN_GPU_JGROUP_SIZE_2LOG2);
3101 }
3102 }
3103}
3104
3105/* Set all atom-pair exclusions from the topology stored in excl
3106 * as masks in the pair-list for simple list i-entry nbl_ci
3107 */
3108static void set_ci_top_excls(const nbnxn_search_t nbs,
3109 nbnxn_pairlist_t *nbl,
3110 gmx_bool diagRemoved,
3111 int na_ci_2log,
3112 int na_cj_2log,
3113 const nbnxn_ci_t *nbl_ci,
3114 const t_blocka *excl)
3115{
3116 const int *cell;
3117 int ci;
3118 int cj_ind_first, cj_ind_last;
3119 int cj_first, cj_last;
3120 int ndirect;
3121 int i, ai, aj, si, eind, ge, se;
3122 int found, cj_ind_0, cj_ind_1, cj_ind_m;
3123 int cj_m;
3124 gmx_bool Found_si;
3125 int si_ind;
3126 nbnxn_excl_t *nbl_excl;
3127 int inner_i, inner_e;
3128
3129 cell = nbs->cell;
3130
3131 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
3132 {
3133 /* Empty list */
3134 return;
3135 }
3136
3137 ci = nbl_ci->ci;
3138
3139 cj_ind_first = nbl_ci->cj_ind_start;
3140 cj_ind_last = nbl->ncj - 1;
3141
3142 cj_first = nbl->cj[cj_ind_first].cj;
3143 cj_last = nbl->cj[cj_ind_last].cj;
3144
3145 /* Determine how many contiguous j-cells we have starting
3146 * from the first i-cell. This number can be used to directly
3147 * calculate j-cell indices for excluded atoms.
3148 */
3149 ndirect = 0;
3150 if (na_ci_2log == na_cj_2log)
3151 {
3152 while (cj_ind_first + ndirect <= cj_ind_last &&
3153 nbl->cj[cj_ind_first+ndirect].cj == ci + ndirect)
3154 {
3155 ndirect++;
3156 }
3157 }
3158#ifdef NBNXN_SEARCH_BB_SIMD4
3159 else
3160 {
3161 while (cj_ind_first + ndirect <= cj_ind_last &&
3162 nbl->cj[cj_ind_first+ndirect].cj == ci_to_cj(na_cj_2log, ci) + ndirect)
3163 {
3164 ndirect++;
3165 }
3166 }
3167#endif
3168
3169 /* Loop over the atoms in the i super-cell */
3170 for (i = 0; i < nbl->na_sc; i++)
3171 {
3172 ai = nbs->a[ci*nbl->na_sc+i];
3173 if (ai >= 0)
3174 {
3175 si = (i>>na_ci_2log);
3176
3177 /* Loop over the topology-based exclusions for this i-atom */
3178 for (eind = excl->index[ai]; eind < excl->index[ai+1]; eind++)
3179 {
3180 aj = excl->a[eind];
3181
3182 if (aj == ai)
3183 {
3184 /* The self exclusion are already set, save some time */
3185 continue;
3186 }
3187
3188 ge = cell[aj];
3189
3190 /* Without shifts we only calculate interactions j>i
3191 * for one-way pair-lists.
3192 */
3193 if (diagRemoved && ge <= ci*nbl->na_sc + i)
3194 {
3195 continue;
3196 }
3197
3198 se = (ge >> na_cj_2log);
3199
3200 /* Could the cluster se be in our list? */
3201 if (se >= cj_first && se <= cj_last)
3202 {
3203 if (se < cj_first + ndirect)
3204 {
3205 /* We can calculate cj_ind directly from se */
3206 found = cj_ind_first + se - cj_first;
3207 }
3208 else
3209 {
3210 /* Search for se using bisection */
3211 found = -1;
3212 cj_ind_0 = cj_ind_first + ndirect;
3213 cj_ind_1 = cj_ind_last + 1;
3214 while (found == -1 && cj_ind_0 < cj_ind_1)
3215 {
3216 cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
3217
3218 cj_m = nbl->cj[cj_ind_m].cj;
3219
3220 if (se == cj_m)
3221 {
3222 found = cj_ind_m;
3223 }
3224 else if (se < cj_m)
3225 {
3226 cj_ind_1 = cj_ind_m;
3227 }
3228 else
3229 {
3230 cj_ind_0 = cj_ind_m + 1;
3231 }
3232 }
3233 }
3234
3235 if (found >= 0)
3236 {
3237 inner_i = i - (si << na_ci_2log);
3238 inner_e = ge - (se << na_cj_2log);
3239
3240 nbl->cj[found].excl &= ~(1U<<((inner_i<<na_cj_2log) + inner_e));
3241/* The next code line is usually not needed. We do not want to version
3242 * away the above line, because there is logic that relies on being
3243 * able to detect easily whether any exclusions exist. */
3244#if (defined GMX_SIMD_IBM_QPX)
3245 nbl->cj[found].interaction_mask_indices[inner_i] &= ~(1U << inner_e);
3246#endif
3247 }
3248 }
3249 }
3250 }
3251 }
3252}
3253
3254/* Add a new i-entry to the FEP list and copy the i-properties */
3255static gmx_inlineinline void fep_list_new_nri_copy(t_nblist *nlist)
3256{
3257 /* Add a new i-entry */
3258 nlist->nri++;
3259
3260 assert(nlist->nri < nlist->maxnri)((void) (0));
3261
3262 /* Duplicate the last i-entry, except for jindex, which continues */
3263 nlist->iinr[nlist->nri] = nlist->iinr[nlist->nri-1];
3264 nlist->shift[nlist->nri] = nlist->shift[nlist->nri-1];
3265 nlist->gid[nlist->nri] = nlist->gid[nlist->nri-1];
3266 nlist->jindex[nlist->nri] = nlist->nrj;
3267}
3268
3269/* For load balancing of the free-energy lists over threads, we set
3270 * the maximum nrj size of an i-entry to 40. This leads to good
3271 * load balancing in the worst case scenario of a single perturbed
3272 * particle on 16 threads, while not introducing significant overhead.
3273 * Note that half of the perturbed pairs will anyhow end up in very small lists,
3274 * since non perturbed i-particles will see few perturbed j-particles).
3275 */
3276const int max_nrj_fep = 40;
3277
3278/* Exclude the perturbed pairs from the Verlet list. This is only done to avoid
3279 * singularities for overlapping particles (0/0), since the charges and
3280 * LJ parameters have been zeroed in the nbnxn data structure.
3281 * Simultaneously make a group pair list for the perturbed pairs.
3282 */
3283static void make_fep_list(const nbnxn_search_t nbs,
3284 const nbnxn_atomdata_t *nbat,
3285 nbnxn_pairlist_t *nbl,
3286 gmx_bool bDiagRemoved,
3287 nbnxn_ci_t *nbl_ci,
3288 const nbnxn_grid_t *gridi,
3289 const nbnxn_grid_t *gridj,
3290 t_nblist *nlist)
3291{
3292 int ci, cj_ind_start, cj_ind_end, cj_ind, cja, cjr;
3293 int nri_max;
3294 int ngid, gid_i = 0, gid_j, gid;
3295 int egp_shift, egp_mask;
3296 int gid_cj = 0;
3297 int i, j, ind_i, ind_j, ai, aj;
3298 int nri;
3299 gmx_bool bFEP_i, bFEP_i_all;
3300
3301 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
3302 {
3303 /* Empty list */
3304 return;
3305 }
3306
3307 ci = nbl_ci->ci;
3308
3309 cj_ind_start = nbl_ci->cj_ind_start;
3310 cj_ind_end = nbl_ci->cj_ind_end;
3311
3312 /* In worst case we have alternating energy groups and create npair lists */
3313 nri_max = nbl->na_ci*(cj_ind_end - cj_ind_start);
3314 if (nlist->nri + nri_max > nlist->maxnri)
3315 {
3316 nlist->maxnri = over_alloc_large(nlist->nri + nri_max)(int)(1.19*(nlist->nri + nri_max) + 1000);
3317 reallocate_nblist(nlist);
3318 }
3319
3320 ngid = nbat->nenergrp;
3321
3322 if (ngid*gridj->na_cj > sizeof(gid_cj)*8)
3323 {
3324 gmx_fatal(FARGS0, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 3324, "The Verlet scheme with %dx%d kernels and free-energy only supports up to %d energy groups",
3325 gridi->na_c, gridj->na_cj, (sizeof(gid_cj)*8)/gridj->na_cj);
3326 }
3327
3328 egp_shift = nbat->neg_2log;
3329 egp_mask = (1<<nbat->neg_2log) - 1;
3330
3331 /* Loop over the atoms in the i sub-cell */
3332 bFEP_i_all = TRUE1;
3333 for (i = 0; i < nbl->na_ci; i++)
3334 {
3335 ind_i = ci*nbl->na_ci + i;
3336 ai = nbs->a[ind_i];
3337 if (ai >= 0)
3338 {
3339 nri = nlist->nri;
3340 nlist->jindex[nri+1] = nlist->jindex[nri];
3341 nlist->iinr[nri] = ai;
3342 /* The actual energy group pair index is set later */
3343 nlist->gid[nri] = 0;
3344 nlist->shift[nri] = nbl_ci->shift & NBNXN_CI_SHIFT127;
3345
3346 bFEP_i = gridi->fep[ci - gridi->cell0] & (1 << i);
3347
3348 bFEP_i_all = bFEP_i_all && bFEP_i;
3349
3350 if ((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj > nlist->maxnrj)
3351 {
3352 nlist->maxnrj = over_alloc_small((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj)(int)(1.19*((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->
na_cj) + 8000);
3353 srenew(nlist->jjnr, nlist->maxnrj)(nlist->jjnr) = save_realloc("nlist->jjnr", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 3353, (nlist->jjnr), (nlist->maxnrj), sizeof(*(nlist->
jjnr)));
3354 srenew(nlist->excl_fep, nlist->maxnrj)(nlist->excl_fep) = save_realloc("nlist->excl_fep", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 3354, (nlist->excl_fep), (nlist->maxnrj), sizeof(*(nlist
->excl_fep)));
3355 }
3356
3357 if (ngid > 1)
3358 {
3359 gid_i = (nbat->energrp[ci] >> (egp_shift*i)) & egp_mask;
3360 }
3361
3362 for (cj_ind = cj_ind_start; cj_ind < cj_ind_end; cj_ind++)
3363 {
3364 unsigned int fep_cj;
3365
3366 cja = nbl->cj[cj_ind].cj;
3367
3368 if (gridj->na_cj == gridj->na_c)
3369 {
3370 cjr = cja - gridj->cell0;
3371 fep_cj = gridj->fep[cjr];
3372 if (ngid > 1)
3373 {
3374 gid_cj = nbat->energrp[cja];
3375 }
3376 }
3377 else if (2*gridj->na_cj == gridj->na_c)
3378 {
3379 cjr = cja - gridj->cell0*2;
3380 /* Extract half of the ci fep/energrp mask */
3381 fep_cj = (gridj->fep[cjr>>1] >> ((cjr&1)*gridj->na_cj)) & ((1<<gridj->na_cj) - 1);
3382 if (ngid > 1)
3383 {
3384 gid_cj = nbat->energrp[cja>>1] >> ((cja&1)*gridj->na_cj*egp_shift) & ((1<<(gridj->na_cj*egp_shift)) - 1);
3385 }
3386 }
3387 else
3388 {
3389 cjr = cja - (gridj->cell0>>1);
3390 /* Combine two ci fep masks/energrp */
3391 fep_cj = gridj->fep[cjr*2] + (gridj->fep[cjr*2+1] << gridj->na_c);
3392 if (ngid > 1)
3393 {
3394 gid_cj = nbat->energrp[cja*2] + (nbat->energrp[cja*2+1] << (gridj->na_c*egp_shift));
3395 }
3396 }
3397
3398 if (bFEP_i || fep_cj != 0)
3399 {
3400 for (j = 0; j < nbl->na_cj; j++)
3401 {
3402 /* Is this interaction perturbed and not excluded? */
3403 ind_j = cja*nbl->na_cj + j;
3404 aj = nbs->a[ind_j];
3405 if (aj >= 0 &&
3406 (bFEP_i || (fep_cj & (1 << j))) &&
3407 (!bDiagRemoved || ind_j >= ind_i))
3408 {
3409 if (ngid > 1)
3410 {
3411 gid_j = (gid_cj >> (j*egp_shift)) & egp_mask;
3412 gid = GID(gid_i, gid_j, ngid)((gid_i < gid_j) ? (gid_i*ngid+gid_j) : (gid_j*ngid+gid_i)
);
3413
3414 if (nlist->nrj > nlist->jindex[nri] &&
3415 nlist->gid[nri] != gid)
3416 {
3417 /* Energy group pair changed: new list */
3418 fep_list_new_nri_copy(nlist);
3419 nri = nlist->nri;
3420 }
3421 nlist->gid[nri] = gid;
3422 }
3423
3424 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
3425 {
3426 fep_list_new_nri_copy(nlist);
3427 nri = nlist->nri;
3428 }
3429
3430 /* Add it to the FEP list */
3431 nlist->jjnr[nlist->nrj] = aj;
3432 nlist->excl_fep[nlist->nrj] = (nbl->cj[cj_ind].excl >> (i*nbl->na_cj + j)) & 1;
3433 nlist->nrj++;
3434
3435 /* Exclude it from the normal list.
3436 * Note that the charge has been set to zero,
3437 * but we need to avoid 0/0, as perturbed atoms
3438 * can be on top of each other.
3439 * (and the LJ parameters have not been zeroed)
3440 */
3441 nbl->cj[cj_ind].excl &= ~(1U << (i*nbl->na_cj + j));
3442 }
3443 }
3444 }
3445 }
3446
3447 if (nlist->nrj > nlist->jindex[nri])
3448 {
3449 nlist->nri++;
3450 nlist->jindex[nlist->nri] = nlist->nrj;
3451 }
3452 }
3453 }
3454
3455 if (bFEP_i_all)
3456 {
3457 /* All interactions are perturbed, we can skip this entry */
3458 nbl_ci->cj_ind_end = cj_ind_start;
3459 }
3460}
3461
3462/* Return the index of atom a within a cluster */
3463static gmx_inlineinline int cj_mod_cj4(int cj)
3464{
3465 return cj & (NBNXN_GPU_JGROUP_SIZE4 - 1);
3466}
3467
3468/* Convert a j-cluster to a cj4 group */
3469static gmx_inlineinline int cj_to_cj4(int cj)
3470{
3471 return cj >> NBNXN_GPU_JGROUP_SIZE_2LOG2;
3472}
3473
3474/* Return the index of an j-atom within a warp */
3475static gmx_inlineinline int a_mod_wj(int a)
3476{
3477 return a & (NBNXN_GPU_CLUSTER_SIZE8/2 - 1);
3478}
3479
3480/* As make_fep_list above, but for super/sub lists. */
3481static void make_fep_list_supersub(const nbnxn_search_t nbs,
3482 const nbnxn_atomdata_t *nbat,
3483 nbnxn_pairlist_t *nbl,
3484 gmx_bool bDiagRemoved,
3485 const nbnxn_sci_t *nbl_sci,
3486 real shx,
3487 real shy,
3488 real shz,
3489 real rlist_fep2,
3490 const nbnxn_grid_t *gridi,
3491 const nbnxn_grid_t *gridj,
3492 t_nblist *nlist)
3493{
3494 int sci, cj4_ind_start, cj4_ind_end, cj4_ind, gcj, cjr;
3495 int nri_max;
3496 int c, c_abs;
3497 int i, j, ind_i, ind_j, ai, aj;
3498 int nri;
3499 gmx_bool bFEP_i;
3500 real xi, yi, zi;
3501 const nbnxn_cj4_t *cj4;
3502
3503 if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
3504 {
3505 /* Empty list */
3506 return;
3507 }
3508
3509 sci = nbl_sci->sci;
3510
3511 cj4_ind_start = nbl_sci->cj4_ind_start;
3512 cj4_ind_end = nbl_sci->cj4_ind_end;
3513
3514 /* No energy groups (yet), so we split lists in max_nrj_fep pairs */
3515 nri_max = nbl->na_sc*(1 + ((cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE4)/max_nrj_fep);
3516 if (nlist->nri + nri_max > nlist->maxnri)
3517 {
3518 nlist->maxnri = over_alloc_large(nlist->nri + nri_max)(int)(1.19*(nlist->nri + nri_max) + 1000);
3519 reallocate_nblist(nlist);
3520 }
3521
3522 /* Loop over the atoms in the i super-cluster */
3523 for (c = 0; c < GPU_NSUBCELL(2*2*2); c++)
3524 {
3525 c_abs = sci*GPU_NSUBCELL(2*2*2) + c;
3526
3527 for (i = 0; i < nbl->na_ci; i++)
3528 {
3529 ind_i = c_abs*nbl->na_ci + i;
3530 ai = nbs->a[ind_i];
3531 if (ai >= 0)
3532 {
3533 nri = nlist->nri;
3534 nlist->jindex[nri+1] = nlist->jindex[nri];
3535 nlist->iinr[nri] = ai;
3536 /* With GPUs, energy groups are not supported */
3537 nlist->gid[nri] = 0;
3538 nlist->shift[nri] = nbl_sci->shift & NBNXN_CI_SHIFT127;
3539
3540 bFEP_i = (gridi->fep[c_abs - gridi->cell0] & (1 << i));
3541
3542 xi = nbat->x[ind_i*nbat->xstride+XX0] + shx;
3543 yi = nbat->x[ind_i*nbat->xstride+YY1] + shy;
3544 zi = nbat->x[ind_i*nbat->xstride+ZZ2] + shz;
3545
3546 if ((nlist->nrj + cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE4*nbl->na_cj > nlist->maxnrj)
3547 {
3548 nlist->maxnrj = over_alloc_small((nlist->nrj + cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE*nbl->na_cj)(int)(1.19*((nlist->nrj + cj4_ind_end - cj4_ind_start)*4*nbl
->na_cj) + 8000);
3549 srenew(nlist->jjnr, nlist->maxnrj)(nlist->jjnr) = save_realloc("nlist->jjnr", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 3549, (nlist->jjnr), (nlist->maxnrj), sizeof(*(nlist->
jjnr)));
3550 srenew(nlist->excl_fep, nlist->maxnrj)(nlist->excl_fep) = save_realloc("nlist->excl_fep", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 3550, (nlist->excl_fep), (nlist->maxnrj), sizeof(*(nlist
->excl_fep)));
3551 }
3552
3553 for (cj4_ind = cj4_ind_start; cj4_ind < cj4_ind_end; cj4_ind++)
3554 {
3555 cj4 = &nbl->cj4[cj4_ind];
3556
3557 for (gcj = 0; gcj < NBNXN_GPU_JGROUP_SIZE4; gcj++)
3558 {
3559 unsigned int fep_cj;
3560
3561 if ((cj4->imei[0].imask & (1U << (gcj*GPU_NSUBCELL(2*2*2) + c))) == 0)
3562 {
3563 /* Skip this ci for this cj */
3564 continue;
3565 }
3566
3567 cjr = cj4->cj[gcj] - gridj->cell0*GPU_NSUBCELL(2*2*2);
3568
3569 fep_cj = gridj->fep[cjr];
3570
3571 if (bFEP_i || fep_cj != 0)
3572 {
3573 for (j = 0; j < nbl->na_cj; j++)
3574 {
3575 /* Is this interaction perturbed and not excluded? */
3576 ind_j = (gridj->cell0*GPU_NSUBCELL(2*2*2) + cjr)*nbl->na_cj + j;
3577 aj = nbs->a[ind_j];
3578 if (aj >= 0 &&
3579 (bFEP_i || (fep_cj & (1 << j))) &&
3580 (!bDiagRemoved || ind_j >= ind_i))
3581 {
3582 nbnxn_excl_t *excl;
3583 int excl_pair;
3584 unsigned int excl_bit;
3585 real dx, dy, dz;
3586
3587 get_nbl_exclusions_1(nbl, cj4_ind, j>>2, &excl);
3588
3589 excl_pair = a_mod_wj(j)*nbl->na_ci + i;
3590 excl_bit = (1U << (gcj*GPU_NSUBCELL(2*2*2) + c));
3591
3592 dx = nbat->x[ind_j*nbat->xstride+XX0] - xi;
3593 dy = nbat->x[ind_j*nbat->xstride+YY1] - yi;
3594 dz = nbat->x[ind_j*nbat->xstride+ZZ2] - zi;
3595
3596 /* The unpruned GPU list has more than 2/3
3597 * of the atom pairs beyond rlist. Using
3598 * this list will cause a lot of overhead
3599 * in the CPU FEP kernels, especially
3600 * relative to the fast GPU kernels.
3601 * So we prune the FEP list here.
3602 */
3603 if (dx*dx + dy*dy + dz*dz < rlist_fep2)
3604 {
3605 if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
3606 {
3607 fep_list_new_nri_copy(nlist);
3608 nri = nlist->nri;
3609 }
3610
3611 /* Add it to the FEP list */
3612 nlist->jjnr[nlist->nrj] = aj;
3613 nlist->excl_fep[nlist->nrj] = (excl->pair[excl_pair] & excl_bit) ? 1 : 0;
3614 nlist->nrj++;
3615 }
3616
3617 /* Exclude it from the normal list.
3618 * Note that the charge and LJ parameters have
3619 * been set to zero, but we need to avoid 0/0,
3620 * as perturbed atoms can be on top of each other.
3621 */
3622 excl->pair[excl_pair] &= ~excl_bit;
3623 }
3624 }
3625
3626 /* Note that we could mask out this pair in imask
3627 * if all i- and/or all j-particles are perturbed.
3628 * But since the perturbed pairs on the CPU will
3629 * take an order of magnitude more time, the GPU
3630 * will finish before the CPU and there is no gain.
3631 */
3632 }
3633 }
3634 }
3635
3636 if (nlist->nrj > nlist->jindex[nri])
3637 {
3638 nlist->nri++;
3639 nlist->jindex[nlist->nri] = nlist->nrj;
3640 }
3641 }
3642 }
3643 }
3644}
3645
3646/* Set all atom-pair exclusions from the topology stored in excl
3647 * as masks in the pair-list for i-super-cell entry nbl_sci
3648 */
3649static void set_sci_top_excls(const nbnxn_search_t nbs,
3650 nbnxn_pairlist_t *nbl,
3651 gmx_bool diagRemoved,
3652 int na_c_2log,
3653 const nbnxn_sci_t *nbl_sci,
3654 const t_blocka *excl)
3655{
3656 const int *cell;
3657 int na_c;
3658 int sci;
3659 int cj_ind_first, cj_ind_last;
3660 int cj_first, cj_last;
3661 int ndirect;
3662 int i, ai, aj, si, eind, ge, se;
3663 int found, cj_ind_0, cj_ind_1, cj_ind_m;
3664 int cj_m;
3665 gmx_bool Found_si;
3666 int si_ind;
3667 nbnxn_excl_t *nbl_excl;
3668 int inner_i, inner_e, w;
3669
3670 cell = nbs->cell;
3671
3672 na_c = nbl->na_ci;
3673
3674 if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
3675 {
3676 /* Empty list */
3677 return;
3678 }
3679
3680 sci = nbl_sci->sci;
3681
3682 cj_ind_first = nbl_sci->cj4_ind_start*NBNXN_GPU_JGROUP_SIZE4;
3683 cj_ind_last = nbl->work->cj_ind - 1;
3684
3685 cj_first = nbl->cj4[nbl_sci->cj4_ind_start].cj[0];
3686 cj_last = nbl_cj(nbl, cj_ind_last);
3687
3688 /* Determine how many contiguous j-clusters we have starting
3689 * from the first i-cluster. This number can be used to directly
3690 * calculate j-cluster indices for excluded atoms.
3691 */
3692 ndirect = 0;
3693 while (cj_ind_first + ndirect <= cj_ind_last &&
3694 nbl_cj(nbl, cj_ind_first+ndirect) == sci*GPU_NSUBCELL(2*2*2) + ndirect)
3695 {
3696 ndirect++;
3697 }
3698
3699 /* Loop over the atoms in the i super-cell */
3700 for (i = 0; i < nbl->na_sc; i++)
3701 {
3702 ai = nbs->a[sci*nbl->na_sc+i];
3703 if (ai >= 0)
3704 {
3705 si = (i>>na_c_2log);
3706
3707 /* Loop over the topology-based exclusions for this i-atom */
3708 for (eind = excl->index[ai]; eind < excl->index[ai+1]; eind++)
3709 {
3710 aj = excl->a[eind];
3711
3712 if (aj == ai)
3713 {
3714 /* The self exclusion are already set, save some time */
3715 continue;
3716 }
3717
3718 ge = cell[aj];
3719
3720 /* Without shifts we only calculate interactions j>i
3721 * for one-way pair-lists.
3722 */
3723 if (diagRemoved && ge <= sci*nbl->na_sc + i)
3724 {
3725 continue;
3726 }
3727
3728 se = ge>>na_c_2log;
3729 /* Could the cluster se be in our list? */
3730 if (se >= cj_first && se <= cj_last)
3731 {
3732 if (se < cj_first + ndirect)
3733 {
3734 /* We can calculate cj_ind directly from se */
3735 found = cj_ind_first + se - cj_first;
3736 }
3737 else
3738 {
3739 /* Search for se using bisection */
3740 found = -1;
3741 cj_ind_0 = cj_ind_first + ndirect;
3742 cj_ind_1 = cj_ind_last + 1;
3743 while (found == -1 && cj_ind_0 < cj_ind_1)
3744 {
3745 cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
3746
3747 cj_m = nbl_cj(nbl, cj_ind_m);
3748
3749 if (se == cj_m)
3750 {
3751 found = cj_ind_m;
3752 }
3753 else if (se < cj_m)
3754 {
3755 cj_ind_1 = cj_ind_m;
3756 }
3757 else
3758 {
3759 cj_ind_0 = cj_ind_m + 1;
3760 }
3761 }
3762 }
3763
3764 if (found >= 0)
3765 {
3766 inner_i = i - si*na_c;
3767 inner_e = ge - se*na_c;
3768
3769 if (nbl_imask0(nbl, found) & (1U << (cj_mod_cj4(found)*GPU_NSUBCELL(2*2*2) + si)))
3770 {
3771 w = (inner_e >> 2);
3772
3773 get_nbl_exclusions_1(nbl, cj_to_cj4(found), w, &nbl_excl);
3774
3775 nbl_excl->pair[a_mod_wj(inner_e)*nbl->na_ci+inner_i] &=
3776 ~(1U << (cj_mod_cj4(found)*GPU_NSUBCELL(2*2*2) + si));
3777 }
3778 }
3779 }
3780 }
3781 }
3782 }
3783}
3784
3785/* Reallocate the simple ci list for at least n entries */
3786static void nb_realloc_ci(nbnxn_pairlist_t *nbl, int n)
3787{
3788 nbl->ci_nalloc = over_alloc_small(n)(int)(1.19*(n) + 8000);
3789 nbnxn_realloc_void((void **)&nbl->ci,
3790 nbl->nci*sizeof(*nbl->ci),
3791 nbl->ci_nalloc*sizeof(*nbl->ci),
3792 nbl->alloc, nbl->free);
3793}
3794
3795/* Reallocate the super-cell sci list for at least n entries */
3796static void nb_realloc_sci(nbnxn_pairlist_t *nbl, int n)
3797{
3798 nbl->sci_nalloc = over_alloc_small(n)(int)(1.19*(n) + 8000);
3799 nbnxn_realloc_void((void **)&nbl->sci,
3800 nbl->nsci*sizeof(*nbl->sci),
3801 nbl->sci_nalloc*sizeof(*nbl->sci),
3802 nbl->alloc, nbl->free);
3803}
3804
3805/* Make a new ci entry at index nbl->nci */
3806static void new_ci_entry(nbnxn_pairlist_t *nbl, int ci, int shift, int flags)
3807{
3808 if (nbl->nci + 1 > nbl->ci_nalloc)
3809 {
3810 nb_realloc_ci(nbl, nbl->nci+1);
3811 }
3812 nbl->ci[nbl->nci].ci = ci;
3813 nbl->ci[nbl->nci].shift = shift;
3814 /* Store the interaction flags along with the shift */
3815 nbl->ci[nbl->nci].shift |= flags;
3816 nbl->ci[nbl->nci].cj_ind_start = nbl->ncj;
3817 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
3818}
3819
3820/* Make a new sci entry at index nbl->nsci */
3821static void new_sci_entry(nbnxn_pairlist_t *nbl, int sci, int shift)
3822{
3823 if (nbl->nsci + 1 > nbl->sci_nalloc)
3824 {
3825 nb_realloc_sci(nbl, nbl->nsci+1);
3826 }
3827 nbl->sci[nbl->nsci].sci = sci;
3828 nbl->sci[nbl->nsci].shift = shift;
3829 nbl->sci[nbl->nsci].cj4_ind_start = nbl->ncj4;
3830 nbl->sci[nbl->nsci].cj4_ind_end = nbl->ncj4;
3831}
3832
3833/* Sort the simple j-list cj on exclusions.
3834 * Entries with exclusions will all be sorted to the beginning of the list.
3835 */
3836static void sort_cj_excl(nbnxn_cj_t *cj, int ncj,
3837 nbnxn_list_work_t *work)
3838{
3839 int jnew, j;
3840
3841 if (ncj > work->cj_nalloc)
3842 {
3843 work->cj_nalloc = over_alloc_large(ncj)(int)(1.19*(ncj) + 1000);
3844 srenew(work->cj, work->cj_nalloc)(work->cj) = save_realloc("work->cj", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 3844, (work->cj), (work->cj_nalloc), sizeof(*(work->
cj)));
3845 }
3846
3847 /* Make a list of the j-cells involving exclusions */
3848 jnew = 0;
3849 for (j = 0; j < ncj; j++)
3850 {
3851 if (cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
3852 {
3853 work->cj[jnew++] = cj[j];
3854 }
3855 }
3856 /* Check if there are exclusions at all or not just the first entry */
3857 if (!((jnew == 0) ||
3858 (jnew == 1 && cj[0].excl != NBNXN_INTERACTION_MASK_ALL)))
3859 {
3860 for (j = 0; j < ncj; j++)
3861 {
3862 if (cj[j].excl == NBNXN_INTERACTION_MASK_ALL)
3863 {
3864 work->cj[jnew++] = cj[j];
3865 }
3866 }
3867 for (j = 0; j < ncj; j++)
3868 {
3869 cj[j] = work->cj[j];
3870 }
3871 }
3872}
3873
3874/* Close this simple list i entry */
3875static void close_ci_entry_simple(nbnxn_pairlist_t *nbl)
3876{
3877 int jlen;
3878
3879 /* All content of the new ci entry have already been filled correctly,
3880 * we only need to increase the count here (for non empty lists).
3881 */
3882 jlen = nbl->ci[nbl->nci].cj_ind_end - nbl->ci[nbl->nci].cj_ind_start;
3883 if (jlen > 0)
3884 {
3885 sort_cj_excl(nbl->cj+nbl->ci[nbl->nci].cj_ind_start, jlen, nbl->work);
3886
3887 /* The counts below are used for non-bonded pair/flop counts
3888 * and should therefore match the available kernel setups.
3889 */
3890 if (!(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_COUL(0)(1<<(9+3*(0)))))
3891 {
3892 nbl->work->ncj_noq += jlen;
3893 }
3894 else if ((nbl->ci[nbl->nci].shift & NBNXN_CI_HALF_LJ(0)(1<<(8+3*(0)))) ||
3895 !(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_LJ(0)(1<<(7+3*(0)))))
3896 {
3897 nbl->work->ncj_hlj += jlen;
3898 }
3899
3900 nbl->nci++;
3901 }
3902}
3903
3904/* Split sci entry for load balancing on the GPU.
3905 * Splitting ensures we have enough lists to fully utilize the whole GPU.
3906 * With progBal we generate progressively smaller lists, which improves
3907 * load balancing. As we only know the current count on our own thread,
3908 * we will need to estimate the current total amount of i-entries.
3909 * As the lists get concatenated later, this estimate depends
3910 * both on nthread and our own thread index.
3911 */
3912static void split_sci_entry(nbnxn_pairlist_t *nbl,
3913 int nsp_max_av, gmx_bool progBal, int nc_bal,
3914 int thread, int nthread)
3915{
3916 int nsci_est;
3917 int nsp_max;
3918 int cj4_start, cj4_end, j4len, cj4;
3919 int sci;
3920 int nsp, nsp_sci, nsp_cj4, nsp_cj4_e, nsp_cj4_p;
3921 int p;
3922
3923 if (progBal)
3924 {
3925 /* Estimate the total numbers of ci's of the nblist combined
3926 * over all threads using the target number of ci's.
3927 */
3928 nsci_est = nc_bal*thread/nthread + nbl->nsci;
3929
3930 /* The first ci blocks should be larger, to avoid overhead.
3931 * The last ci blocks should be smaller, to improve load balancing.
3932 */
3933 nsp_max = max(1,(((1) > (nsp_max_av*nc_bal*3/(2*(nsci_est - 1 + nc_bal))))
? (1) : (nsp_max_av*nc_bal*3/(2*(nsci_est - 1 + nc_bal))) )
3934 nsp_max_av*nc_bal*3/(2*(nsci_est - 1 + nc_bal)))(((1) > (nsp_max_av*nc_bal*3/(2*(nsci_est - 1 + nc_bal))))
? (1) : (nsp_max_av*nc_bal*3/(2*(nsci_est - 1 + nc_bal))) );
3935 }
3936 else
3937 {
3938 nsp_max = nsp_max_av;
3939 }
3940
3941 cj4_start = nbl->sci[nbl->nsci-1].cj4_ind_start;
3942 cj4_end = nbl->sci[nbl->nsci-1].cj4_ind_end;
3943 j4len = cj4_end - cj4_start;
3944
3945 if (j4len > 1 && j4len*GPU_NSUBCELL(2*2*2)*NBNXN_GPU_JGROUP_SIZE4 > nsp_max)
3946 {
3947 /* Remove the last ci entry and process the cj4's again */
3948 nbl->nsci -= 1;
3949
3950 sci = nbl->nsci;
3951 nsp = 0;
3952 nsp_sci = 0;
3953 nsp_cj4_e = 0;
3954 nsp_cj4 = 0;
3955 for (cj4 = cj4_start; cj4 < cj4_end; cj4++)
3956 {
3957 nsp_cj4_p = nsp_cj4;
3958 /* Count the number of cluster pairs in this cj4 group */
3959 nsp_cj4 = 0;
3960 for (p = 0; p < GPU_NSUBCELL(2*2*2)*NBNXN_GPU_JGROUP_SIZE4; p++)
3961 {
3962 nsp_cj4 += (nbl->cj4[cj4].imei[0].imask >> p) & 1;
3963 }
3964
3965 if (nsp_cj4 > 0 && nsp + nsp_cj4 > nsp_max)
3966 {
3967 /* Split the list at cj4 */
3968 nbl->sci[sci].cj4_ind_end = cj4;
3969 /* Create a new sci entry */
3970 sci++;
3971 nbl->nsci++;
3972 if (nbl->nsci+1 > nbl->sci_nalloc)
3973 {
3974 nb_realloc_sci(nbl, nbl->nsci+1);
3975 }
3976 nbl->sci[sci].sci = nbl->sci[nbl->nsci-1].sci;
3977 nbl->sci[sci].shift = nbl->sci[nbl->nsci-1].shift;
3978 nbl->sci[sci].cj4_ind_start = cj4;
3979 nsp_sci = nsp;
3980 nsp_cj4_e = nsp_cj4_p;
3981 nsp = 0;
3982 }
3983 nsp += nsp_cj4;
3984 }
3985
3986 /* Put the remaining cj4's in the last sci entry */
3987 nbl->sci[sci].cj4_ind_end = cj4_end;
3988
3989 /* Possibly balance out the last two sci's
3990 * by moving the last cj4 of the second last sci.
3991 */
3992 if (nsp_sci - nsp_cj4_e >= nsp + nsp_cj4_e)
3993 {
3994 nbl->sci[sci-1].cj4_ind_end--;
3995 nbl->sci[sci].cj4_ind_start--;
3996 }
3997
3998 nbl->nsci++;
3999 }
4000}
4001
4002/* Clost this super/sub list i entry */
4003static void close_ci_entry_supersub(nbnxn_pairlist_t *nbl,
4004 int nsp_max_av,
4005 gmx_bool progBal, int nc_bal,
4006 int thread, int nthread)
4007{
4008 int j4len, tlen;
4009 int nb, b;
4010
4011 /* All content of the new ci entry have already been filled correctly,
4012 * we only need to increase the count here (for non empty lists).
4013 */
4014 j4len = nbl->sci[nbl->nsci].cj4_ind_end - nbl->sci[nbl->nsci].cj4_ind_start;
4015 if (j4len > 0)
4016 {
4017 /* We can only have complete blocks of 4 j-entries in a list,
4018 * so round the count up before closing.
4019 */
4020 nbl->ncj4 = ((nbl->work->cj_ind + NBNXN_GPU_JGROUP_SIZE4 - 1) >> NBNXN_GPU_JGROUP_SIZE_2LOG2);
4021 nbl->work->cj_ind = nbl->ncj4*NBNXN_GPU_JGROUP_SIZE4;
4022
4023 nbl->nsci++;
4024
4025 if (nsp_max_av > 0)
4026 {
4027 /* Measure the size of the new entry and potentially split it */
4028 split_sci_entry(nbl, nsp_max_av, progBal, nc_bal, thread, nthread);
4029 }
4030 }
4031}
4032
4033/* Syncs the working array before adding another grid pair to the list */
4034static void sync_work(nbnxn_pairlist_t *nbl)
4035{
4036 if (!nbl->bSimple)
4037 {
4038 nbl->work->cj_ind = nbl->ncj4*NBNXN_GPU_JGROUP_SIZE4;
4039 nbl->work->cj4_init = nbl->ncj4;
4040 }
4041}
4042
4043/* Clears an nbnxn_pairlist_t data structure */
4044static void clear_pairlist(nbnxn_pairlist_t *nbl)
4045{
4046 nbl->nci = 0;
4047 nbl->nsci = 0;
4048 nbl->ncj = 0;
4049 nbl->ncj4 = 0;
4050 nbl->nci_tot = 0;
4051 nbl->nexcl = 1;
4052
4053 nbl->work->ncj_noq = 0;
4054 nbl->work->ncj_hlj = 0;
4055}
4056
4057/* Clears a group scheme pair list */
4058static void clear_pairlist_fep(t_nblist *nl)
4059{
4060 nl->nri = 0;
4061 nl->nrj = 0;
4062 if (nl->jindex == NULL((void*)0))
4063 {
4064 snew(nl->jindex, 1)(nl->jindex) = save_calloc("nl->jindex", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 4064, (1), sizeof(*(nl->jindex)));
4065 }
4066 nl->jindex[0] = 0;
4067}
4068
4069/* Sets a simple list i-cell bounding box, including PBC shift */
4070static gmx_inlineinline void set_icell_bb_simple(const nbnxn_bb_t *bb, int ci,
4071 real shx, real shy, real shz,
4072 nbnxn_bb_t *bb_ci)
4073{
4074 bb_ci->lower[BB_X0] = bb[ci].lower[BB_X0] + shx;
4075 bb_ci->lower[BB_Y1] = bb[ci].lower[BB_Y1] + shy;
4076 bb_ci->lower[BB_Z2] = bb[ci].lower[BB_Z2] + shz;
4077 bb_ci->upper[BB_X0] = bb[ci].upper[BB_X0] + shx;
4078 bb_ci->upper[BB_Y1] = bb[ci].upper[BB_Y1] + shy;
4079 bb_ci->upper[BB_Z2] = bb[ci].upper[BB_Z2] + shz;
4080}
4081
4082#ifdef NBNXN_BBXXXX
4083/* Sets a super-cell and sub cell bounding boxes, including PBC shift */
4084static void set_icell_bbxxxx_supersub(const float *bb, int ci,
4085 real shx, real shy, real shz,
4086 float *bb_ci)
4087{
4088 int ia, m, i;
4089
4090 ia = ci*(GPU_NSUBCELL(2*2*2)>>STRIDE_PBB_2LOG2)*NNBSBB_XXXX(2*3*4);
4091 for (m = 0; m < (GPU_NSUBCELL(2*2*2)>>STRIDE_PBB_2LOG2)*NNBSBB_XXXX(2*3*4); m += NNBSBB_XXXX(2*3*4))
77
←
Loop condition is true. Entering loop body
→
4092 {
4093 for (i = 0; i < STRIDE_PBB4; i++)
78
←
Loop condition is true. Entering loop body
→
4094 {
4095 bb_ci[m+0*STRIDE_PBB4+i] = bb[ia+m+0*STRIDE_PBB4+i] + shx;
79
←
Array access (from variable 'bb') results in a null pointer dereference
4096 bb_ci[m+1*STRIDE_PBB4+i] = bb[ia+m+1*STRIDE_PBB4+i] + shy;
4097 bb_ci[m+2*STRIDE_PBB4+i] = bb[ia+m+2*STRIDE_PBB4+i] + shz;
4098 bb_ci[m+3*STRIDE_PBB4+i] = bb[ia+m+3*STRIDE_PBB4+i] + shx;
4099 bb_ci[m+4*STRIDE_PBB4+i] = bb[ia+m+4*STRIDE_PBB4+i] + shy;
4100 bb_ci[m+5*STRIDE_PBB4+i] = bb[ia+m+5*STRIDE_PBB4+i] + shz;
4101 }
4102 }
4103}
4104#endif
4105
4106/* Sets a super-cell and sub cell bounding boxes, including PBC shift */
4107static void set_icell_bb_supersub(const nbnxn_bb_t *bb, int ci,
4108 real shx, real shy, real shz,
4109 nbnxn_bb_t *bb_ci)
4110{
4111 int i;
4112
4113 for (i = 0; i < GPU_NSUBCELL(2*2*2); i++)
4114 {
4115 set_icell_bb_simple(bb, ci*GPU_NSUBCELL(2*2*2)+i,
4116 shx, shy, shz,
4117 &bb_ci[i]);
4118 }
4119}
4120
4121/* Copies PBC shifted i-cell atom coordinates x,y,z to working array */
4122static void icell_set_x_simple(int ci,
4123 real shx, real shy, real shz,
4124 int gmx_unused__attribute__ ((unused)) na_c,
4125 int stride, const real *x,
4126 nbnxn_list_work_t *work)
4127{
4128 int ia, i;
4129
4130 ia = ci*NBNXN_CPU_CLUSTER_I_SIZE4;
4131
4132 for (i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE4; i++)
4133 {
4134 work->x_ci[i*STRIDE_XYZ3+XX0] = x[(ia+i)*stride+XX0] + shx;
4135 work->x_ci[i*STRIDE_XYZ3+YY1] = x[(ia+i)*stride+YY1] + shy;
4136 work->x_ci[i*STRIDE_XYZ3+ZZ2] = x[(ia+i)*stride+ZZ2] + shz;
4137 }
4138}
4139
4140/* Copies PBC shifted super-cell atom coordinates x,y,z to working array */
4141static void icell_set_x_supersub(int ci,
4142 real shx, real shy, real shz,
4143 int na_c,
4144 int stride, const real *x,
4145 nbnxn_list_work_t *work)
4146{
4147 int ia, i;
4148 real *x_ci;
4149
4150 x_ci = work->x_ci;
4151
4152 ia = ci*GPU_NSUBCELL(2*2*2)*na_c;
4153 for (i = 0; i < GPU_NSUBCELL(2*2*2)*na_c; i++)
4154 {
4155 x_ci[i*DIM3 + XX0] = x[(ia+i)*stride + XX0] + shx;
4156 x_ci[i*DIM3 + YY1] = x[(ia+i)*stride + YY1] + shy;
4157 x_ci[i*DIM3 + ZZ2] = x[(ia+i)*stride + ZZ2] + shz;
4158 }
4159}
4160
4161#ifdef NBNXN_SEARCH_BB_SIMD4
4162/* Copies PBC shifted super-cell packed atom coordinates to working array */
4163static void icell_set_x_supersub_simd4(int ci,
4164 real shx, real shy, real shz,
4165 int na_c,
4166 int stride, const real *x,
4167 nbnxn_list_work_t *work)
4168{
4169 int si, io, ia, i, j;
4170 real *x_ci;
4171
4172 x_ci = work->x_ci;
4173
4174 for (si = 0; si < GPU_NSUBCELL(2*2*2); si++)
4175 {
4176 for (i = 0; i < na_c; i += STRIDE_PBB4)
4177 {
4178 io = si*na_c + i;
4179 ia = ci*GPU_NSUBCELL(2*2*2)*na_c + io;
4180 for (j = 0; j < STRIDE_PBB4; j++)
4181 {
4182 x_ci[io*DIM3 + j + XX0*STRIDE_PBB4] = x[(ia+j)*stride+XX0] + shx;
4183 x_ci[io*DIM3 + j + YY1*STRIDE_PBB4] = x[(ia+j)*stride+YY1] + shy;
4184 x_ci[io*DIM3 + j + ZZ2*STRIDE_PBB4] = x[(ia+j)*stride+ZZ2] + shz;
4185 }
4186 }
4187 }
4188}
4189#endif
4190
4191static real minimum_subgrid_size_xy(const nbnxn_grid_t *grid)
4192{
4193 if (grid->bSimple)
4194 {
4195 return min(grid->sx, grid->sy)(((grid->sx) < (grid->sy)) ? (grid->sx) : (grid->
sy) );
4196 }
4197 else
4198 {
4199 return min(grid->sx/GPU_NSUBCELL_X, grid->sy/GPU_NSUBCELL_Y)(((grid->sx/2) < (grid->sy/2)) ? (grid->sx/2) : (
grid->sy/2) );
4200 }
4201}
4202
4203static real effective_buffer_1x1_vs_MxN(const nbnxn_grid_t *gridi,
4204 const nbnxn_grid_t *gridj)
4205{
4206 const real eff_1x1_buffer_fac_overest = 0.1;
4207
4208 /* Determine an atom-pair list cut-off buffer size for atom pairs,
4209 * to be added to rlist (including buffer) used for MxN.
4210 * This is for converting an MxN list to a 1x1 list. This means we can't
4211 * use the normal buffer estimate, as we have an MxN list in which
4212 * some atom pairs beyond rlist are missing. We want to capture
4213 * the beneficial effect of buffering by extra pairs just outside rlist,
4214 * while removing the useless pairs that are further away from rlist.
4215 * (Also the buffer could have been set manually not using the estimate.)
4216 * This buffer size is an overestimate.
4217 * We add 10% of the smallest grid sub-cell dimensions.
4218 * Note that the z-size differs per cell and we don't use this,
4219 * so we overestimate.
4220 * With PME, the 10% value gives a buffer that is somewhat larger
4221 * than the effective buffer with a tolerance of 0.005 kJ/mol/ps.
4222 * Smaller tolerances or using RF lead to a smaller effective buffer,
4223 * so 10% gives a safe overestimate.
4224 */
4225 return eff_1x1_buffer_fac_overest*(minimum_subgrid_size_xy(gridi) +
4226 minimum_subgrid_size_xy(gridj));
4227}
4228
4229/* Clusters at the cut-off only increase rlist by 60% of their size */
4230static real nbnxn_rlist_inc_outside_fac = 0.6;
4231
4232/* Due to the cluster size the effective pair-list is longer than
4233 * that of a simple atom pair-list. This function gives the extra distance.
4234 */
4235real nbnxn_get_rlist_effective_inc(int cluster_size_j, real atom_density)
4236{
4237 int cluster_size_i;
4238 real vol_inc_i, vol_inc_j;
4239
4240 /* We should get this from the setup, but currently it's the same for
4241 * all setups, including GPUs.
4242 */
4243 cluster_size_i = NBNXN_CPU_CLUSTER_I_SIZE4;
4244
4245 vol_inc_i = (cluster_size_i - 1)/atom_density;
4246 vol_inc_j = (cluster_size_j - 1)/atom_density;
4247
4248 return nbnxn_rlist_inc_outside_fac*pow(vol_inc_i + vol_inc_j, 1.0/3.0);
4249}
4250
4251/* Estimates the interaction volume^2 for non-local interactions */
4252static real nonlocal_vol2(const gmx_domdec_zones_t *zones, rvec ls, real r)
4253{
4254 int z, d;
4255 real cl, ca, za;
4256 real vold_est;
4257 real vol2_est_tot;
4258
4259 vol2_est_tot = 0;
4260
4261 /* Here we simply add up the volumes of 1, 2 or 3 1D decomposition
4262 * not home interaction volume^2. As these volumes are not additive,
4263 * this is an overestimate, but it would only be significant in the limit
4264 * of small cells, where we anyhow need to split the lists into
4265 * as small parts as possible.
4266 */
4267
4268 for (z = 0; z < zones->n; z++)
4269 {
4270 if (zones->shift[z][XX0] + zones->shift[z][YY1] + zones->shift[z][ZZ2] == 1)
4271 {
4272 cl = 0;
4273 ca = 1;
4274 za = 1;
4275 for (d = 0; d < DIM3; d++)
4276 {
4277 if (zones->shift[z][d] == 0)
4278 {
4279 cl += 0.5*ls[d];
4280 ca *= ls[d];
4281 za *= zones->size[z].x1[d] - zones->size[z].x0[d];
4282 }
4283 }
4284
4285 /* 4 octants of a sphere */
4286 vold_est = 0.25*M_PI3.14159265358979323846*r*r*r*r;
4287 /* 4 quarter pie slices on the edges */
4288 vold_est += 4*cl*M_PI3.14159265358979323846/6.0*r*r*r;
4289 /* One rectangular volume on a face */
4290 vold_est += ca*0.5*r*r;
4291
4292 vol2_est_tot += vold_est*za;
4293 }
4294 }
4295
4296 return vol2_est_tot;
4297}
4298
4299/* Estimates the average size of a full j-list for super/sub setup */
4300static int get_nsubpair_max(const nbnxn_search_t nbs,
4301 int iloc,
4302 real rlist,
4303 int min_ci_balanced)
4304{
4305 const nbnxn_grid_t *grid;
4306 rvec ls;
4307 real xy_diag2, r_eff_sup, vol_est, nsp_est, nsp_est_nl;
4308 int nsubpair_max;
4309
4310 grid = &nbs->grid[0];
4311
4312 ls[XX0] = (grid->c1[XX0] - grid->c0[XX0])/(grid->ncx*GPU_NSUBCELL_X2);
4313 ls[YY1] = (grid->c1[YY1] - grid->c0[YY1])/(grid->ncy*GPU_NSUBCELL_Y2);
4314 ls[ZZ2] = (grid->c1[ZZ2] - grid->c0[ZZ2])*grid->ncx*grid->ncy/(grid->nc*GPU_NSUBCELL_Z2);
4315
4316 /* The average squared length of the diagonal of a sub cell */
4317 xy_diag2 = ls[XX0]*ls[XX0] + ls[YY1]*ls[YY1] + ls[ZZ2]*ls[ZZ2];
4318
4319 /* The formulas below are a heuristic estimate of the average nsj per si*/
4320 r_eff_sup = rlist + nbnxn_rlist_inc_outside_fac*sqr((grid->na_c - 1.0)/grid->na_c)*sqrt(xy_diag2/3);
4321
4322 if (!nbs->DomDec || nbs->zones->n == 1)
4323 {
4324 nsp_est_nl = 0;
4325 }
4326 else
4327 {
4328 nsp_est_nl =
4329 sqr(grid->atom_density/grid->na_c)*
4330 nonlocal_vol2(nbs->zones, ls, r_eff_sup);
4331 }
4332
4333 if (LOCAL_I(iloc)((iloc) == eintLocal))
4334 {
4335 /* Sub-cell interacts with itself */
4336 vol_est = ls[XX0]*ls[YY1]*ls[ZZ2];
4337 /* 6/2 rectangular volume on the faces */
4338 vol_est += (ls[XX0]*ls[YY1] + ls[XX0]*ls[ZZ2] + ls[YY1]*ls[ZZ2])*r_eff_sup;
4339 /* 12/2 quarter pie slices on the edges */
4340 vol_est += 2*(ls[XX0] + ls[YY1] + ls[ZZ2])*0.25*M_PI3.14159265358979323846*sqr(r_eff_sup);
4341 /* 4 octants of a sphere */
4342 vol_est += 0.5*4.0/3.0*M_PI3.14159265358979323846*pow(r_eff_sup, 3);
4343
4344 nsp_est = grid->nsubc_tot*vol_est*grid->atom_density/grid->na_c;
4345
4346 /* Subtract the non-local pair count */
4347 nsp_est -= nsp_est_nl;
4348
4349 if (debug)
4350 {
4351 fprintf(debug, "nsp_est local %5.1f non-local %5.1f\n",
4352 nsp_est, nsp_est_nl);
4353 }
4354 }
4355 else
4356 {
4357 nsp_est = nsp_est_nl;
4358 }
4359
4360 if (min_ci_balanced <= 0 || grid->nc >= min_ci_balanced || grid->nc == 0)
4361 {
4362 /* We don't need to worry */
4363 nsubpair_max = -1;
4364 }
4365 else
4366 {
4367 /* Thus the (average) maximum j-list size should be as follows */
4368 nsubpair_max = max(1, (int)(nsp_est/min_ci_balanced+0.5))(((1) > ((int)(nsp_est/min_ci_balanced+0.5))) ? (1) : ((int
)(nsp_est/min_ci_balanced+0.5)) );
4369
4370 /* Since the target value is a maximum (this avoids high outliers,
4371 * which lead to load imbalance), not average, we add half the
4372 * number of pairs in a cj4 block to get the average about right.
4373 */
4374 nsubpair_max += GPU_NSUBCELL(2*2*2)*NBNXN_GPU_JGROUP_SIZE4/2;
4375 }
4376
4377 if (debug)
4378 {
4379 fprintf(debug, "nbl nsp estimate %.1f, nsubpair_max %d\n",
4380 nsp_est, nsubpair_max);
4381 }
4382
4383 return nsubpair_max;
4384}
4385
4386/* Debug list print function */
4387static void print_nblist_ci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
4388{
4389 int i, j;
4390
4391 for (i = 0; i < nbl->nci; i++)
4392 {
4393 fprintf(fp, "ci %4d shift %2d ncj %3d\n",
4394 nbl->ci[i].ci, nbl->ci[i].shift,
4395 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start);
4396
4397 for (j = nbl->ci[i].cj_ind_start; j < nbl->ci[i].cj_ind_end; j++)
4398 {
4399 fprintf(fp, " cj %5d imask %x\n",
4400 nbl->cj[j].cj,
4401 nbl->cj[j].excl);
4402 }
4403 }
4404}
4405
4406/* Debug list print function */
4407static void print_nblist_sci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
4408{
4409 int i, j4, j, ncp, si;
4410
4411 for (i = 0; i < nbl->nsci; i++)
4412 {
4413 fprintf(fp, "ci %4d shift %2d ncj4 %2d\n",
4414 nbl->sci[i].sci, nbl->sci[i].shift,
4415 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start);
4416
4417 ncp = 0;
4418 for (j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
4419 {
4420 for (j = 0; j < NBNXN_GPU_JGROUP_SIZE4; j++)
4421 {
4422 fprintf(fp, " sj %5d imask %x\n",
4423 nbl->cj4[j4].cj[j],
4424 nbl->cj4[j4].imei[0].imask);
4425 for (si = 0; si < GPU_NSUBCELL(2*2*2); si++)
4426 {
4427 if (nbl->cj4[j4].imei[0].imask & (1U << (j*GPU_NSUBCELL(2*2*2) + si)))
4428 {
4429 ncp++;
4430 }
4431 }
4432 }
4433 }
4434 fprintf(fp, "ci %4d shift %2d ncj4 %2d ncp %3d\n",
4435 nbl->sci[i].sci, nbl->sci[i].shift,
4436 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start,
4437 ncp);
4438 }
4439}
4440
4441/* Combine pair lists *nbl generated on multiple threads nblc */
4442static void combine_nblists(int nnbl, nbnxn_pairlist_t **nbl,
4443 nbnxn_pairlist_t *nblc)
4444{
4445 int nsci, ncj4, nexcl;
4446 int n, i;
4447
4448 if (nblc->bSimple)
4449 {
4450 gmx_incons("combine_nblists does not support simple lists")_gmx_error("incons", "combine_nblists does not support simple lists"
, "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 4450);
4451 }
4452
4453 nsci = nblc->nsci;
4454 ncj4 = nblc->ncj4;
4455 nexcl = nblc->nexcl;
4456 for (i = 0; i < nnbl; i++)
4457 {
4458 nsci += nbl[i]->nsci;
4459 ncj4 += nbl[i]->ncj4;
4460 nexcl += nbl[i]->nexcl;
4461 }
4462
4463 if (nsci > nblc->sci_nalloc)
4464 {
4465 nb_realloc_sci(nblc, nsci);
4466 }
4467 if (ncj4 > nblc->cj4_nalloc)
4468 {
4469 nblc->cj4_nalloc = over_alloc_small(ncj4)(int)(1.19*(ncj4) + 8000);
4470 nbnxn_realloc_void((void **)&nblc->cj4,
4471 nblc->ncj4*sizeof(*nblc->cj4),
4472 nblc->cj4_nalloc*sizeof(*nblc->cj4),
4473 nblc->alloc, nblc->free);
4474 }
4475 if (nexcl > nblc->excl_nalloc)
4476 {
4477 nblc->excl_nalloc = over_alloc_small(nexcl)(int)(1.19*(nexcl) + 8000);
4478 nbnxn_realloc_void((void **)&nblc->excl,
4479 nblc->nexcl*sizeof(*nblc->excl),
4480 nblc->excl_nalloc*sizeof(*nblc->excl),
4481 nblc->alloc, nblc->free);
4482 }
4483
4484 /* Each thread should copy its own data to the combined arrays,
4485 * as otherwise data will go back and forth between different caches.
4486 */
4487#pragma omp parallel for num_threads(gmx_omp_nthreads_get(emntPairsearch)) schedule(static)
4488 for (n = 0; n < nnbl; n++)
4489 {
4490 int sci_offset;
4491 int cj4_offset;
4492 int ci_offset;
4493 int excl_offset;
4494 int i, j4;
4495 const nbnxn_pairlist_t *nbli;
4496
4497 /* Determine the offset in the combined data for our thread */
4498 sci_offset = nblc->nsci;
4499 cj4_offset = nblc->ncj4;
4500 ci_offset = nblc->nci_tot;
4501 excl_offset = nblc->nexcl;
4502
4503 for (i = 0; i < n; i++)
4504 {
4505 sci_offset += nbl[i]->nsci;
4506 cj4_offset += nbl[i]->ncj4;
4507 ci_offset += nbl[i]->nci_tot;
4508 excl_offset += nbl[i]->nexcl;
4509 }
4510
4511 nbli = nbl[n];
4512
4513 for (i = 0; i < nbli->nsci; i++)
4514 {
4515 nblc->sci[sci_offset+i] = nbli->sci[i];
4516 nblc->sci[sci_offset+i].cj4_ind_start += cj4_offset;
4517 nblc->sci[sci_offset+i].cj4_ind_end += cj4_offset;
4518 }
4519
4520 for (j4 = 0; j4 < nbli->ncj4; j4++)
4521 {
4522 nblc->cj4[cj4_offset+j4] = nbli->cj4[j4];
4523 nblc->cj4[cj4_offset+j4].imei[0].excl_ind += excl_offset;
4524 nblc->cj4[cj4_offset+j4].imei[1].excl_ind += excl_offset;
4525 }
4526
4527 for (j4 = 0; j4 < nbli->nexcl; j4++)
4528 {
4529 nblc->excl[excl_offset+j4] = nbli->excl[j4];
4530 }
4531 }
4532
4533 for (n = 0; n < nnbl; n++)
4534 {
4535 nblc->nsci += nbl[n]->nsci;
4536 nblc->ncj4 += nbl[n]->ncj4;
4537 nblc->nci_tot += nbl[n]->nci_tot;
4538 nblc->nexcl += nbl[n]->nexcl;
4539 }
4540}
4541
4542static void balance_fep_lists(const nbnxn_search_t nbs,
4543 nbnxn_pairlist_set_t *nbl_lists)
4544{
4545 int nnbl, th;
4546 int nri_tot, nrj_tot, nrj_target;
4547 int th_dest;
4548 t_nblist *nbld;
4549
4550 nnbl = nbl_lists->nnbl;
4551
4552 if (nnbl == 1)
4553 {
4554 /* Nothing to balance */
4555 return;
4556 }
4557
4558 /* Count the total i-lists and pairs */
4559 nri_tot = 0;
4560 nrj_tot = 0;
4561 for (th = 0; th < nnbl; th++)
4562 {
4563 nri_tot += nbl_lists->nbl_fep[th]->nri;
4564 nrj_tot += nbl_lists->nbl_fep[th]->nrj;
4565 }
4566
4567 nrj_target = (nrj_tot + nnbl - 1)/nnbl;
4568
4569 assert(gmx_omp_nthreads_get(emntNonbonded) == nnbl)((void) (0));
4570
4571#pragma omp parallel for schedule(static) num_threads(nnbl)
4572 for (th = 0; th < nnbl; th++)
4573 {
4574 t_nblist *nbl;
4575
4576 nbl = nbs->work[th].nbl_fep;
4577
4578 /* Note that here we allocate for the total size, instead of
4579 * a per-thread esimate (which is hard to obtain).
4580 */
4581 if (nri_tot > nbl->maxnri)
4582 {
4583 nbl->maxnri = over_alloc_large(nri_tot)(int)(1.19*(nri_tot) + 1000);
4584 reallocate_nblist(nbl);
4585 }
4586 if (nri_tot > nbl->maxnri || nrj_tot > nbl->maxnrj)
4587 {
4588 nbl->maxnrj = over_alloc_small(nrj_tot)(int)(1.19*(nrj_tot) + 8000);
4589 srenew(nbl->jjnr, nbl->maxnrj)(nbl->jjnr) = save_realloc("nbl->jjnr", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 4589, (nbl->jjnr), (nbl->maxnrj), sizeof(*(nbl->jjnr
)));
4590 srenew(nbl->excl_fep, nbl->maxnrj)(nbl->excl_fep) = save_realloc("nbl->excl_fep", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 4590, (nbl->excl_fep), (nbl->maxnrj), sizeof(*(nbl->
excl_fep)));
4591 }
4592
4593 clear_pairlist_fep(nbl);
4594 }
4595
4596 /* Loop over the source lists and assign and copy i-entries */
4597 th_dest = 0;
4598 nbld = nbs->work[th_dest].nbl_fep;
4599 for (th = 0; th < nnbl; th++)
4600 {
4601 t_nblist *nbls;
4602 int i, j;
4603
4604 nbls = nbl_lists->nbl_fep[th];
4605
4606 for (i = 0; i < nbls->nri; i++)
4607 {
4608 int nrj;
4609
4610 /* The number of pairs in this i-entry */
4611 nrj = nbls->jindex[i+1] - nbls->jindex[i];
4612
4613 /* Decide if list th_dest is too large and we should procede
4614 * to the next destination list.
4615 */
4616 if (th_dest+1 < nnbl && nbld->nrj > 0 &&
4617 nbld->nrj + nrj - nrj_target > nrj_target - nbld->nrj)
4618 {
4619 th_dest++;
4620 nbld = nbs->work[th_dest].nbl_fep;
4621 }
4622
4623 nbld->iinr[nbld->nri] = nbls->iinr[i];
4624 nbld->gid[nbld->nri] = nbls->gid[i];
4625 nbld->shift[nbld->nri] = nbls->shift[i];
4626
4627 for (j = nbls->jindex[i]; j < nbls->jindex[i+1]; j++)
4628 {
4629 nbld->jjnr[nbld->nrj] = nbls->jjnr[j];
4630 nbld->excl_fep[nbld->nrj] = nbls->excl_fep[j];
4631 nbld->nrj++;
4632 }
4633 nbld->nri++;
4634 nbld->jindex[nbld->nri] = nbld->nrj;
4635 }
4636 }
4637
4638 /* Swap the list pointers */
4639 for (th = 0; th < nnbl; th++)
4640 {
4641 t_nblist *nbl_tmp;
4642
4643 nbl_tmp = nbl_lists->nbl_fep[th];
4644 nbl_lists->nbl_fep[th] = nbs->work[th].nbl_fep;
4645 nbs->work[th].nbl_fep = nbl_tmp;
4646
4647 if (debug)
4648 {
4649 fprintf(debug, "nbl_fep[%d] nri %4d nrj %4d\n",
4650 th,
4651 nbl_lists->nbl_fep[th]->nri,
4652 nbl_lists->nbl_fep[th]->nrj);
4653 }
4654 }
4655}
4656
4657/* Returns the next ci to be processes by our thread */
4658static gmx_bool next_ci(const nbnxn_grid_t *grid,
4659 int conv,
4660 int nth, int ci_block,
4661 int *ci_x, int *ci_y,
4662 int *ci_b, int *ci)
4663{
4664 (*ci_b)++;
4665 (*ci)++;
4666
4667 if (*ci_b == ci_block)
4668 {
4669 /* Jump to the next block assigned to this task */
4670 *ci += (nth - 1)*ci_block;
4671 *ci_b = 0;
4672 }
4673
4674 if (*ci >= grid->nc*conv)
4675 {
4676 return FALSE0;
4677 }
4678
4679 while (*ci >= grid->cxy_ind[*ci_x*grid->ncy + *ci_y + 1]*conv)
4680 {
4681 *ci_y += 1;
4682 if (*ci_y == grid->ncy)
4683 {
4684 *ci_x += 1;
4685 *ci_y = 0;
4686 }
4687 }
4688
4689 return TRUE1;
4690}
4691
4692/* Returns the distance^2 for which we put cell pairs in the list
4693 * without checking atom pair distances. This is usually < rlist^2.
4694 */
4695static float boundingbox_only_distance2(const nbnxn_grid_t *gridi,
4696 const nbnxn_grid_t *gridj,
4697 real rlist,
4698 gmx_bool simple)
4699{
4700 /* If the distance between two sub-cell bounding boxes is less
4701 * than this distance, do not check the distance between
4702 * all particle pairs in the sub-cell, since then it is likely
4703 * that the box pair has atom pairs within the cut-off.
4704 * We use the nblist cut-off minus 0.5 times the average x/y diagonal
4705 * spacing of the sub-cells. Around 40% of the checked pairs are pruned.
4706 * Using more than 0.5 gains at most 0.5%.
4707 * If forces are calculated more than twice, the performance gain
4708 * in the force calculation outweighs the cost of checking.
4709 * Note that with subcell lists, the atom-pair distance check
4710 * is only performed when only 1 out of 8 sub-cells in within range,
4711 * this is because the GPU is much faster than the cpu.
4712 */
4713 real bbx, bby;
4714 real rbb2;
4715
4716 bbx = 0.5*(gridi->sx + gridj->sx);
4717 bby = 0.5*(gridi->sy + gridj->sy);
4718 if (!simple)
4719 {
4720 bbx /= GPU_NSUBCELL_X2;
4721 bby /= GPU_NSUBCELL_Y2;
4722 }
4723
4724 rbb2 = sqr(max(0, rlist - 0.5*sqrt(bbx*bbx + bby*bby))(((0) > (rlist - 0.5*sqrt(bbx*bbx + bby*bby))) ? (0) : (rlist
- 0.5*sqrt(bbx*bbx + bby*bby)) ));
4725
4726#ifndef GMX_DOUBLE
4727 return rbb2;
4728#else
4729 return (float)((1+GMX_FLOAT_EPS5.96046448E-08)*rbb2);
4730#endif
4731}
4732
4733static int get_ci_block_size(const nbnxn_grid_t *gridi,
4734 gmx_bool bDomDec, int nth)
4735{
4736 const int ci_block_enum = 5;
4737 const int ci_block_denom = 11;
4738 const int ci_block_min_atoms = 16;
4739 int ci_block;
4740
4741 /* Here we decide how to distribute the blocks over the threads.
4742 * We use prime numbers to try to avoid that the grid size becomes
4743 * a multiple of the number of threads, which would lead to some
4744 * threads getting "inner" pairs and others getting boundary pairs,
4745 * which in turns will lead to load imbalance between threads.
4746 * Set the block size as 5/11/ntask times the average number of cells
4747 * in a y,z slab. This should ensure a quite uniform distribution
4748 * of the grid parts of the different thread along all three grid
4749 * zone boundaries with 3D domain decomposition. At the same time
4750 * the blocks will not become too small.
4751 */
4752 ci_block = (gridi->nc*ci_block_enum)/(ci_block_denom*gridi->ncx*nth);
4753
4754 /* Ensure the blocks are not too small: avoids cache invalidation */
4755 if (ci_block*gridi->na_sc < ci_block_min_atoms)
4756 {
4757 ci_block = (ci_block_min_atoms + gridi->na_sc - 1)/gridi->na_sc;
4758 }
4759
4760 /* Without domain decomposition
4761 * or with less than 3 blocks per task, divide in nth blocks.
4762 */
4763 if (!bDomDec || ci_block*3*nth > gridi->nc)
4764 {
4765 ci_block = (gridi->nc + nth - 1)/nth;
4766 }
4767
4768 return ci_block;
4769}
4770
4771/* Generates the part of pair-list nbl assigned to our thread */
4772static void nbnxn_make_pairlist_part(const nbnxn_search_t nbs,
4773 const nbnxn_grid_t *gridi,
4774 const nbnxn_grid_t *gridj,
4775 nbnxn_search_work_t *work,
4776 const nbnxn_atomdata_t *nbat,
4777 const t_blocka *excl,
4778 real rlist,
4779 int nb_kernel_type,
4780 int ci_block,
4781 gmx_bool bFBufferFlag,
4782 int nsubpair_max,
4783 gmx_bool progBal,
4784 int min_ci_balanced,
4785 int th, int nth,
4786 nbnxn_pairlist_t *nbl,
4787 t_nblist *nbl_fep)
4788{
4789 int na_cj_2log;
4790 matrix box;
4791 real rl2, rl_fep2 = 0;
4792 float rbb2;
4793 int d;
4794 int ci_b, ci, ci_x, ci_y, ci_xy, cj;
4795 ivec shp;
4796 int tx, ty, tz;
4797 int shift;
4798 gmx_bool bMakeList;
4799 real shx, shy, shz;
4800 int conv_i, cell0_i;
4801 const nbnxn_bb_t *bb_i = NULL((void*)0);
4802#ifdef NBNXN_BBXXXX
4803 const float *pbb_i = NULL((void*)0);
1
'pbb_i' initialized to a null pointer value
→
4804#endif
4805 const float *bbcz_i, *bbcz_j;
4806 const int *flags_i;
4807 real bx0, bx1, by0, by1, bz0, bz1;
4808 real bz1_frac;
4809 real d2cx, d2z, d2z_cx, d2z_cy, d2zx, d2zxy, d2xy;
4810 int cxf, cxl, cyf, cyf_x, cyl;
4811 int cx, cy;
4812 int c0, c1, cs, cf, cl;
4813 int ndistc;
4814 int ncpcheck;
4815 int gridi_flag_shift = 0, gridj_flag_shift = 0;
4816 unsigned int *gridj_flag = NULL((void*)0);
4817 int ncj_old_i, ncj_old_j;
4818
4819 nbs_cycle_start(&work->cc[enbsCCsearch]);
4820
4821 if (gridj->bSimple != nbl->bSimple)
2
←
Taking false branch
→
4822 {
4823 gmx_incons("Grid incompatible with pair-list")_gmx_error("incons", "Grid incompatible with pair-list", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 4823);
4824 }
4825
4826 sync_work(nbl);
4827 nbl->na_sc = gridj->na_sc;
4828 nbl->na_ci = gridj->na_c;
4829 nbl->na_cj = nbnxn_kernel_to_cj_size(nb_kernel_type);
4830 na_cj_2log = get_2log(nbl->na_cj);
4831
4832 nbl->rlist = rlist;
4833
4834 if (bFBufferFlag)
3
←
Assuming 'bFBufferFlag' is 0
→
4
←
Taking false branch
→
4835 {
4836 /* Determine conversion of clusters to flag blocks */
4837 gridi_flag_shift = 0;
4838 while ((nbl->na_ci<<gridi_flag_shift) < NBNXN_BUFFERFLAG_SIZE16)
4839 {
4840 gridi_flag_shift++;
4841 }
4842 gridj_flag_shift = 0;
4843 while ((nbl->na_cj<<gridj_flag_shift) < NBNXN_BUFFERFLAG_SIZE16)
4844 {
4845 gridj_flag_shift++;
4846 }
4847
4848 gridj_flag = work->buffer_flags.flag;
4849 }
4850
4851 copy_mat(nbs->box, box);
4852
4853 rl2 = nbl->rlist*nbl->rlist;
4854
4855 if (nbs->bFEP && !nbl->bSimple)
4856 {
4857 /* Determine an atom-pair list cut-off distance for FEP atom pairs.
4858 * We should not simply use rlist, since then we would not have
4859 * the small, effective buffering of the NxN lists.
4860 * The buffer is on overestimate, but the resulting cost for pairs
4861 * beyond rlist is neglible compared to the FEP pairs within rlist.
4862 */
4863 rl_fep2 = nbl->rlist + effective_buffer_1x1_vs_MxN(gridi, gridj);
4864
4865 if (debug)
4866 {
4867 fprintf(debug, "nbl_fep atom-pair rlist %f\n", rl_fep2);
4868 }
4869 rl_fep2 = rl_fep2*rl_fep2;
4870 }
4871
4872 rbb2 = boundingbox_only_distance2(gridi, gridj, nbl->rlist, nbl->bSimple);
4873
4874 if (debug)
5
←
Assuming 'debug' is null
→
6
←
Taking false branch
→
4875 {
4876 fprintf(debug, "nbl bounding box only distance %f\n", sqrt(rbb2));
4877 }
4878
4879 /* Set the shift range */
4880 for (d = 0; d < DIM3; d++)
7
←
Loop condition is true. Entering loop body
→
10
←
Loop condition is true. Entering loop body
→
12
←
Loop condition is true. Entering loop body
→
14
←
Loop condition is false. Execution continues on line 4903
→
4881 {
4882 /* Check if we need periodicity shifts.
4883 * Without PBC or with domain decomposition we don't need them.
4884 */
4885 if (d >= ePBC2npbcdim(nbs->ePBC) || nbs->dd_dim[d])
8
←
Taking false branch
→
11
←
Taking false branch
→
13
←
Taking false branch
→
4886 {
4887 shp[d] = 0;
4888 }
4889 else
4890 {
4891 if (d == XX0 &&
9
←
Taking false branch
→
4892 box[XX0][XX0] - fabs(box[YY1][XX0]) - fabs(box[ZZ2][XX0]) < sqrt(rl2))
4893 {
4894 shp[d] = 2;
4895 }
4896 else
4897 {
4898 shp[d] = 1;
4899 }
4900 }
4901 }
4902
4903 if (nbl->bSimple && !gridi->bSimple)
15
←
Taking false branch
→
4904 {
4905 conv_i = gridi->na_sc/gridj->na_sc;
4906 bb_i = gridi->bb_simple;
4907 bbcz_i = gridi->bbcz_simple;
4908 flags_i = gridi->flags_simple;
4909 }
4910 else
4911 {
4912 conv_i = 1;
4913#ifdef NBNXN_BBXXXX
4914 if (gridi->bSimple)
16
←
Taking true branch
→
4915 {
4916 bb_i = gridi->bb;
4917 }
4918 else
4919 {
4920 pbb_i = gridi->pbb;
4921 }
4922#else
4923 /* We use the normal bounding box format for both grid types */
4924 bb_i = gridi->bb;
4925#endif
4926 bbcz_i = gridi->bbcz;
4927 flags_i = gridi->flags;
4928 }
4929 cell0_i = gridi->cell0*conv_i;
4930
4931 bbcz_j = gridj->bbcz;
4932
4933 if (conv_i != 1)
17
←
Taking false branch
→
4934 {
4935 /* Blocks of the conversion factor - 1 give a large repeat count
4936 * combined with a small block size. This should result in good
4937 * load balancing for both small and large domains.
4938 */
4939 ci_block = conv_i - 1;
4940 }
4941 if (debug)
18
←
Assuming 'debug' is null
→
19
←
Taking false branch
→
4942 {
4943 fprintf(debug, "nbl nc_i %d col.av. %.1f ci_block %d\n",
4944 gridi->nc, gridi->nc/(double)(gridi->ncx*gridi->ncy), ci_block);
4945 }
4946
4947 ndistc = 0;
4948 ncpcheck = 0;
4949
4950 /* Initially ci_b and ci to 1 before where we want them to start,
4951 * as they will both be incremented in next_ci.
4952 */
4953 ci_b = -1;
4954 ci = th*ci_block - 1;
4955 ci_x = 0;
4956 ci_y = 0;
4957 while (next_ci(gridi, conv_i, nth, ci_block, &ci_x, &ci_y, &ci_b, &ci))
20
←
Loop condition is true. Entering loop body
→
4958 {
4959 if (nbl->bSimple && flags_i[ci] == 0)
21
←
Taking false branch
→
4960 {
4961 continue;
4962 }
4963
4964 ncj_old_i = nbl->ncj;
4965
4966 d2cx = 0;
4967 if (gridj != gridi && shp[XX0] == 0)
22
←
Assuming 'gridj' is equal to 'gridi'
→
4968 {
4969 if (nbl->bSimple)
4970 {
4971 bx1 = bb_i[ci].upper[BB_X0];
4972 }
4973 else
4974 {
4975 bx1 = gridi->c0[XX0] + (ci_x+1)*gridi->sx;
4976 }
4977 if (bx1 < gridj->c0[XX0])
4978 {
4979 d2cx = sqr(gridj->c0[XX0] - bx1);
4980
4981 if (d2cx >= rl2)
4982 {
4983 continue;
4984 }
4985 }
4986 }
4987
4988 ci_xy = ci_x*gridi->ncy + ci_y;
4989
4990 /* Loop over shift vectors in three dimensions */
4991 for (tz = -shp[ZZ2]; tz <= shp[ZZ2]; tz++)
23
←
Loop condition is true. Entering loop body
→
4992 {
4993 shz = tz*box[ZZ2][ZZ2];
4994
4995 bz0 = bbcz_i[ci*NNBSBB_D2 ] + shz;
4996 bz1 = bbcz_i[ci*NNBSBB_D2+1] + shz;
4997
4998 if (tz == 0)
24
←
Taking false branch
→
4999 {
5000 d2z = 0;
5001 }
5002 else if (tz < 0)
25
←
Taking true branch
→
5003 {
5004 d2z = sqr(bz1);
5005 }
5006 else
5007 {
5008 d2z = sqr(bz0 - box[ZZ2][ZZ2]);
5009 }
5010
5011 d2z_cx = d2z + d2cx;
5012
5013 if (d2z_cx >= rl2)
26
←
Taking false branch
→
5014 {
5015 continue;
5016 }
5017
5018 bz1_frac =
5019 bz1/((real)(gridi->cxy_ind[ci_xy+1] - gridi->cxy_ind[ci_xy]));
5020 if (bz1_frac < 0)
27
←
Taking false branch
→
5021 {
5022 bz1_frac = 0;
5023 }
5024 /* The check with bz1_frac close to or larger than 1 comes later */
5025
5026 for (ty = -shp[YY1]; ty <= shp[YY1]; ty++)
28
←
Loop condition is true. Entering loop body
→
5027 {
5028 shy = ty*box[YY1][YY1] + tz*box[ZZ2][YY1];
5029
5030 if (nbl->bSimple)
29
←
Taking true branch
→
5031 {
5032 by0 = bb_i[ci].lower[BB_Y1] + shy;
5033 by1 = bb_i[ci].upper[BB_Y1] + shy;
5034 }
5035 else
5036 {
5037 by0 = gridi->c0[YY1] + (ci_y )*gridi->sy + shy;
5038 by1 = gridi->c0[YY1] + (ci_y+1)*gridi->sy + shy;
5039 }
5040
5041 get_cell_range(by0, by1,
5042 gridj->ncy, gridj->c0[YY1], gridj->sy, gridj->inv_sy,
5043 d2z_cx, rl2,
5044 &cyf, &cyl);
5045
5046 if (cyf > cyl)
30
←
Assuming 'cyf' is <= 'cyl'
→
31
←
Taking false branch
→
5047 {
5048 continue;
5049 }
5050
5051 d2z_cy = d2z;
5052 if (by1 < gridj->c0[YY1])
32
←
Taking false branch
→
5053 {
5054 d2z_cy += sqr(gridj->c0[YY1] - by1);
5055 }
5056 else if (by0 > gridj->c1[YY1])
33
←
Taking false branch
→
5057 {
5058 d2z_cy += sqr(by0 - gridj->c1[YY1]);
5059 }
5060
5061 for (tx = -shp[XX0]; tx <= shp[XX0]; tx++)
34
←
Loop condition is true. Entering loop body
→
54
←
Loop condition is true. Entering loop body
→
70
←
Loop condition is true. Entering loop body
→
5062 {
5063 shift = XYZ2IS(tx, ty, tz)((2*2 +1)*((2*1 +1)*((tz)+1)+(ty)+1)+(tx)+2);
5064
5065#ifdef NBNXN_SHIFT_BACKWARD
5066 if (gridi == gridj && shift > CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2))
35
←
Assuming 'gridi' is not equal to 'gridj'
→
5067 {
5068 continue;
5069 }
5070#endif
5071
5072 shx = tx*box[XX0][XX0] + ty*box[YY1][XX0] + tz*box[ZZ2][XX0];
5073
5074 if (nbl->bSimple)
36
←
Taking true branch
→
55
←
Taking true branch
→
71
←
Taking true branch
→
5075 {
5076 bx0 = bb_i[ci].lower[BB_X0] + shx;
5077 bx1 = bb_i[ci].upper[BB_X0] + shx;
5078 }
5079 else
5080 {
5081 bx0 = gridi->c0[XX0] + (ci_x )*gridi->sx + shx;
5082 bx1 = gridi->c0[XX0] + (ci_x+1)*gridi->sx + shx;
5083 }
5084
5085 get_cell_range(bx0, bx1,
5086 gridj->ncx, gridj->c0[XX0], gridj->sx, gridj->inv_sx,
5087 d2z_cy, rl2,
5088 &cxf, &cxl);
5089
5090 if (cxf > cxl)
37
←
Assuming 'cxf' is <= 'cxl'
→
38
←
Taking false branch
→
56
←
Taking false branch
→
72
←
Taking false branch
→
5091 {
5092 continue;
5093 }
5094
5095 if (nbl->bSimple)
39
←
Taking true branch
→
57
←
Taking true branch
→
73
←
Taking true branch
→
5096 {
5097 new_ci_entry(nbl, cell0_i+ci, shift, flags_i[ci]);
5098 }
5099 else
5100 {
5101 new_sci_entry(nbl, cell0_i+ci, shift);
5102 }
5103
5104#ifndef NBNXN_SHIFT_BACKWARD
5105 if (cxf < ci_x)
5106#else
5107 if (shift == CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2) && gridi == gridj &&
5108 cxf < ci_x)
5109#endif
5110 {
5111 /* Leave the pairs with i > j.
5112 * x is the major index, so skip half of it.
5113 */
5114 cxf = ci_x;
5115 }
5116
5117 if (nbl->bSimple)
40
←
Taking true branch
→
58
←
Taking true branch
→
74
←
Taking false branch
→
5118 {
5119 set_icell_bb_simple(bb_i, ci, shx, shy, shz,
5120 nbl->work->bb_ci);
5121 }
5122 else
5123 {
5124#ifdef NBNXN_BBXXXX
5125 set_icell_bbxxxx_supersub(pbb_i, ci, shx, shy, shz,
75
←
Passing null pointer value via 1st parameter 'bb'
→
76
←
Calling 'set_icell_bbxxxx_supersub'
→
5126 nbl->work->pbb_ci);
5127#else
5128 set_icell_bb_supersub(bb_i, ci, shx, shy, shz,
5129 nbl->work->bb_ci);
5130#endif
5131 }
5132
5133 nbs->icell_set_x(cell0_i+ci, shx, shy, shz,
5134 gridi->na_c, nbat->xstride, nbat->x,
5135 nbl->work);
5136
5137 for (cx = cxf; cx <= cxl; cx++)
41
←
Loop condition is true. Entering loop body
→
49
←
Assuming 'cx' is > 'cxl'
→
50
←
Loop condition is false. Execution continues on line 5331
→
59
←
Loop condition is true. Entering loop body
→
66
←
Loop condition is false. Execution continues on line 5331
→
5138 {
5139 d2zx = d2z;
5140 if (gridj->c0[XX0] + cx*gridj->sx > bx1)
42
←
Taking false branch
→
60
←
Taking false branch
→
5141 {
5142 d2zx += sqr(gridj->c0[XX0] + cx*gridj->sx - bx1);
5143 }
5144 else if (gridj->c0[XX0] + (cx+1)*gridj->sx < bx0)
43
←
Taking false branch
→
61
←
Taking false branch
→
5145 {
5146 d2zx += sqr(gridj->c0[XX0] + (cx+1)*gridj->sx - bx0);
5147 }
5148
5149#ifndef NBNXN_SHIFT_BACKWARD
5150 if (gridi == gridj &&
5151 cx == 0 && cyf < ci_y)
5152#else
5153 if (gridi == gridj &&
5154 cx == 0 && shift == CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2) && cyf < ci_y)
5155#endif
5156 {
5157 /* Leave the pairs with i > j.
5158 * Skip half of y when i and j have the same x.
5159 */
5160 cyf_x = ci_y;
5161 }
5162 else
5163 {
5164 cyf_x = cyf;
5165 }
5166
5167 for (cy = cyf_x; cy <= cyl; cy++)
44
←
Loop condition is true. Entering loop body
→
47
←
Assuming 'cy' is > 'cyl'
→
48
←
Loop condition is false. Execution continues on line 5137
→
62
←
Loop condition is true. Entering loop body
→
65
←
Loop condition is false. Execution continues on line 5137
→
5168 {
5169 c0 = gridj->cxy_ind[cx*gridj->ncy+cy];
5170 c1 = gridj->cxy_ind[cx*gridj->ncy+cy+1];
5171#ifdef NBNXN_SHIFT_BACKWARD
5172 if (gridi == gridj &&
5173 shift == CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2) && c0 < ci)
5174 {
5175 c0 = ci;
5176 }
5177#endif
5178
5179 d2zxy = d2zx;
5180 if (gridj->c0[YY1] + cy*gridj->sy > by1)
45
←
Taking false branch
→
63
←
Taking false branch
→
5181 {
5182 d2zxy += sqr(gridj->c0[YY1] + cy*gridj->sy - by1);
5183 }
5184 else if (gridj->c0[YY1] + (cy+1)*gridj->sy < by0)
46
←
Taking false branch
→
64
←
Taking false branch
→
5185 {
5186 d2zxy += sqr(gridj->c0[YY1] + (cy+1)*gridj->sy - by0);
5187 }
5188 if (c1 > c0 && d2zxy < rl2)
5189 {
5190 cs = c0 + (int)(bz1_frac*(c1 - c0));
5191 if (cs >= c1)
5192 {
5193 cs = c1 - 1;
5194 }
5195
5196 d2xy = d2zxy - d2z;
5197
5198 /* Find the lowest cell that can possibly
5199 * be within range.
5200 */
5201 cf = cs;
5202 while (cf > c0 &&
5203 (bbcz_j[cf*NNBSBB_D2+1] >= bz0 ||
5204 d2xy + sqr(bbcz_j[cf*NNBSBB_D2+1] - bz0) < rl2))
5205 {
5206 cf--;
5207 }
5208
5209 /* Find the highest cell that can possibly
5210 * be within range.
5211 */
5212 cl = cs;
5213 while (cl < c1-1 &&
5214 (bbcz_j[cl*NNBSBB_D2] <= bz1 ||
5215 d2xy + sqr(bbcz_j[cl*NNBSBB_D2] - bz1) < rl2))
5216 {
5217 cl++;
5218 }
5219
5220#ifdef NBNXN_REFCODE
5221 {
5222 /* Simple reference code, for debugging,
5223 * overrides the more complex code above.
5224 */
5225 int k;
5226 cf = c1;
5227 cl = -1;
5228 for (k = c0; k < c1; k++)
5229 {
5230 if (box_dist2(bx0, bx1, by0, by1, bz0, bz1, bb+k) < rl2 &&
5231 k < cf)
5232 {
5233 cf = k;
5234 }
5235 if (box_dist2(bx0, bx1, by0, by1, bz0, bz1, bb+k) < rl2 &&
5236 k > cl)
5237 {
5238 cl = k;
5239 }
5240 }
5241 }
5242#endif
5243
5244 if (gridi == gridj)
5245 {
5246 /* We want each atom/cell pair only once,
5247 * only use cj >= ci.
5248 */
5249#ifndef NBNXN_SHIFT_BACKWARD
5250 cf = max(cf, ci)(((cf) > (ci)) ? (cf) : (ci) );
5251#else
5252 if (shift == CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2))
5253 {
5254 cf = max(cf, ci)(((cf) > (ci)) ? (cf) : (ci) );
5255 }
5256#endif
5257 }
5258
5259 if (cf <= cl)
5260 {
5261 /* For f buffer flags with simple lists */
5262 ncj_old_j = nbl->ncj;
5263
5264 switch (nb_kernel_type)
5265 {
5266 case nbnxnk4x4_PlainC:
5267 check_subcell_list_space_simple(nbl, cl-cf+1);
5268
5269 make_cluster_list_simple(gridj,
5270 nbl, ci, cf, cl,
5271 (gridi == gridj && shift == CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2)),
5272 nbat->x,
5273 rl2, rbb2,
5274 &ndistc);
5275 break;
5276#ifdef GMX_NBNXN_SIMD_4XN
5277 case nbnxnk4xN_SIMD_4xN:
5278 check_subcell_list_space_simple(nbl, ci_to_cj(na_cj_2log, cl-cf)+2);
5279 make_cluster_list_simd_4xn(gridj,
5280 nbl, ci, cf, cl,
5281 (gridi == gridj && shift == CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2)),
5282 nbat->x,
5283 rl2, rbb2,
5284 &ndistc);
5285 break;
5286#endif
5287#ifdef GMX_NBNXN_SIMD_2XNN
5288 case nbnxnk4xN_SIMD_2xNN:
5289 check_subcell_list_space_simple(nbl, ci_to_cj(na_cj_2log, cl-cf)+2);
5290 make_cluster_list_simd_2xnn(gridj,
5291 nbl, ci, cf, cl,
5292 (gridi == gridj && shift == CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2)),
5293 nbat->x,
5294 rl2, rbb2,
5295 &ndistc);
5296 break;
5297#endif
5298 case nbnxnk8x8x8_PlainC:
5299 case nbnxnk8x8x8_CUDA:
5300 check_subcell_list_space_supersub(nbl, cl-cf+1);
5301 for (cj = cf; cj <= cl; cj++)
5302 {
5303 make_cluster_list_supersub(gridi, gridj,
5304 nbl, ci, cj,
5305 (gridi == gridj && shift == CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2) && ci == cj),
5306 nbat->xstride, nbat->x,
5307 rl2, rbb2,
5308 &ndistc);
5309 }
5310 break;
5311 }
5312 ncpcheck += cl - cf + 1;
5313
5314 if (bFBufferFlag && nbl->ncj > ncj_old_j)
5315 {
5316 int cbf, cbl, cb;
5317
5318 cbf = nbl->cj[ncj_old_j].cj >> gridj_flag_shift;
5319 cbl = nbl->cj[nbl->ncj-1].cj >> gridj_flag_shift;
5320 for (cb = cbf; cb <= cbl; cb++)
5321 {
5322 gridj_flag[cb] = 1U<<th;
5323 }
5324 }
5325 }
5326 }
5327 }
5328 }
5329
5330 /* Set the exclusions for this ci list */
5331 if (nbl->bSimple)
51
←
Taking true branch
→
67
←
Taking true branch
→
5332 {
5333 set_ci_top_excls(nbs,
5334 nbl,
5335 shift == CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2) && gridi == gridj,
5336 gridj->na_c_2log,
5337 na_cj_2log,
5338 &(nbl->ci[nbl->nci]),
5339 excl);
5340
5341 if (nbs->bFEP)
52
←
Taking false branch
→
68
←
Taking false branch
→
5342 {
5343 make_fep_list(nbs, nbat, nbl,
5344 shift == CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2) && gridi == gridj,
5345 &(nbl->ci[nbl->nci]),
5346 gridi, gridj, nbl_fep);
5347 }
5348 }
5349 else
5350 {
5351 set_sci_top_excls(nbs,
5352 nbl,
5353 shift == CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2) && gridi == gridj,
5354 gridj->na_c_2log,
5355 &(nbl->sci[nbl->nsci]),
5356 excl);
5357
5358 if (nbs->bFEP)
5359 {
5360 make_fep_list_supersub(nbs, nbat, nbl,
5361 shift == CENTRAL(((2*1 +1)*(2*1 +1)*(2*2 +1))/2) && gridi == gridj,
5362 &(nbl->sci[nbl->nsci]),
5363 shx, shy, shz,
5364 rl_fep2,
5365 gridi, gridj, nbl_fep);
5366 }
5367 }
5368
5369 /* Close this ci list */
5370 if (nbl->bSimple)
53
←
Taking true branch
→
69
←
Taking true branch
→
5371 {
5372 close_ci_entry_simple(nbl);
5373 }
5374 else
5375 {
5376 close_ci_entry_supersub(nbl,
5377 nsubpair_max,
5378 progBal, min_ci_balanced,
5379 th, nth);
5380 }
5381 }
5382 }
5383 }
5384
5385 if (bFBufferFlag && nbl->ncj > ncj_old_i)
5386 {
5387 work->buffer_flags.flag[(gridi->cell0+ci)>>gridi_flag_shift] = 1U<<th;
5388 }
5389 }
5390
5391 work->ndistc = ndistc;
5392
5393 nbs_cycle_stop(&work->cc[enbsCCsearch]);
5394
5395 if (debug)
5396 {
5397 fprintf(debug, "number of distance checks %d\n", ndistc);
5398 fprintf(debug, "ncpcheck %s %d\n", gridi == gridj ? "local" : "non-local",
5399 ncpcheck);
5400
5401 if (nbl->bSimple)
5402 {
5403 print_nblist_statistics_simple(debug, nbl, nbs, rlist);
5404 }
5405 else
5406 {
5407 print_nblist_statistics_supersub(debug, nbl, nbs, rlist);
5408 }
5409
5410 if (nbs->bFEP)
5411 {
5412 fprintf(debug, "nbl FEP list pairs: %d\n", nbl_fep->nrj);
5413 }
5414 }
5415}
5416
5417static void reduce_buffer_flags(const nbnxn_search_t nbs,
5418 int nsrc,
5419 const nbnxn_buffer_flags_t *dest)
5420{
5421 int s, b;
5422 const unsigned int *flag;
5423
5424 for (s = 0; s < nsrc; s++)
5425 {
5426 flag = nbs->work[s].buffer_flags.flag;
5427
5428 for (b = 0; b < dest->nflag; b++)
5429 {
5430 dest->flag[b] |= flag[b];
5431 }
5432 }
5433}
5434
5435static void print_reduction_cost(const nbnxn_buffer_flags_t *flags, int nout)
5436{
5437 int nelem, nkeep, ncopy, nred, b, c, out;
5438
5439 nelem = 0;
5440 nkeep = 0;
5441 ncopy = 0;
5442 nred = 0;
5443 for (b = 0; b < flags->nflag; b++)
5444 {
5445 if (flags->flag[b] == 1)
5446 {
5447 /* Only flag 0 is set, no copy of reduction required */
5448 nelem++;
5449 nkeep++;
5450 }
5451 else if (flags->flag[b] > 0)
5452 {
5453 c = 0;
5454 for (out = 0; out < nout; out++)
5455 {
5456 if (flags->flag[b] & (1U<<out))
5457 {
5458 c++;
5459 }
5460 }
5461 nelem += c;
5462 if (c == 1)
5463 {
5464 ncopy++;
5465 }
5466 else
5467 {
5468 nred += c;
5469 }
5470 }
5471 }
5472
5473 fprintf(debug, "nbnxn reduction: #flag %d #list %d elem %4.2f, keep %4.2f copy %4.2f red %4.2f\n",
5474 flags->nflag, nout,
5475 nelem/(double)(flags->nflag),
5476 nkeep/(double)(flags->nflag),
5477 ncopy/(double)(flags->nflag),
5478 nred/(double)(flags->nflag));
5479}
5480
5481/* Perform a count (linear) sort to sort the smaller lists to the end.
5482 * This avoids load imbalance on the GPU, as large lists will be
5483 * scheduled and executed first and the smaller lists later.
5484 * Load balancing between multi-processors only happens at the end
5485 * and there smaller lists lead to more effective load balancing.
5486 * The sorting is done on the cj4 count, not on the actual pair counts.
5487 * Not only does this make the sort faster, but it also results in
5488 * better load balancing than using a list sorted on exact load.
5489 * This function swaps the pointer in the pair list to avoid a copy operation.
5490 */
5491static void sort_sci(nbnxn_pairlist_t *nbl)
5492{
5493 nbnxn_list_work_t *work;
5494 int m, i, s, s0, s1;
5495 nbnxn_sci_t *sci_sort;
5496
5497 if (nbl->ncj4 <= nbl->nsci)
5498 {
5499 /* nsci = 0 or all sci have size 1, sorting won't change the order */
5500 return;
5501 }
5502
5503 work = nbl->work;
5504
5505 /* We will distinguish differences up to double the average */
5506 m = (2*nbl->ncj4)/nbl->nsci;
5507
5508 if (m + 1 > work->sort_nalloc)
5509 {
5510 work->sort_nalloc = over_alloc_large(m + 1)(int)(1.19*(m + 1) + 1000);
5511 srenew(work->sort, work->sort_nalloc)(work->sort) = save_realloc("work->sort", "/home/alexxy/Develop/gromacs/src/gromacs/mdlib/nbnxn_search.c"
, 5511, (work->sort), (work->sort_nalloc), sizeof(*(work
->sort)));
5512 }
5513
5514 if (work->sci_sort_nalloc != nbl->sci_nalloc)
5515 {
5516 work->sci_sort_nalloc = nbl->sci_nalloc;
5517 nbnxn_realloc_void((void **)&work->sci_sort,
5518 0,
5519 work->sci_sort_nalloc*sizeof(*work->sci_sort),
5520 nbl->alloc, nbl->free);
5521 }
5522
5523 /* Count the entries of each size */
5524 for (i = 0; i <= m; i++)
5525 {
5526 work->sort[i] = 0;
5527 }
5528 for (s = 0; s < nbl->nsci; s++)
5529 {
5530 i = min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start)(((m) < (nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start
)) ? (m) : (nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start
) );
5531 work->sort[i]++;
5532 }
5533 /* Calculate the offset for each count */
5534 s0 = work->sort[m];
5535 work->sort[m] = 0;
5536 for (i = m - 1; i >= 0; i--)
5537 {
5538 s1 = work->sort[i];
5539 work->sort[i] = work->sort[i + 1] + s0;
5540 s0 = s1;
5541 }
5542
5543 /* Sort entries directly into place */
5544 sci_sort = work->sci_sort;
5545 for (s = 0; s < nbl->nsci; s++)
5546 {
5547 i = min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start)(((m) < (nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start
)) ? (m) : (nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start
) );
5548 sci_sort[work->sort[i]++] = nbl->sci[s];
5549 }
5550
5551 /* Swap the sci pointers so we use the new, sorted list */
5552 work->sci_sort = nbl->sci;
5553 nbl->sci = sci_sort;
5554}
5555
5556/* Make a local or non-local pair-list, depending on iloc */
5557void nbnxn_make_pairlist(const nbnxn_search_t nbs,
5558 nbnxn_atomdata_t *nbat,
5559 const t_blocka *excl,
5560 real rlist,
5561 int min_ci_balanced,
5562 nbnxn_pairlist_set_t *nbl_list,
5563 int iloc,
5564 int nb_kernel_type,
5565 t_nrnb *nrnb)
5566{
5567 nbnxn_grid_t *gridi, *gridj;
5568 gmx_bool bGPUCPU;
5569 int nzi, zi, zj0, zj1, zj;
5570 int nsubpair_max;
5571 int th;
5572 int nnbl;
5573 nbnxn_pairlist_t **nbl;
5574 int ci_block;
5575 gmx_bool CombineNBLists;
5576 gmx_bool progBal;
5577 int np_tot, np_noq, np_hlj, nap;
5578
5579 /* Check if we are running hybrid GPU + CPU nbnxn mode */
5580 bGPUCPU = (!nbs->grid[0].bSimple && nbl_list->bSimple);
5581
5582 nnbl = nbl_list->nnbl;
5583 nbl = nbl_list->nbl;
5584 CombineNBLists = nbl_list->bCombined;
5585
5586 if (debug)
5587 {
5588 fprintf(debug, "ns making %d nblists\n", nnbl);
5589 }
5590
5591 nbat->bUseBufferFlags = (nbat->nout > 1);
5592 /* We should re-init the flags before making the first list */
5593 if (nbat->bUseBufferFlags && (LOCAL_I(iloc)((iloc) == eintLocal) || bGPUCPU))
5594 {
5595 init_buffer_flags(&nbat->buffer_flags, nbat->natoms);
5596 }
5597
5598 if (nbl_list->bSimple)
5599 {
5600 switch (nb_kernel_type)
5601 {
5602#ifdef GMX_NBNXN_SIMD_4XN
5603 case nbnxnk4xN_SIMD_4xN:
5604 nbs->icell_set_x = icell_set_x_simd_4xn;
5605 break;
5606#endif
5607#ifdef GMX_NBNXN_SIMD_2XNN
5608 case nbnxnk4xN_SIMD_2xNN:
5609 nbs->icell_set_x = icell_set_x_simd_2xnn;
5610 break;
5611#endif
5612 default:
5613 nbs->icell_set_x = icell_set_x_simple;
5614 break;
5615 }
5616 }
5617 else
5618 {
5619#ifdef NBNXN_SEARCH_BB_SIMD4
5620 nbs->icell_set_x = icell_set_x_supersub_simd4;
5621#else
5622 nbs->icell_set_x = icell_set_x_supersub;
5623#endif
5624 }
5625
5626 if (LOCAL_I(iloc)((iloc) == eintLocal))
5627 {
5628 /* Only zone (grid) 0 vs 0 */
5629 nzi = 1;
5630 zj0 = 0;
5631 zj1 = 1;
5632 }
5633 else
5634 {
5635 nzi = nbs->zones->nizone;
5636 }
5637
5638 if (!nbl_list->bSimple && min_ci_balanced > 0)
5639 {
5640 nsubpair_max = get_nsubpair_max(nbs, iloc, rlist, min_ci_balanced);
5641 }
5642 else
5643 {
5644 nsubpair_max = 0;
5645 }
5646
5647 /* Clear all pair-lists */
5648 for (th = 0; th < nnbl; th++)
5649 {
5650 clear_pairlist(nbl[th]);
5651
5652 if (nbs->bFEP)
5653 {
5654 clear_pairlist_fep(nbl_list->nbl_fep[th]);
5655 }
5656 }
5657
5658 for (zi = 0; zi < nzi; zi++)
5659 {
5660 gridi = &nbs->grid[zi];
5661
5662 if (NONLOCAL_I(iloc)((iloc) == eintNonlocal))
5663 {
5664 zj0 = nbs->zones->izone[zi].j0;
5665 zj1 = nbs->zones->izone[zi].j1;
5666 if (zi == 0)
5667 {
5668 zj0++;
5669 }
5670 }
5671 for (zj = zj0; zj < zj1; zj++)
5672 {
5673 gridj = &nbs->grid[zj];
5674
5675 if (debug)
5676 {
5677 fprintf(debug, "ns search grid %d vs %d\n", zi, zj);
5678 }
5679
5680 nbs_cycle_start(&nbs->cc[enbsCCsearch]);
5681
5682 if (nbl[0]->bSimple && !gridi->bSimple)
5683 {
5684 /* Hybrid list, determine blocking later */
5685 ci_block = 0;
5686 }
5687 else
5688 {
5689 ci_block = get_ci_block_size(gridi, nbs->DomDec, nnbl);
5690 }
5691
5692 /* With GPU: generate progressively smaller lists for
5693 * load balancing for local only or non-local with 2 zones.
5694 */
5695 progBal = (LOCAL_I(iloc)((iloc) == eintLocal) || nbs->zones->n <= 2);
5696
5697#pragma omp parallel for num_threads(nnbl) schedule(static)
5698 for (th = 0; th < nnbl; th++)
5699 {
5700 /* Re-init the thread-local work flag data before making
5701 * the first list (not an elegant conditional).
5702 */
5703 if (nbat->bUseBufferFlags && ((zi == 0 && zj == 0) ||
5704 (bGPUCPU && zi == 0 && zj == 1)))
5705 {
5706 init_buffer_flags(&nbs->work[th].buffer_flags, nbat->natoms);
5707 }
5708
5709 if (CombineNBLists && th > 0)
5710 {
5711 clear_pairlist(nbl[th]);
5712 }
5713
5714 /* Divide the i super cell equally over the nblists */
5715 nbnxn_make_pairlist_part(nbs, gridi, gridj,
5716 &nbs->work[th], nbat, excl,
5717 rlist,
5718 nb_kernel_type,
5719 ci_block,
5720 nbat->bUseBufferFlags,
5721 nsubpair_max,
5722 progBal, min_ci_balanced,
5723 th, nnbl,
5724 nbl[th],
5725 nbl_list->nbl_fep[th]);
5726 }
5727 nbs_cycle_stop(&nbs->cc[enbsCCsearch]);
5728
5729 np_tot = 0;
5730 np_noq = 0;
5731 np_hlj = 0;
5732 for (th = 0; th < nnbl; th++)
5733 {
5734 inc_nrnb(nrnb, eNR_NBNXN_DIST2, nbs->work[th].ndistc)(nrnb)->n[eNR_NBNXN_DIST2] += nbs->work[th].ndistc;
5735
5736 if (nbl_list->bSimple)
5737 {
5738 np_tot += nbl[th]->ncj;
5739 np_noq += nbl[th]->work->ncj_noq;
5740 np_hlj += nbl[th]->work->ncj_hlj;
5741 }
5742 else
5743 {
5744 /* This count ignores potential subsequent pair pruning */
5745 np_tot += nbl[th]->nci_tot;
5746 }
5747 }
5748 nap = nbl[0]->na_ci*nbl[0]->na_cj;
5749 nbl_list->natpair_ljq = (np_tot - np_noq)*nap - np_hlj*nap/2;
5750 nbl_list->natpair_lj = np_noq*nap;
5751 nbl_list->natpair_q = np_hlj*nap/2;
5752
5753 if (CombineNBLists && nnbl > 1)
5754 {
5755 nbs_cycle_start(&nbs->cc[enbsCCcombine]);
5756
5757 combine_nblists(nnbl-1, nbl+1, nbl[0]);
5758
5759 nbs_cycle_stop(&nbs->cc[enbsCCcombine]);
5760 }
5761 }
5762 }
5763
5764 if (!nbl_list->bSimple)
5765 {
5766 /* Sort the entries on size, large ones first */
5767 if (CombineNBLists || nnbl == 1)
5768 {
5769 sort_sci(nbl[0]);
5770 }
5771 else
5772 {
5773#pragma omp parallel for num_threads(nnbl) schedule(static)
5774 for (th = 0; th < nnbl; th++)
5775 {
5776 sort_sci(nbl[th]);
5777 }
5778 }
5779 }
5780
5781 if (nbat->bUseBufferFlags)
5782 {
5783 reduce_buffer_flags(nbs, nnbl, &nbat->buffer_flags);
5784 }
5785
5786 if (nbs->bFEP)
5787 {
5788 /* Balance the free-energy lists over all the threads */
5789 balance_fep_lists(nbs, nbl_list);
5790 }
5791
5792 /* Special performance logging stuff (env.var. GMX_NBNXN_CYCLE) */
5793 if (LOCAL_I(iloc)((iloc) == eintLocal))
5794 {
5795 nbs->search_count++;
5796 }
5797 if (nbs->print_cycles &&
5798 (!nbs->DomDec || (nbs->DomDec && !LOCAL_I(iloc)((iloc) == eintLocal))) &&
5799 nbs->search_count % 100 == 0)
5800 {
5801 nbs_cycle_print(stderrstderr, nbs);
5802 }
5803
5804 if (debug && (CombineNBLists && nnbl > 1))
5805 {
5806 if (nbl[0]->bSimple)
5807 {
5808 print_nblist_statistics_simple(debug, nbl[0], nbs, rlist);
5809 }
5810 else
5811 {
5812 print_nblist_statistics_supersub(debug, nbl[0], nbs, rlist);
5813 }
5814 }
5815
5816 if (debug)
5817 {
5818 if (gmx_debug_at)
5819 {
5820 if (nbl[0]->bSimple)
5821 {
5822 print_nblist_ci_cj(debug, nbl[0]);
5823 }
5824 else
5825 {
5826 print_nblist_sci_cj(debug, nbl[0]);
5827 }
5828 }
5829
5830 if (nbat->bUseBufferFlags)
5831 {
5832 print_reduction_cost(&nbat->buffer_flags, nnbl);
5833 }
5834 }
5835}