?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
made pair-search with GPU 20% faster
[gromacs.git] / src / mdlib / nbnxn_search.c
blob68617556a7be743eb6eb32ac08f8e17cea645d64
1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
2 *
3 *
4 * This source code is part of
5 *
6 * G R O M A C S
7 *
8 * GROningen MAchine for Chemical Simulations
9 *
10 * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
11 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
12 * Copyright (c) 2001-2012, The GROMACS development team,
13 * check out http://www.gromacs.org for more information.
14
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version 2
18 * of the License, or (at your option) any later version.
19 *
20 * If you want to redistribute modifications, please consider that
21 * scientific software is very special. Version control is crucial -
22 * bugs must be traceable. We will be happy to consider code for
23 * inclusion in the official distribution, but derived work must not
24 * be called official GROMACS. Details are found in the README & COPYING
25 * files - if they are missing, get the official version at www.gromacs.org.
26 *
27 * To help us fund GROMACS development, we humbly ask that you cite
28 * the papers on the package - you can find them in the top README file.
29 *
30 * For more info, check our website at http://www.gromacs.org
31 */
32
33 #ifdef HAVE_CONFIG_H
34 #include <config.h>
35 #endif
36
37 #include <math.h>
38 #include <string.h>
39 #include "sysstuff.h"
40 #include "smalloc.h"
41 #include "macros.h"
42 #include "maths.h"
43 #include "vec.h"
44 #include "pbc.h"
45 #include "nbnxn_search.h"
46 #include "nbnxn_consts.h"
47 #include "gmx_cyclecounter.h"
48 #include "gmxfio.h"
49 #include "gmx_omp_nthreads.h"
50 #include "nrnb.h"
51
52 #ifdef GMX_X86_SSE2
53 #define NBNXN_SEARCH_SSE
54
55 #ifndef GMX_DOUBLE
56 #define NBNXN_SEARCH_SSE_SINGLE
57 #include "gmx_x86_simd_single.h"
58 #else
59 #include "gmx_x86_simd_double.h"
60 #endif
61
62 #if defined NBNXN_SEARCH_SSE_SINGLE && GPU_NSUBCELL == 8
63 #define NBNXN_8BB_SSE
64 #endif
65
66 /* The width of SSE/AVX128 with single precision for bounding boxes with GPU.
67 * Here AVX-256 turns out to be slightly slower than AVX-128.
68 */
69 #define STRIDE_8BB 4
70 #define STRIDE_8BB_2LOG 2
71
72 #endif /* NBNXN_SEARCH_SSE */
73
74 /* Pair search box lower and upper corner in x,y,z.
75 * Store this in 4 iso 3 reals, which is useful with SSE.
76 * To avoid complicating the code we also use 4 without SSE.
77 */
78 #define NNBSBB_C 4
79 #define NNBSBB_B (2*NNBSBB_C)
80 /* Pair search box lower and upper bound in z only. */
81 #define NNBSBB_D 2
82 /* Pair search box lower and upper corner x,y,z indices */
83 #define BBL_X 0
84 #define BBL_Y 1
85 #define BBL_Z 2
86 #define BBU_X 4
87 #define BBU_Y 5
88 #define BBU_Z 6
89
90 /* Strides for x/f with xyz and xyzq coordinate (and charge) storage */
91 #define STRIDE_XYZ 3
92 #define STRIDE_XYZQ 4
93 /* Size of packs of x, y or z with SSE/AVX packed coords/forces */
94 #define PACK_X4 4
95 #define PACK_X8 8
96 /* Strides for a pack of 4 and 8 coordinates/forces */
97 #define STRIDE_P4 (DIM*PACK_X4)
98 #define STRIDE_P8 (DIM*PACK_X8)
99
100 /* Index of atom a into the SSE/AVX coordinate/force array */
101 #define X4_IND_A(a) (STRIDE_P4*((a) >> 2) + ((a) & (PACK_X4 - 1)))
102 #define X8_IND_A(a) (STRIDE_P8*((a) >> 3) + ((a) & (PACK_X8 - 1)))
103
104
105 #ifdef NBNXN_SEARCH_SSE
106
107 /* The functions below are macros as they are performance sensitive */
108
109 /* 4x4 list, pack=4: no complex conversion required */
110 /* i-cluster to j-cluster conversion */
111 #define CI_TO_CJ_J4(ci) (ci)
112 /* cluster index to coordinate array index conversion */
113 #define X_IND_CI_J4(ci) ((ci)*STRIDE_P4)
114 #define X_IND_CJ_J4(cj) ((cj)*STRIDE_P4)
115
116 /* 4x2 list, pack=4: j-cluster size is half the packing width */
117 /* i-cluster to j-cluster conversion */
118 #define CI_TO_CJ_J2(ci) ((ci)<<1)
119 /* cluster index to coordinate array index conversion */
120 #define X_IND_CI_J2(ci) ((ci)*STRIDE_P4)
121 #define X_IND_CJ_J2(cj) (((cj)>>1)*STRIDE_P4 + ((cj) & 1)*(PACK_X4>>1))
122
123 /* 4x8 list, pack=8: i-cluster size is half the packing width */
124 /* i-cluster to j-cluster conversion */
125 #define CI_TO_CJ_J8(ci) ((ci)>>1)
126 /* cluster index to coordinate array index conversion */
127 #define X_IND_CI_J8(ci) (((ci)>>1)*STRIDE_P8 + ((ci) & 1)*(PACK_X8>>1))
128 #define X_IND_CJ_J8(cj) ((cj)*STRIDE_P8)
129
130 /* The j-cluster size is matched to the SIMD width */
131 #ifndef GMX_DOUBLE
132 /* 128 bits can hold 4 floats */
133 #define CI_TO_CJ_S128(ci) CI_TO_CJ_J4(ci)
134 #define X_IND_CI_S128(ci) X_IND_CI_J4(ci)
135 #define X_IND_CJ_S128(cj) X_IND_CJ_J4(cj)
136 /* 256 bits can hold 8 floats */
137 #define CI_TO_CJ_S256(ci) CI_TO_CJ_J8(ci)
138 #define X_IND_CI_S256(ci) X_IND_CI_J8(ci)
139 #define X_IND_CJ_S256(cj) X_IND_CJ_J8(cj)
140 #else
141 /* 128 bits can hold 2 doubles */
142 #define CI_TO_CJ_S128(ci) CI_TO_CJ_J2(ci)
143 #define X_IND_CI_S128(ci) X_IND_CI_J2(ci)
144 #define X_IND_CJ_S128(cj) X_IND_CJ_J2(cj)
145 /* 256 bits can hold 4 doubles */
146 #define CI_TO_CJ_S256(ci) CI_TO_CJ_J4(ci)
147 #define X_IND_CI_S256(ci) X_IND_CI_J4(ci)
148 #define X_IND_CJ_S256(cj) X_IND_CJ_J4(cj)
149 #endif
150
151 #endif /* NBNXN_SEARCH_SSE */
152
153
154 /* Interaction masks for 4xN atom interactions.
155 * Bit i*CJ_SIZE + j tells if atom i and j interact.
156 */
157 /* All interaction mask is the same for all kernels */
158 #define NBNXN_INT_MASK_ALL 0xffffffff
159 /* 4x4 kernel diagonal mask */
160 #define NBNXN_INT_MASK_DIAG 0x08ce
161 /* 4x2 kernel diagonal masks */
162 #define NBNXN_INT_MASK_DIAG_J2_0 0x0002
163 #define NBNXN_INT_MASK_DIAG_J2_1 0x002F
164 /* 4x8 kernel diagonal masks */
165 #define NBNXN_INT_MASK_DIAG_J8_0 0xf0f8fcfe
166 #define NBNXN_INT_MASK_DIAG_J8_1 0x0080c0e0
167
168
169 #ifdef NBNXN_SEARCH_SSE
170 /* Store bounding boxes corners as quadruplets: xxxxyyyyzzzz */
171 #define NBNXN_BBXXXX
172 /* Size of bounding box corners quadruplet */
173 #define NNBSBB_XXXX (NNBSBB_D*DIM*STRIDE_8BB)
174 #endif
175
176 /* We shift the i-particles backward for PBC.
177 * This leads to more conditionals than shifting forward.
178 * We do this to get more balanced pair lists.
179 */
180 #define NBNXN_SHIFT_BACKWARD
181
182
183 /* This define is a lazy way to avoid interdependence of the grid
184 * and searching data structures.
185 */
186 #define NBNXN_NA_SC_MAX (GPU_NSUBCELL*NBNXN_GPU_CLUSTER_SIZE)
187
188 #ifdef NBNXN_SEARCH_SSE
189 #define GMX_MM128_HERE
190 #include "gmx_x86_simd_macros.h"
191 typedef struct nbnxn_x_ci_x86_simd128 {
192 /* The i-cluster coordinates for simple search */
193 gmx_mm_pr ix_SSE0,iy_SSE0,iz_SSE0;
194 gmx_mm_pr ix_SSE1,iy_SSE1,iz_SSE1;
195 gmx_mm_pr ix_SSE2,iy_SSE2,iz_SSE2;
196 gmx_mm_pr ix_SSE3,iy_SSE3,iz_SSE3;
197 } nbnxn_x_ci_x86_simd128_t;
198 #undef GMX_MM128_HERE
199 #ifdef GMX_X86_AVX_256
200 #define GMX_MM256_HERE
201 #include "gmx_x86_simd_macros.h"
202 typedef struct nbnxn_x_ci_x86_simd256 {
203 /* The i-cluster coordinates for simple search */
204 gmx_mm_pr ix_SSE0,iy_SSE0,iz_SSE0;
205 gmx_mm_pr ix_SSE1,iy_SSE1,iz_SSE1;
206 gmx_mm_pr ix_SSE2,iy_SSE2,iz_SSE2;
207 gmx_mm_pr ix_SSE3,iy_SSE3,iz_SSE3;
208 } nbnxn_x_ci_x86_simd256_t;
209 #undef GMX_MM256_HERE
210 #endif
211 #endif
212
213 /* Working data for the actual i-supercell during pair search */
214 typedef struct nbnxn_list_work {
215 gmx_cache_protect_t cp0; /* Protect cache between threads */
216
217 float *bb_ci; /* The bounding boxes, pbc shifted, for each cluster */
218 real *x_ci; /* The coordinates, pbc shifted, for each atom */
219 #ifdef NBNXN_SEARCH_SSE
220 nbnxn_x_ci_x86_simd128_t *x_ci_x86_simd128;
221 #ifdef GMX_X86_AVX_256
222 nbnxn_x_ci_x86_simd256_t *x_ci_x86_simd256;
223 #endif
224 #endif
225 int cj_ind; /* The current cj_ind index for the current list */
226 int cj4_init; /* The first unitialized cj4 block */
227
228 float *d2; /* Bounding box distance work array */
229
230 nbnxn_cj_t *cj; /* The j-cell list */
231 int cj_nalloc; /* Allocation size of cj */
232
233 int ncj_noq; /* Nr. of cluster pairs without Coul for flop count */
234 int ncj_hlj; /* Nr. of cluster pairs with 1/2 LJ for flop count */
235
236 gmx_cache_protect_t cp1; /* Protect cache between threads */
237 } nbnxn_list_work_t;
238
239 /* Function type for setting the i-atom coordinate working data */
240 typedef void
241 gmx_icell_set_x_t(int ci,
242 real shx,real shy,real shz,
243 int na_c,
244 int stride,const real *x,
245 nbnxn_list_work_t *work);
246
247 static gmx_icell_set_x_t icell_set_x_simple;
248 #ifdef NBNXN_SEARCH_SSE
249 static gmx_icell_set_x_t icell_set_x_simple_x86_simd128;
250 #ifdef GMX_X86_AVX_256
251 static gmx_icell_set_x_t icell_set_x_simple_x86_simd256;
252 #endif
253 #endif
254 static gmx_icell_set_x_t icell_set_x_supersub;
255 #ifdef NBNXN_SEARCH_SSE
256 static gmx_icell_set_x_t icell_set_x_supersub_sse8;
257 #endif
258
259 /* Local cycle count struct for profiling */
260 typedef struct {
261 int count;
262 gmx_cycles_t c;
263 gmx_cycles_t start;
264 } nbnxn_cycle_t;
265
266 /* Local cycle count enum for profiling */
267 enum { enbsCCgrid, enbsCCsearch, enbsCCcombine, enbsCCreducef, enbsCCnr };
268
269 /* A pair-search grid struct for one domain decomposition zone */
270 typedef struct {
271 rvec c0; /* The lower corner of the (local) grid */
272 rvec c1; /* The upper corner of the (local) grid */
273 real atom_density; /* The atom number density for the local grid */
274
275 gmx_bool bSimple; /* Is this grid simple or super/sub */
276 int na_c; /* Number of atoms per cluster */
277 int na_cj; /* Number of atoms for list j-clusters */
278 int na_sc; /* Number of atoms per super-cluster */
279 int na_c_2log; /* 2log of na_c */
280
281 int ncx; /* Number of (super-)cells along x */
282 int ncy; /* Number of (super-)cells along y */
283 int nc; /* Total number of (super-)cells */
284
285 real sx; /* x-size of a (super-)cell */
286 real sy; /* y-size of a (super-)cell */
287 real inv_sx; /* 1/sx */
288 real inv_sy; /* 1/sy */
289
290 int cell0; /* Index in nbs->cell corresponding to cell 0 */
291
292 int *cxy_na; /* The number of atoms for each column in x,y */
293 int *cxy_ind; /* Grid (super)cell index, offset from cell0 */
294 int cxy_nalloc; /* Allocation size for cxy_na and cxy_ind */
295
296 int *nsubc; /* The number of sub cells for each super cell */
297 float *bbcz; /* Bounding boxes in z for the super cells */
298 float *bb; /* 3D bounding boxes for the sub cells */
299 float *bbj; /* 3D j-b.boxes for SSE-double or AVX-single */
300 int *flags; /* Flag for the super cells */
301 int nc_nalloc; /* Allocation size for the pointers above */
302
303 float *bbcz_simple; /* bbcz for simple grid converted from super */
304 float *bb_simple; /* bb for simple grid converted from super */
305 int *flags_simple; /* flags for simple grid converted from super */
306 int nc_nalloc_simple; /* Allocation size for the pointers above */
307
308 int nsubc_tot; /* Total number of subcell, used for printing */
309 } nbnxn_grid_t;
310
311 /* Thread-local work struct, contains part of nbnxn_grid_t */
312 typedef struct {
313 gmx_cache_protect_t cp0;
314
315 int *cxy_na;
316 int cxy_na_nalloc;
317
318 int *sort_work;
319 int sort_work_nalloc;
320
321 int ndistc; /* Number of distance checks for flop counting */
322
323 nbnxn_cycle_t cc[enbsCCnr];
324
325 gmx_cache_protect_t cp1;
326 } nbnxn_search_work_t;
327
328 /* Main pair-search struct, contains the grid(s), not the pair-list(s) */
329 typedef struct nbnxn_search {
330 int ePBC; /* PBC type enum */
331 matrix box; /* The periodic unit-cell */
332
333 gmx_bool DomDec; /* Are we doing domain decomposition? */
334 ivec dd_dim; /* Are we doing DD in x,y,z? */
335 gmx_domdec_zones_t *zones; /* The domain decomposition zones */
336
337 int ngrid; /* The number of grids, equal to #DD-zones */
338 nbnxn_grid_t *grid; /* Array of grids, size ngrid */
339 int *cell; /* Actual allocated cell array for all grids */
340 int cell_nalloc; /* Allocation size of cell */
341 int *a; /* Atom index for grid, the inverse of cell */
342 int a_nalloc; /* Allocation size of a */
343
344 int natoms_local; /* The local atoms run from 0 to natoms_local */
345 int natoms_nonlocal; /* The non-local atoms run from natoms_local
346 * to natoms_nonlocal */
347
348 gmx_bool print_cycles;
349 int search_count;
350 nbnxn_cycle_t cc[enbsCCnr];
351
352 gmx_icell_set_x_t *icell_set_x; /* Function for setting i-coords */
353
354 int nthread_max; /* Maximum number of threads for pair-search */
355 nbnxn_search_work_t *work; /* Work array, size nthread_max */
356 } nbnxn_search_t_t;
357
358
359 static void nbs_cycle_clear(nbnxn_cycle_t *cc)
360 {
361 int i;
362
363 for(i=0; i<enbsCCnr; i++)
364 {
365 cc[i].count = 0;
366 cc[i].c = 0;
367 }
368 }
369
370 static void nbs_cycle_start(nbnxn_cycle_t *cc)
371 {
372 cc->start = gmx_cycles_read();
373 }
374
375 static void nbs_cycle_stop(nbnxn_cycle_t *cc)
376 {
377 cc->c += gmx_cycles_read() - cc->start;
378 cc->count++;
379 }
380
381 static double Mcyc_av(const nbnxn_cycle_t *cc)
382 {
383 return (double)cc->c*1e-6/cc->count;
384 }
385
386 static void nbs_cycle_print(FILE *fp,const nbnxn_search_t nbs)
387 {
388 int n;
389 int t;
390
391 fprintf(fp,"\n");
392 fprintf(fp,"ns %4d grid %4.1f search %4.1f red.f %5.3f",
393 nbs->cc[enbsCCgrid].count,
394 Mcyc_av(&nbs->cc[enbsCCgrid]),
395 Mcyc_av(&nbs->cc[enbsCCsearch]),
396 Mcyc_av(&nbs->cc[enbsCCreducef]));
397
398 if (nbs->nthread_max > 1)
399 {
400 if (nbs->cc[enbsCCcombine].count > 0)
401 {
402 fprintf(fp," comb %5.2f",
403 Mcyc_av(&nbs->cc[enbsCCcombine]));
404 }
405 fprintf(fp," s. th");
406 for(t=0; t<nbs->nthread_max; t++)
407 {
408 fprintf(fp," %4.1f",
409 Mcyc_av(&nbs->work[t].cc[enbsCCsearch]));
410 }
411 }
412 fprintf(fp,"\n");
413 }
414
415 static void nbnxn_grid_init(nbnxn_grid_t * grid)
416 {
417 grid->cxy_na = NULL;
418 grid->cxy_ind = NULL;
419 grid->cxy_nalloc = 0;
420 grid->bb = NULL;
421 grid->bbj = NULL;
422 grid->nc_nalloc = 0;
423 }
424
425 static int get_2log(int n)
426 {
427 int log2;
428
429 log2 = 0;
430 while ((1<<log2) < n)
431 {
432 log2++;
433 }
434 if ((1<<log2) != n)
435 {
436 gmx_fatal(FARGS,"nbnxn na_c (%d) is not a power of 2",n);
437 }
438
439 return log2;
440 }
441
442 static int kernel_to_ci_size(int nb_kernel_type)
443 {
444 switch (nb_kernel_type)
445 {
446 case nbk4x4_PlainC:
447 case nbk4xN_X86_SIMD128:
448 case nbk4xN_X86_SIMD256:
449 return NBNXN_CPU_CLUSTER_I_SIZE;
450 case nbk8x8x8_CUDA:
451 case nbk8x8x8_PlainC:
452 /* The cluster size for super/sub lists is only set here.
453 * Any value should work for the pair-search and atomdata code.
454 * The kernels, of course, might require a particular value.
455 */
456 return NBNXN_GPU_CLUSTER_SIZE;
457 default:
458 gmx_incons("unknown kernel type");
459 }
460
461 return 0;
462 }
463
464 static int kernel_to_cj_size(int nb_kernel_type)
465 {
466 switch (nb_kernel_type)
467 {
468 case nbk4x4_PlainC:
469 return NBNXN_CPU_CLUSTER_I_SIZE;
470 case nbk4xN_X86_SIMD128:
471 /* Number of reals that fit in SIMD (128 bits = 16 bytes) */
472 return 16/sizeof(real);
473 case nbk4xN_X86_SIMD256:
474 /* Number of reals that fit in SIMD (256 bits = 32 bytes) */
475 return 32/sizeof(real);
476 case nbk8x8x8_CUDA:
477 case nbk8x8x8_PlainC:
478 return kernel_to_ci_size(nb_kernel_type);
479 default:
480 gmx_incons("unknown kernel type");
481 }
482
483 return 0;
484 }
485
486 static int ci_to_cj(int na_cj_2log,int ci)
487 {
488 switch (na_cj_2log)
489 {
490 case 2: return ci; break;
491 case 1: return (ci<<1); break;
492 case 3: return (ci>>1); break;
493 }
494
495 return 0;
496 }
497
498 gmx_bool nbnxn_kernel_pairlist_simple(int nb_kernel_type)
499 {
500 if (nb_kernel_type == nbkNotSet)
501 {
502 gmx_fatal(FARGS, "Non-bonded kernel type not set for Verlet-style pair-list.");
503 }
504
505 switch (nb_kernel_type)
506 {
507 case nbk8x8x8_CUDA:
508 case nbk8x8x8_PlainC:
509 return FALSE;
510
511 case nbk4x4_PlainC:
512 case nbk4xN_X86_SIMD128:
513 case nbk4xN_X86_SIMD256:
514 return TRUE;
515
516 default:
517 gmx_incons("Invalid nonbonded kernel type passed!");
518 return FALSE;
519 }
520 }
521
522 void nbnxn_init_search(nbnxn_search_t * nbs_ptr,
523 ivec *n_dd_cells,
524 gmx_domdec_zones_t *zones,
525 int nthread_max)
526 {
527 nbnxn_search_t nbs;
528 int d,g,t;
529
530 snew(nbs,1);
531 *nbs_ptr = nbs;
532
533 nbs->DomDec = (n_dd_cells != NULL);
534
535 clear_ivec(nbs->dd_dim);
536 nbs->ngrid = 1;
537 if (nbs->DomDec)
538 {
539 nbs->zones = zones;
540
541 for(d=0; d<DIM; d++)
542 {
543 if ((*n_dd_cells)[d] > 1)
544 {
545 nbs->dd_dim[d] = 1;
546 /* Each grid matches a DD zone */
547 nbs->ngrid *= 2;
548 }
549 }
550 }
551
552 snew(nbs->grid,nbs->ngrid);
553 for(g=0; g<nbs->ngrid; g++)
554 {
555 nbnxn_grid_init(&nbs->grid[g]);
556 }
557 nbs->cell = NULL;
558 nbs->cell_nalloc = 0;
559 nbs->a = NULL;
560 nbs->a_nalloc = 0;
561
562 nbs->nthread_max = nthread_max;
563
564 /* Initialize the work data structures for each thread */
565 snew(nbs->work,nbs->nthread_max);
566 for(t=0; t<nbs->nthread_max; t++)
567 {
568 nbs->work[t].cxy_na = NULL;
569 nbs->work[t].cxy_na_nalloc = 0;
570 nbs->work[t].sort_work = NULL;
571 nbs->work[t].sort_work_nalloc = 0;
572 }
573
574 /* Initialize detailed nbsearch cycle counting */
575 nbs->print_cycles = (getenv("GMX_NBNXN_CYCLE") != 0);
576 nbs->search_count = 0;
577 nbs_cycle_clear(nbs->cc);
578 for(t=0; t<nbs->nthread_max; t++)
579 {
580 nbs_cycle_clear(nbs->work[t].cc);
581 }
582 }
583
584 static real grid_atom_density(int n,rvec corner0,rvec corner1)
585 {
586 rvec size;
587
588 rvec_sub(corner1,corner0,size);
589
590 return n/(size[XX]*size[YY]*size[ZZ]);
591 }
592
593 static int set_grid_size_xy(const nbnxn_search_t nbs,
594 nbnxn_grid_t *grid,
595 int n,rvec corner0,rvec corner1,
596 real atom_density,
597 int XFormat)
598 {
599 rvec size;
600 int na_c;
601 real adens,tlen,tlen_x,tlen_y,nc_max;
602 int t;
603
604 rvec_sub(corner1,corner0,size);
605
606 if (n > grid->na_sc)
607 {
608 /* target cell length */
609 if (grid->bSimple)
610 {
611 /* To minimize the zero interactions, we should make
612 * the largest of the i/j cell cubic.
613 */
614 na_c = max(grid->na_c,grid->na_cj);
615
616 /* Approximately cubic cells */
617 tlen = pow(na_c/atom_density,1.0/3.0);
618 tlen_x = tlen;
619 tlen_y = tlen;
620 }
621 else
622 {
623 /* Approximately cubic sub cells */
624 tlen = pow(grid->na_c/atom_density,1.0/3.0);
625 tlen_x = tlen*GPU_NSUBCELL_X;
626 tlen_y = tlen*GPU_NSUBCELL_Y;
627 }
628 /* We round ncx and ncy down, because we get less cell pairs
629 * in the nbsist when the fixed cell dimensions (x,y) are
630 * larger than the variable one (z) than the other way around.
631 */
632 grid->ncx = max(1,(int)(size[XX]/tlen_x));
633 grid->ncy = max(1,(int)(size[YY]/tlen_y));
634 }
635 else
636 {
637 grid->ncx = 1;
638 grid->ncy = 1;
639 }
640
641 /* We need one additional cell entry for particles moved by DD */
642 if (grid->ncx*grid->ncy+1 > grid->cxy_nalloc)
643 {
644 grid->cxy_nalloc = over_alloc_large(grid->ncx*grid->ncy+1);
645 srenew(grid->cxy_na,grid->cxy_nalloc);
646 srenew(grid->cxy_ind,grid->cxy_nalloc+1);
647 }
648 for(t=0; t<nbs->nthread_max; t++)
649 {
650 if (grid->ncx*grid->ncy+1 > nbs->work[t].cxy_na_nalloc)
651 {
652 nbs->work[t].cxy_na_nalloc = over_alloc_large(grid->ncx*grid->ncy+1);
653 srenew(nbs->work[t].cxy_na,nbs->work[t].cxy_na_nalloc);
654 }
655 }
656
657 /* Worst case scenario of 1 atom in each last cell */
658 if (grid->na_cj <= grid->na_c)
659 {
660 nc_max = n/grid->na_sc + grid->ncx*grid->ncy;
661 }
662 else
663 {
664 nc_max = n/grid->na_sc + grid->ncx*grid->ncy*grid->na_cj/grid->na_c;
665 }
666
667 if (nc_max > grid->nc_nalloc)
668 {
669 int bb_nalloc;
670
671 grid->nc_nalloc = over_alloc_large(nc_max);
672 srenew(grid->nsubc,grid->nc_nalloc);
673 srenew(grid->bbcz,grid->nc_nalloc*NNBSBB_D);
674 #ifdef NBNXN_8BB_SSE
675 bb_nalloc = grid->nc_nalloc*GPU_NSUBCELL/STRIDE_8BB*NNBSBB_XXXX;
676 #else
677 bb_nalloc = grid->nc_nalloc*GPU_NSUBCELL*NNBSBB_B;
678 #endif
679 sfree_aligned(grid->bb);
680 /* This snew also zeros the contents, this avoid possible
681 * floating exceptions in SSE with the unused bb elements.
682 */
683 snew_aligned(grid->bb,bb_nalloc,16);
684
685 if (grid->bSimple)
686 {
687 if (grid->na_cj == grid->na_c)
688 {
689 grid->bbj = grid->bb;
690 }
691 else
692 {
693 sfree_aligned(grid->bbj);
694 snew_aligned(grid->bbj,bb_nalloc*grid->na_c/grid->na_cj,16);
695 }
696 }
697
698 srenew(grid->flags,grid->nc_nalloc);
699 }
700
701 copy_rvec(corner0,grid->c0);
702 copy_rvec(corner1,grid->c1);
703 grid->sx = size[XX]/grid->ncx;
704 grid->sy = size[YY]/grid->ncy;
705 grid->inv_sx = 1/grid->sx;
706 grid->inv_sy = 1/grid->sy;
707
708 return nc_max;
709 }
710
711 #define SORT_GRID_OVERSIZE 2
712 #define SGSF (SORT_GRID_OVERSIZE + 1)
713
714 static void sort_atoms(int dim,gmx_bool Backwards,
715 int *a,int n,rvec *x,
716 real h0,real invh,int nsort,int *sort)
717 {
718 int i,c;
719 int zi,zim;
720 int cp,tmp;
721
722 if (n <= 1)
723 {
724 /* Nothing to do */
725 return;
726 }
727
728 /* For small oversize factors clearing the whole area is fastest.
729 * For large oversize we should clear the used elements after use.
730 */
731 for(i=0; i<nsort; i++)
732 {
733 sort[i] = -1;
734 }
735 /* Sort the particles using a simple index sort */
736 for(i=0; i<n; i++)
737 {
738 /* The cast takes care of float-point rounding effects below zero.
739 * This code assumes particles are less than 1/SORT_GRID_OVERSIZE
740 * times the box height out of the box.
741 */
742 zi = (int)((x[a[i]][dim] - h0)*invh);
743
744 #ifdef DEBUG_NBNXN_GRIDDING
745 if (zi < 0 || zi >= nsort)
746 {
747 gmx_fatal(FARGS,"(int)((x[%d][%c]=%f - %f)*%f) = %d, not in 0 - %d\n",
748 a[i],'x'+dim,x[a[i]][dim],h0,invh,zi,nsort);
749 }
750 #endif
751
752 /* Ideally this particle should go in sort cell zi,
753 * but that might already be in use,
754 * in that case find the first empty cell higher up
755 */
756 if (sort[zi] < 0)
757 {
758 sort[zi] = a[i];
759 }
760 else
761 {
762 /* We have multiple atoms in the same sorting slot.
763 * Sort on real z for minimal bounding box size.
764 * There is an extra check for identical z to ensure
765 * well-defined output order, independent of input order
766 * to ensure binary reproducibility after restarts.
767 */
768 while(sort[zi] >= 0 && ( x[a[i]][dim] > x[sort[zi]][dim] ||
769 (x[a[i]][dim] == x[sort[zi]][dim] &&
770 a[i] > sort[zi])))
771 {
772 zi++;
773 }
774
775 if (sort[zi] >= 0)
776 {
777 /* Shift all elements by one slot until we find an empty slot */
778 cp = sort[zi];
779 zim = zi + 1;
780 while (sort[zim] >= 0)
781 {
782 tmp = sort[zim];
783 sort[zim] = cp;
784 cp = tmp;
785 zim++;
786 }
787 sort[zim] = cp;
788 }
789 sort[zi] = a[i];
790 }
791 }
792
793 c = 0;
794 if (!Backwards)
795 {
796 for(zi=0; zi<nsort; zi++)
797 {
798 if (sort[zi] >= 0)
799 {
800 a[c++] = sort[zi];
801 }
802 }
803 }
804 else
805 {
806 for(zi=nsort-1; zi>=0; zi--)
807 {
808 if (sort[zi] >= 0)
809 {
810 a[c++] = sort[zi];
811 }
812 }
813 }
814 if (c < n)
815 {
816 gmx_incons("Lost particles while sorting");
817 }
818 }
819
820 #ifdef GMX_DOUBLE
821 #define R2F_D(x) ((float)((x) >= 0 ? ((1-GMX_FLOAT_EPS)*(x)) : ((1+GMX_FLOAT_EPS)*(x))))
822 #define R2F_U(x) ((float)((x) >= 0 ? ((1+GMX_FLOAT_EPS)*(x)) : ((1-GMX_FLOAT_EPS)*(x))))
823 #else
824 #define R2F_D(x) (x)
825 #define R2F_U(x) (x)
826 #endif
827
828 /* Coordinate order x,y,z, bb order xyz0 */
829 static void calc_bounding_box(int na,int stride,const real *x,float *bb)
830 {
831 int i,j;
832 real xl,xh,yl,yh,zl,zh;
833
834 i = 0;
835 xl = x[i+XX];
836 xh = x[i+XX];
837 yl = x[i+YY];
838 yh = x[i+YY];
839 zl = x[i+ZZ];
840 zh = x[i+ZZ];
841 i += stride;
842 for(j=1; j<na; j++)
843 {
844 xl = min(xl,x[i+XX]);
845 xh = max(xh,x[i+XX]);
846 yl = min(yl,x[i+YY]);
847 yh = max(yh,x[i+YY]);
848 zl = min(zl,x[i+ZZ]);
849 zh = max(zh,x[i+ZZ]);
850 i += stride;
851 }
852 /* Note: possible double to float conversion here */
853 bb[BBL_X] = R2F_D(xl);
854 bb[BBL_Y] = R2F_D(yl);
855 bb[BBL_Z] = R2F_D(zl);
856 bb[BBU_X] = R2F_U(xh);
857 bb[BBU_Y] = R2F_U(yh);
858 bb[BBU_Z] = R2F_U(zh);
859 }
860
861 /* Packed coordinates, bb order xyz0 */
862 static void calc_bounding_box_x_x4(int na,const real *x,float *bb)
863 {
864 int j;
865 real xl,xh,yl,yh,zl,zh;
866
867 xl = x[XX*PACK_X4];
868 xh = x[XX*PACK_X4];
869 yl = x[YY*PACK_X4];
870 yh = x[YY*PACK_X4];
871 zl = x[ZZ*PACK_X4];
872 zh = x[ZZ*PACK_X4];
873 for(j=1; j<na; j++)
874 {
875 xl = min(xl,x[j+XX*PACK_X4]);
876 xh = max(xh,x[j+XX*PACK_X4]);
877 yl = min(yl,x[j+YY*PACK_X4]);
878 yh = max(yh,x[j+YY*PACK_X4]);
879 zl = min(zl,x[j+ZZ*PACK_X4]);
880 zh = max(zh,x[j+ZZ*PACK_X4]);
881 }
882 /* Note: possible double to float conversion here */
883 bb[BBL_X] = R2F_D(xl);
884 bb[BBL_Y] = R2F_D(yl);
885 bb[BBL_Z] = R2F_D(zl);
886 bb[BBU_X] = R2F_U(xh);
887 bb[BBU_Y] = R2F_U(yh);
888 bb[BBU_Z] = R2F_U(zh);
889 }
890
891 /* Packed coordinates, bb order xyz0 */
892 static void calc_bounding_box_x_x8(int na,const real *x,float *bb)
893 {
894 int j;
895 real xl,xh,yl,yh,zl,zh;
896
897 xl = x[XX*PACK_X8];
898 xh = x[XX*PACK_X8];
899 yl = x[YY*PACK_X8];
900 yh = x[YY*PACK_X8];
901 zl = x[ZZ*PACK_X8];
902 zh = x[ZZ*PACK_X8];
903 for(j=1; j<na; j++)
904 {
905 xl = min(xl,x[j+XX*PACK_X8]);
906 xh = max(xh,x[j+XX*PACK_X8]);
907 yl = min(yl,x[j+YY*PACK_X8]);
908 yh = max(yh,x[j+YY*PACK_X8]);
909 zl = min(zl,x[j+ZZ*PACK_X8]);
910 zh = max(zh,x[j+ZZ*PACK_X8]);
911 }
912 /* Note: possible double to float conversion here */
913 bb[BBL_X] = R2F_D(xl);
914 bb[BBL_Y] = R2F_D(yl);
915 bb[BBL_Z] = R2F_D(zl);
916 bb[BBU_X] = R2F_U(xh);
917 bb[BBU_Y] = R2F_U(yh);
918 bb[BBU_Z] = R2F_U(zh);
919 }
920
921 #ifdef NBNXN_SEARCH_SSE
922
923 /* Packed coordinates, bb order xyz0 */
924 static void calc_bounding_box_x_x4_halves(int na,const real *x,
925 float *bb,float *bbj)
926 {
927 calc_bounding_box_x_x4(min(na,2),x,bbj);
928
929 if (na > 2)
930 {
931 calc_bounding_box_x_x4(min(na-2,2),x+(PACK_X4>>1),bbj+NNBSBB_B);
932 }
933 else
934 {
935 /* Set the "empty" bounding box to the same as the first one,
936 * so we don't need to treat special cases in the rest of the code.
937 */
938 _mm_store_ps(bbj+NNBSBB_B ,_mm_load_ps(bbj));
939 _mm_store_ps(bbj+NNBSBB_B+NNBSBB_C,_mm_load_ps(bbj+NNBSBB_C));
940 }
941
942 _mm_store_ps(bb ,_mm_min_ps(_mm_load_ps(bbj),
943 _mm_load_ps(bbj+NNBSBB_B)));
944 _mm_store_ps(bb+NNBSBB_C,_mm_max_ps(_mm_load_ps(bbj+NNBSBB_C),
945 _mm_load_ps(bbj+NNBSBB_B+NNBSBB_C)));
946 }
947
948 /* Coordinate order xyz, bb order xxxxyyyyzzzz */
949 static void calc_bounding_box_xxxx(int na,int stride,const real *x,float *bb)
950 {
951 int i,j;
952 real xl,xh,yl,yh,zl,zh;
953
954 i = 0;
955 xl = x[i+XX];
956 xh = x[i+XX];
957 yl = x[i+YY];
958 yh = x[i+YY];
959 zl = x[i+ZZ];
960 zh = x[i+ZZ];
961 i += stride;
962 for(j=1; j<na; j++)
963 {
964 xl = min(xl,x[i+XX]);
965 xh = max(xh,x[i+XX]);
966 yl = min(yl,x[i+YY]);
967 yh = max(yh,x[i+YY]);
968 zl = min(zl,x[i+ZZ]);
969 zh = max(zh,x[i+ZZ]);
970 i += stride;
971 }
972 /* Note: possible double to float conversion here */
973 bb[0*STRIDE_8BB] = R2F_D(xl);
974 bb[1*STRIDE_8BB] = R2F_D(yl);
975 bb[2*STRIDE_8BB] = R2F_D(zl);
976 bb[3*STRIDE_8BB] = R2F_U(xh);
977 bb[4*STRIDE_8BB] = R2F_U(yh);
978 bb[5*STRIDE_8BB] = R2F_U(zh);
979 }
980
981 #endif /* NBNXN_SEARCH_SSE */
982
983 #ifdef NBNXN_SEARCH_SSE_SINGLE
984
985 /* Coordinate order xyz?, bb order xyz0 */
986 static void calc_bounding_box_sse(int na,const float *x,float *bb)
987 {
988 __m128 bb_0_SSE,bb_1_SSE;
989 __m128 x_SSE;
990
991 int i;
992
993 bb_0_SSE = _mm_load_ps(x);
994 bb_1_SSE = bb_0_SSE;
995
996 for(i=1; i<na; i++)
997 {
998 x_SSE = _mm_load_ps(x+i*NNBSBB_C);
999 bb_0_SSE = _mm_min_ps(bb_0_SSE,x_SSE);
1000 bb_1_SSE = _mm_max_ps(bb_1_SSE,x_SSE);
1001 }
1002
1003 _mm_store_ps(bb ,bb_0_SSE);
1004 _mm_store_ps(bb+4,bb_1_SSE);
1005 }
1006
1007 /* Coordinate order xyz?, bb order xxxxyyyyzzzz */
1008 static void calc_bounding_box_xxxx_sse(int na,const float *x,
1009 float *bb_work,
1010 real *bb)
1011 {
1012 calc_bounding_box_sse(na,x,bb_work);
1013
1014 bb[0*STRIDE_8BB] = bb_work[BBL_X];
1015 bb[1*STRIDE_8BB] = bb_work[BBL_Y];
1016 bb[2*STRIDE_8BB] = bb_work[BBL_Z];
1017 bb[3*STRIDE_8BB] = bb_work[BBU_X];
1018 bb[4*STRIDE_8BB] = bb_work[BBU_Y];
1019 bb[5*STRIDE_8BB] = bb_work[BBU_Z];
1020 }
1021
1022 #endif /* NBNXN_SEARCH_SSE_SINGLE */
1023
1024 #ifdef NBNXN_SEARCH_SSE
1025
1026 /* Combines pairs of consecutive bounding boxes */
1027 static void combine_bounding_box_pairs(nbnxn_grid_t *grid,const float *bb)
1028 {
1029 int i,j,sc2,nc2,c2;
1030 __m128 min_SSE,max_SSE;
1031
1032 for(i=0; i<grid->ncx*grid->ncy; i++)
1033 {
1034 /* Starting bb in a column is expected to be 2-aligned */
1035 sc2 = grid->cxy_ind[i]>>1;
1036 /* For odd numbers skip the last bb here */
1037 nc2 = (grid->cxy_na[i]+3)>>(2+1);
1038 for(c2=sc2; c2<sc2+nc2; c2++)
1039 {
1040 min_SSE = _mm_min_ps(_mm_load_ps(bb+(c2*4+0)*NNBSBB_C),
1041 _mm_load_ps(bb+(c2*4+2)*NNBSBB_C));
1042 max_SSE = _mm_max_ps(_mm_load_ps(bb+(c2*4+1)*NNBSBB_C),
1043 _mm_load_ps(bb+(c2*4+3)*NNBSBB_C));
1044 _mm_store_ps(grid->bbj+(c2*2+0)*NNBSBB_C,min_SSE);
1045 _mm_store_ps(grid->bbj+(c2*2+1)*NNBSBB_C,max_SSE);
1046 }
1047 if (((grid->cxy_na[i]+3)>>2) & 1)
1048 {
1049 /* Copy the last bb for odd bb count in this column */
1050 for(j=0; j<NNBSBB_C; j++)
1051 {
1052 grid->bbj[(c2*2+0)*NNBSBB_C+j] = bb[(c2*4+0)*NNBSBB_C+j];
1053 grid->bbj[(c2*2+1)*NNBSBB_C+j] = bb[(c2*4+1)*NNBSBB_C+j];
1054 }
1055 }
1056 }
1057 }
1058
1059 #endif
1060
1061
1062 /* Prints the average bb size, used for debug output */
1063 static void print_bbsizes_simple(FILE *fp,
1064 const nbnxn_search_t nbs,
1065 const nbnxn_grid_t *grid)
1066 {
1067 int c,d;
1068 dvec ba;
1069
1070 clear_dvec(ba);
1071 for(c=0; c<grid->nc; c++)
1072 {
1073 for(d=0; d<DIM; d++)
1074 {
1075 ba[d] += grid->bb[c*NNBSBB_B+NNBSBB_C+d] - grid->bb[c*NNBSBB_B+d];
1076 }
1077 }
1078 dsvmul(1.0/grid->nc,ba,ba);
1079
1080 fprintf(fp,"ns bb: %4.2f %4.2f %4.2f %4.2f %4.2f %4.2f rel %4.2f %4.2f %4.2f\n",
1081 nbs->box[XX][XX]/grid->ncx,
1082 nbs->box[YY][YY]/grid->ncy,
1083 nbs->box[ZZ][ZZ]*grid->ncx*grid->ncy/grid->nc,
1084 ba[XX],ba[YY],ba[ZZ],
1085 ba[XX]*grid->ncx/nbs->box[XX][XX],
1086 ba[YY]*grid->ncy/nbs->box[YY][YY],
1087 ba[ZZ]*grid->nc/(grid->ncx*grid->ncy*nbs->box[ZZ][ZZ]));
1088 }
1089
1090 /* Prints the average bb size, used for debug output */
1091 static void print_bbsizes_supersub(FILE *fp,
1092 const nbnxn_search_t nbs,
1093 const nbnxn_grid_t *grid)
1094 {
1095 int ns,c,s;
1096 dvec ba;
1097
1098 clear_dvec(ba);
1099 ns = 0;
1100 for(c=0; c<grid->nc; c++)
1101 {
1102 #ifdef NBNXN_BBXXXX
1103 for(s=0; s<grid->nsubc[c]; s+=STRIDE_8BB)
1104 {
1105 int cs_w,i,d;
1106
1107 cs_w = (c*GPU_NSUBCELL + s)/STRIDE_8BB;
1108 for(i=0; i<STRIDE_8BB; i++)
1109 {
1110 for(d=0; d<DIM; d++)
1111 {
1112 ba[d] +=
1113 grid->bb[cs_w*NNBSBB_XXXX+(DIM+d)*STRIDE_8BB+i] -
1114 grid->bb[cs_w*NNBSBB_XXXX+ d *STRIDE_8BB+i];
1115 }
1116 }
1117 }
1118 #else
1119 for(s=0; s<grid->nsubc[c]; s++)
1120 {
1121 int cs,d;
1122
1123 cs = c*GPU_NSUBCELL + s;
1124 for(d=0; d<DIM; d++)
1125 {
1126 ba[d] +=
1127 grid->bb[cs*NNBSBB_B+NNBSBB_C+d] -
1128 grid->bb[cs*NNBSBB_B +d];
1129 }
1130 }
1131 #endif
1132 ns += grid->nsubc[c];
1133 }
1134 dsvmul(1.0/ns,ba,ba);
1135
1136 fprintf(fp,"ns bb: %4.2f %4.2f %4.2f %4.2f %4.2f %4.2f rel %4.2f %4.2f %4.2f\n",
1137 nbs->box[XX][XX]/(grid->ncx*GPU_NSUBCELL_X),
1138 nbs->box[YY][YY]/(grid->ncy*GPU_NSUBCELL_Y),
1139 nbs->box[ZZ][ZZ]*grid->ncx*grid->ncy/(grid->nc*GPU_NSUBCELL_Z),
1140 ba[XX],ba[YY],ba[ZZ],
1141 ba[XX]*grid->ncx*GPU_NSUBCELL_X/nbs->box[XX][XX],
1142 ba[YY]*grid->ncy*GPU_NSUBCELL_Y/nbs->box[YY][YY],
1143 ba[ZZ]*grid->nc*GPU_NSUBCELL_Z/(grid->ncx*grid->ncy*nbs->box[ZZ][ZZ]));
1144 }
1145
1146 static void copy_int_to_nbat_int(const int *a,int na,int na_round,
1147 const int *in,int fill,int *innb)
1148 {
1149 int i,j;
1150
1151 j = 0;
1152 for(i=0; i<na; i++)
1153 {
1154 innb[j++] = in[a[i]];
1155 }
1156 /* Complete the partially filled last cell with fill */
1157 for(; i<na_round; i++)
1158 {
1159 innb[j++] = fill;
1160 }
1161 }
1162
1163 static void clear_nbat_real(int na,int nbatFormat,real *xnb,int a0)
1164 {
1165 int a,d,j,c;
1166
1167 switch (nbatFormat)
1168 {
1169 case nbatXYZ:
1170 for(a=0; a<na; a++)
1171 {
1172 for(d=0; d<DIM; d++)
1173 {
1174 xnb[(a0+a)*STRIDE_XYZ+d] = 0;
1175 }
1176 }
1177 break;
1178 case nbatXYZQ:
1179 for(a=0; a<na; a++)
1180 {
1181 for(d=0; d<DIM; d++)
1182 {
1183 xnb[(a0+a)*STRIDE_XYZQ+d] = 0;
1184 }
1185 }
1186 break;
1187 case nbatX4:
1188 j = X4_IND_A(a0);
1189 c = a0 & (PACK_X4-1);
1190 for(a=0; a<na; a++)
1191 {
1192 xnb[j+XX*PACK_X4] = 0;
1193 xnb[j+YY*PACK_X4] = 0;
1194 xnb[j+ZZ*PACK_X4] = 0;
1195 j++;
1196 c++;
1197 if (c == PACK_X4)
1198 {
1199 j += (DIM-1)*PACK_X4;
1200 c = 0;
1201 }
1202 }
1203 break;
1204 case nbatX8:
1205 j = X8_IND_A(a0);
1206 c = a0 & (PACK_X8-1);
1207 for(a=0; a<na; a++)
1208 {
1209 xnb[j+XX*PACK_X8] = 0;
1210 xnb[j+YY*PACK_X8] = 0;
1211 xnb[j+ZZ*PACK_X8] = 0;
1212 j++;
1213 c++;
1214 if (c == PACK_X8)
1215 {
1216 j += (DIM-1)*PACK_X8;
1217 c = 0;
1218 }
1219 }
1220 break;
1221 }
1222 }
1223
1224 static void copy_rvec_to_nbat_real(const int *a,int na,int na_round,
1225 rvec *x,int nbatFormat,real *xnb,int a0,
1226 int cx,int cy,int cz)
1227 {
1228 int i,j,c;
1229
1230 /* We might need to place filler particles to fill up the cell to na_round.
1231 * The coefficients (LJ and q) for such particles are zero.
1232 * But we might still get NaN as 0*NaN when distances are too small.
1233 * We hope that -107 nm is far away enough from to zero
1234 * to avoid accidental short distances to particles shifted down for pbc.
1235 */
1236 #define NBAT_FAR_AWAY 107
1237
1238 switch (nbatFormat)
1239 {
1240 case nbatXYZ:
1241 j = a0*STRIDE_XYZ;
1242 for(i=0; i<na; i++)
1243 {
1244 xnb[j++] = x[a[i]][XX];
1245 xnb[j++] = x[a[i]][YY];
1246 xnb[j++] = x[a[i]][ZZ];
1247 }
1248 /* Complete the partially filled last cell with copies of the last element.
1249 * This simplifies the bounding box calculation and avoid
1250 * numerical issues with atoms that are coincidentally close.
1251 */
1252 for(; i<na_round; i++)
1253 {
1254 xnb[j++] = -NBAT_FAR_AWAY*(1 + cx);
1255 xnb[j++] = -NBAT_FAR_AWAY*(1 + cy);
1256 xnb[j++] = -NBAT_FAR_AWAY*(1 + cz + i);
1257 }
1258 break;
1259 case nbatXYZQ:
1260 j = a0*STRIDE_XYZQ;
1261 for(i=0; i<na; i++)
1262 {
1263 xnb[j++] = x[a[i]][XX];
1264 xnb[j++] = x[a[i]][YY];
1265 xnb[j++] = x[a[i]][ZZ];
1266 j++;
1267 }
1268 /* Complete the partially filled last cell with particles far apart */
1269 for(; i<na_round; i++)
1270 {
1271 xnb[j++] = -NBAT_FAR_AWAY*(1 + cx);
1272 xnb[j++] = -NBAT_FAR_AWAY*(1 + cy);
1273 xnb[j++] = -NBAT_FAR_AWAY*(1 + cz + i);
1274 j++;
1275 }
1276 break;
1277 case nbatX4:
1278 j = X4_IND_A(a0);
1279 c = a0 & (PACK_X4-1);
1280 for(i=0; i<na; i++)
1281 {
1282 xnb[j+XX*PACK_X4] = x[a[i]][XX];
1283 xnb[j+YY*PACK_X4] = x[a[i]][YY];
1284 xnb[j+ZZ*PACK_X4] = x[a[i]][ZZ];
1285 j++;
1286 c++;
1287 if (c == PACK_X4)
1288 {
1289 j += (DIM-1)*PACK_X4;
1290 c = 0;
1291 }
1292 }
1293 /* Complete the partially filled last cell with particles far apart */
1294 for(; i<na_round; i++)
1295 {
1296 xnb[j+XX*PACK_X4] = -NBAT_FAR_AWAY*(1 + cx);
1297 xnb[j+YY*PACK_X4] = -NBAT_FAR_AWAY*(1 + cy);
1298 xnb[j+ZZ*PACK_X4] = -NBAT_FAR_AWAY*(1 + cz + i);
1299 j++;
1300 c++;
1301 if (c == PACK_X4)
1302 {
1303 j += (DIM-1)*PACK_X4;
1304 c = 0;
1305 }
1306 }
1307 break;
1308 case nbatX8:
1309 j = X8_IND_A(a0);
1310 c = a0 & (PACK_X8 - 1);
1311 for(i=0; i<na; i++)
1312 {
1313 xnb[j+XX*PACK_X8] = x[a[i]][XX];
1314 xnb[j+YY*PACK_X8] = x[a[i]][YY];
1315 xnb[j+ZZ*PACK_X8] = x[a[i]][ZZ];
1316 j++;
1317 c++;
1318 if (c == PACK_X8)
1319 {
1320 j += (DIM-1)*PACK_X8;
1321 c = 0;
1322 }
1323 }
1324 /* Complete the partially filled last cell with particles far apart */
1325 for(; i<na_round; i++)
1326 {
1327 xnb[j+XX*PACK_X8] = -NBAT_FAR_AWAY*(1 + cx);
1328 xnb[j+YY*PACK_X8] = -NBAT_FAR_AWAY*(1 + cy);
1329 xnb[j+ZZ*PACK_X8] = -NBAT_FAR_AWAY*(1 + cz + i);
1330 j++;
1331 c++;
1332 if (c == PACK_X8)
1333 {
1334 j += (DIM-1)*PACK_X8;
1335 c = 0;
1336 }
1337 }
1338 break;
1339 default:
1340 gmx_incons("Unsupported stride");
1341 }
1342 }
1343
1344 /* Potentially sorts atoms on LJ coefficients !=0 and ==0.
1345 * Also sets interaction flags.
1346 */
1347 void sort_on_lj(nbnxn_atomdata_t *nbat,int na_c,
1348 int a0,int a1,const int *atinfo,
1349 int *order,
1350 int *flags)
1351 {
1352 int subc,s,a,n1,n2,a_lj_max,i,j;
1353 int sort1[NBNXN_NA_SC_MAX/GPU_NSUBCELL];
1354 int sort2[NBNXN_NA_SC_MAX/GPU_NSUBCELL];
1355 gmx_bool haveQ;
1356
1357 *flags = 0;
1358
1359 subc = 0;
1360 for(s=a0; s<a1; s+=na_c)
1361 {
1362 /* Make lists for this (sub-)cell on atoms with and without LJ */
1363 n1 = 0;
1364 n2 = 0;
1365 haveQ = FALSE;
1366 a_lj_max = -1;
1367 for(a=s; a<min(s+na_c,a1); a++)
1368 {
1369 haveQ = haveQ || GET_CGINFO_HAS_Q(atinfo[order[a]]);
1370
1371 if (GET_CGINFO_HAS_VDW(atinfo[order[a]]))
1372 {
1373 sort1[n1++] = order[a];
1374 a_lj_max = a;
1375 }
1376 else
1377 {
1378 sort2[n2++] = order[a];
1379 }
1380 }
1381
1382 /* If we don't have atom with LJ, there's nothing to sort */
1383 if (n1 > 0)
1384 {
1385 *flags |= NBNXN_CI_DO_LJ(subc);
1386
1387 if (2*n1 <= na_c)
1388 {
1389 /* Only sort when strictly necessary. Ordering particles
1390 * Ordering particles can lead to less accurate summation
1391 * due to rounding, both for LJ and Coulomb interactions.
1392 */
1393 if (2*(a_lj_max - s) >= na_c)
1394 {
1395 for(i=0; i<n1; i++)
1396 {
1397 order[a0+i] = sort1[i];
1398 }
1399 for(j=0; j<n2; j++)
1400 {
1401 order[a0+n1+j] = sort2[j];
1402 }
1403 }
1404
1405 *flags |= NBNXN_CI_HALF_LJ(subc);
1406 }
1407 }
1408 if (haveQ)
1409 {
1410 *flags |= NBNXN_CI_DO_COUL(subc);
1411 }
1412 subc++;
1413 }
1414 }
1415
1416 /* Fill a pair search cell with atoms.
1417 * Potentially sorts atoms and sets the interaction flags.
1418 */
1419 void fill_cell(const nbnxn_search_t nbs,
1420 nbnxn_grid_t *grid,
1421 nbnxn_atomdata_t *nbat,
1422 int a0,int a1,
1423 const int *atinfo,
1424 rvec *x,
1425 int sx,int sy, int sz,
1426 float *bb_work)
1427 {
1428 int na,a;
1429 size_t offset;
1430 float *bb_ptr;
1431
1432 na = a1 - a0;
1433
1434 if (grid->bSimple)
1435 {
1436 sort_on_lj(nbat,grid->na_c,a0,a1,atinfo,nbs->a,
1437 grid->flags+(a0>>grid->na_c_2log)-grid->cell0);
1438 }
1439
1440 /* Now we have sorted the atoms, set the cell indices */
1441 for(a=a0; a<a1; a++)
1442 {
1443 nbs->cell[nbs->a[a]] = a;
1444 }
1445
1446 copy_rvec_to_nbat_real(nbs->a+a0,a1-a0,grid->na_c,x,
1447 nbat->XFormat,nbat->x,a0,
1448 sx,sy,sz);
1449
1450 if (nbat->XFormat == nbatX4)
1451 {
1452 /* Store the bounding boxes as xyz.xyz. */
1453 offset = ((a0 - grid->cell0*grid->na_sc)>>grid->na_c_2log)*NNBSBB_B;
1454 bb_ptr = grid->bb + offset;
1455
1456 #if defined GMX_DOUBLE && defined NBNXN_SEARCH_SSE
1457 if (2*grid->na_cj == grid->na_c)
1458 {
1459 calc_bounding_box_x_x4_halves(na,nbat->x+X4_IND_A(a0),bb_ptr,
1460 grid->bbj+offset*2);
1461 }
1462 else
1463 #endif
1464 {
1465 calc_bounding_box_x_x4(na,nbat->x+X4_IND_A(a0),bb_ptr);
1466 }
1467 }
1468 else if (nbat->XFormat == nbatX8)
1469 {
1470 /* Store the bounding boxes as xyz.xyz. */
1471 offset = ((a0 - grid->cell0*grid->na_sc)>>grid->na_c_2log)*NNBSBB_B;
1472 bb_ptr = grid->bb + offset;
1473
1474 calc_bounding_box_x_x8(na,nbat->x+X8_IND_A(a0),bb_ptr);
1475 }
1476 #ifdef NBNXN_BBXXXX
1477 else if (!grid->bSimple)
1478 {
1479 /* Store the bounding boxes in a format convenient
1480 * for SSE calculations: xxxxyyyyzzzz...
1481 */
1482 bb_ptr =
1483 grid->bb +
1484 ((a0-grid->cell0*grid->na_sc)>>(grid->na_c_2log+STRIDE_8BB_2LOG))*NNBSBB_XXXX +
1485 (((a0-grid->cell0*grid->na_sc)>>grid->na_c_2log) & (STRIDE_8BB-1));
1486
1487 #ifdef NBNXN_SEARCH_SSE_SINGLE
1488 if (nbat->XFormat == nbatXYZQ)
1489 {
1490 calc_bounding_box_xxxx_sse(na,nbat->x+a0*nbat->xstride,
1491 bb_work,bb_ptr);
1492 }
1493 else
1494 #endif
1495 {
1496 calc_bounding_box_xxxx(na,nbat->xstride,nbat->x+a0*nbat->xstride,
1497 bb_ptr);
1498 }
1499 if (gmx_debug_at)
1500 {
1501 fprintf(debug,"%2d %2d %2d bb %5.2f %5.2f %5.2f %5.2f %5.2f %5.2f\n",
1502 sx,sy,sz,
1503 bb_ptr[0*STRIDE_8BB],bb_ptr[3*STRIDE_8BB],
1504 bb_ptr[1*STRIDE_8BB],bb_ptr[4*STRIDE_8BB],
1505 bb_ptr[2*STRIDE_8BB],bb_ptr[5*STRIDE_8BB]);
1506 }
1507 }
1508 #endif
1509 else
1510 {
1511 /* Store the bounding boxes as xyz.xyz. */
1512 bb_ptr = grid->bb+((a0-grid->cell0*grid->na_sc)>>grid->na_c_2log)*NNBSBB_B;
1513
1514 calc_bounding_box(na,nbat->xstride,nbat->x+a0*nbat->xstride,
1515 bb_ptr);
1516
1517 if (gmx_debug_at)
1518 {
1519 int bbo;
1520 bbo = (a0 - grid->cell0*grid->na_sc)/grid->na_c;
1521 fprintf(debug,"%2d %2d %2d bb %5.2f %5.2f %5.2f %5.2f %5.2f %5.2f\n",
1522 sx,sy,sz,
1523 (grid->bb+bbo*NNBSBB_B)[BBL_X],
1524 (grid->bb+bbo*NNBSBB_B)[BBU_X],
1525 (grid->bb+bbo*NNBSBB_B)[BBL_Y],
1526 (grid->bb+bbo*NNBSBB_B)[BBU_Y],
1527 (grid->bb+bbo*NNBSBB_B)[BBL_Z],
1528 (grid->bb+bbo*NNBSBB_B)[BBU_Z]);
1529 }
1530 }
1531 }
1532
1533 /* Spatially sort the atoms within one grid column */
1534 static void sort_columns_simple(const nbnxn_search_t nbs,
1535 int dd_zone,
1536 nbnxn_grid_t *grid,
1537 int a0,int a1,
1538 const int *atinfo,
1539 rvec *x,
1540 nbnxn_atomdata_t *nbat,
1541 int cxy_start,int cxy_end,
1542 int *sort_work)
1543 {
1544 int cxy;
1545 int cx,cy,cz,ncz,cfilled,c;
1546 int na,ash,ind,a;
1547 int na_c,ash_c;
1548
1549 if (debug)
1550 {
1551 fprintf(debug,"cell0 %d sorting columns %d - %d, atoms %d - %d\n",
1552 grid->cell0,cxy_start,cxy_end,a0,a1);
1553 }
1554
1555 /* Sort the atoms within each x,y column in 3 dimensions */
1556 for(cxy=cxy_start; cxy<cxy_end; cxy++)
1557 {
1558 cx = cxy/grid->ncy;
1559 cy = cxy - cx*grid->ncy;
1560
1561 na = grid->cxy_na[cxy];
1562 ncz = grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy];
1563 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1564
1565 /* Sort the atoms within each x,y column on z coordinate */
1566 sort_atoms(ZZ,FALSE,
1567 nbs->a+ash,na,x,
1568 grid->c0[ZZ],
1569 ncz*grid->na_sc*SORT_GRID_OVERSIZE/nbs->box[ZZ][ZZ],
1570 ncz*grid->na_sc*SGSF,sort_work);
1571
1572 /* Fill the ncz cells in this column */
1573 cfilled = grid->cxy_ind[cxy];
1574 for(cz=0; cz<ncz; cz++)
1575 {
1576 c = grid->cxy_ind[cxy] + cz ;
1577
1578 ash_c = ash + cz*grid->na_sc;
1579 na_c = min(grid->na_sc,na-(ash_c-ash));
1580
1581 fill_cell(nbs,grid,nbat,
1582 ash_c,ash_c+na_c,atinfo,x,
1583 grid->na_sc*cx + (dd_zone >> 2),
1584 grid->na_sc*cy + (dd_zone & 3),
1585 grid->na_sc*cz,
1586 NULL);
1587
1588 /* This copy to bbcz is not really necessary.
1589 * But it allows to use the same grid search code
1590 * for the simple and supersub cell setups.
1591 */
1592 if (na_c > 0)
1593 {
1594 cfilled = c;
1595 }
1596 grid->bbcz[c*NNBSBB_D ] = grid->bb[cfilled*NNBSBB_B+2];
1597 grid->bbcz[c*NNBSBB_D+1] = grid->bb[cfilled*NNBSBB_B+6];
1598 }
1599
1600 /* Set the unused atom indices to -1 */
1601 for(ind=na; ind<ncz*grid->na_sc; ind++)
1602 {
1603 nbs->a[ash+ind] = -1;
1604 }
1605 }
1606 }
1607
1608 /* Spatially sort the atoms within one grid column */
1609 static void sort_columns_supersub(const nbnxn_search_t nbs,
1610 int dd_zone,
1611 nbnxn_grid_t *grid,
1612 int a0,int a1,
1613 const int *atinfo,
1614 rvec *x,
1615 nbnxn_atomdata_t *nbat,
1616 int cxy_start,int cxy_end,
1617 int *sort_work)
1618 {
1619 int cxy;
1620 int cx,cy,cz=-1,c=-1,ncz;
1621 int na,ash,na_c,ind,a;
1622 int subdiv_z,sub_z,na_z,ash_z;
1623 int subdiv_y,sub_y,na_y,ash_y;
1624 int subdiv_x,sub_x,na_x,ash_x;
1625
1626 /* cppcheck-suppress unassignedVariable */
1627 float bb_work_array[NNBSBB_B+3],*bb_work_align;
1628
1629 bb_work_align = (float *)(((size_t)(bb_work_array+3)) & (~((size_t)15)));
1630
1631 if (debug)
1632 {
1633 fprintf(debug,"cell0 %d sorting columns %d - %d, atoms %d - %d\n",
1634 grid->cell0,cxy_start,cxy_end,a0,a1);
1635 }
1636
1637 subdiv_x = grid->na_c;
1638 subdiv_y = GPU_NSUBCELL_X*subdiv_x;
1639 subdiv_z = GPU_NSUBCELL_Y*subdiv_y;
1640
1641 /* Sort the atoms within each x,y column in 3 dimensions */
1642 for(cxy=cxy_start; cxy<cxy_end; cxy++)
1643 {
1644 cx = cxy/grid->ncy;
1645 cy = cxy - cx*grid->ncy;
1646
1647 na = grid->cxy_na[cxy];
1648 ncz = grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy];
1649 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
1650
1651 /* Sort the atoms within each x,y column on z coordinate */
1652 sort_atoms(ZZ,FALSE,
1653 nbs->a+ash,na,x,
1654 grid->c0[ZZ],
1655 ncz*grid->na_sc*SORT_GRID_OVERSIZE/nbs->box[ZZ][ZZ],
1656 ncz*grid->na_sc*SGSF,sort_work);
1657
1658 /* This loop goes over the supercells and subcells along z at once */
1659 for(sub_z=0; sub_z<ncz*GPU_NSUBCELL_Z; sub_z++)
1660 {
1661 ash_z = ash + sub_z*subdiv_z;
1662 na_z = min(subdiv_z,na-(ash_z-ash));
1663
1664 /* We have already sorted on z */
1665
1666 if (sub_z % GPU_NSUBCELL_Z == 0)
1667 {
1668 cz = sub_z/GPU_NSUBCELL_Z;
1669 c = grid->cxy_ind[cxy] + cz ;
1670
1671 /* The number of atoms in this supercell */
1672 na_c = min(grid->na_sc,na-(ash_z-ash));
1673
1674 grid->nsubc[c] = min(GPU_NSUBCELL,(na_c+grid->na_c-1)/grid->na_c);
1675
1676 /* Store the z-boundaries of the super cell */
1677 grid->bbcz[c*NNBSBB_D ] = x[nbs->a[ash_z]][ZZ];
1678 grid->bbcz[c*NNBSBB_D+1] = x[nbs->a[ash_z+na_c-1]][ZZ];
1679 }
1680
1681 #if GPU_NSUBCELL_Y > 1
1682 /* Sort the atoms along y */
1683 sort_atoms(YY,(sub_z & 1),
1684 nbs->a+ash_z,na_z,x,
1685 grid->c0[YY]+cy*grid->sy,grid->inv_sy,
1686 subdiv_y*SGSF,sort_work);
1687 #endif
1688
1689 for(sub_y=0; sub_y<GPU_NSUBCELL_Y; sub_y++)
1690 {
1691 ash_y = ash_z + sub_y*subdiv_y;
1692 na_y = min(subdiv_y,na-(ash_y-ash));
1693
1694 #if GPU_NSUBCELL_X > 1
1695 /* Sort the atoms along x */
1696 sort_atoms(XX,((cz*GPU_NSUBCELL_Y + sub_y) & 1),
1697 nbs->a+ash_y,na_y,x,
1698 grid->c0[XX]+cx*grid->sx,grid->inv_sx,
1699 subdiv_x*SGSF,sort_work);
1700 #endif
1701
1702 for(sub_x=0; sub_x<GPU_NSUBCELL_X; sub_x++)
1703 {
1704 ash_x = ash_y + sub_x*subdiv_x;
1705 na_x = min(subdiv_x,na-(ash_x-ash));
1706
1707 fill_cell(nbs,grid,nbat,
1708 ash_x,ash_x+na_x,atinfo,x,
1709 grid->na_c*(cx*GPU_NSUBCELL_X+sub_x) + (dd_zone >> 2),
1710 grid->na_c*(cy*GPU_NSUBCELL_Y+sub_y) + (dd_zone & 3),
1711 grid->na_c*sub_z,
1712 bb_work_align);
1713 }
1714 }
1715 }
1716
1717 /* Set the unused atom indices to -1 */
1718 for(ind=na; ind<ncz*grid->na_sc; ind++)
1719 {
1720 nbs->a[ash+ind] = -1;
1721 }
1722 }
1723 }
1724
1725 /* Determine in which grid column atoms should go */
1726 static void calc_column_indices(nbnxn_grid_t *grid,
1727 int a0,int a1,
1728 rvec *x,const int *move,
1729 int thread,int nthread,
1730 int *cell,
1731 int *cxy_na)
1732 {
1733 int n0,n1,i;
1734 int cx,cy;
1735
1736 /* We add one extra cell for particles which moved during DD */
1737 for(i=0; i<grid->ncx*grid->ncy+1; i++)
1738 {
1739 cxy_na[i] = 0;
1740 }
1741
1742 n0 = a0 + (int)((thread+0)*(a1 - a0))/nthread;
1743 n1 = a0 + (int)((thread+1)*(a1 - a0))/nthread;
1744 for(i=n0; i<n1; i++)
1745 {
1746 if (move == NULL || move[i] >= 0)
1747 {
1748 /* We need to be careful with rounding,
1749 * particles might be a few bits outside the local box.
1750 * The int cast takes care of the lower bound,
1751 * we need to explicitly take care of the upper bound.
1752 */
1753 cx = (int)((x[i][XX] - grid->c0[XX])*grid->inv_sx);
1754 if (cx == grid->ncx)
1755 {
1756 cx = grid->ncx - 1;
1757 }
1758 cy = (int)((x[i][YY] - grid->c0[YY])*grid->inv_sy);
1759 if (cy == grid->ncy)
1760 {
1761 cy = grid->ncy - 1;
1762 }
1763 /* For the moment cell contains only the, grid local,
1764 * x and y indices, not z.
1765 */
1766 cell[i] = cx*grid->ncy + cy;
1767
1768 #ifdef DEBUG_NBNXN_GRIDDING
1769 if (cell[i] < 0 || cell[i] >= grid->ncx*grid->ncy)
1770 {
1771 gmx_fatal(FARGS,
1772 "grid cell cx %d cy %d out of range (max %d %d)",
1773 cx,cy,grid->ncx,grid->ncy);
1774 }
1775 #endif
1776 }
1777 else
1778 {
1779 /* Put this moved particle after the end of the grid,
1780 * so we can process it later without using conditionals.
1781 */
1782 cell[i] = grid->ncx*grid->ncy;
1783 }
1784
1785 cxy_na[cell[i]]++;
1786 }
1787 }
1788
1789 /* Determine in which grid cells the atoms should go */
1790 static void calc_cell_indices(const nbnxn_search_t nbs,
1791 int dd_zone,
1792 nbnxn_grid_t *grid,
1793 int a0,int a1,
1794 const int *atinfo,
1795 rvec *x,
1796 const int *move,
1797 nbnxn_atomdata_t *nbat)
1798 {
1799 int n0,n1,i;
1800 int cx,cy,cxy,ncz_max,ncz;
1801 int nthread,thread;
1802 int *cxy_na,cxy_na_i;
1803
1804 nthread = gmx_omp_nthreads_get(emntPairsearch);
1805
1806 #pragma omp parallel for num_threads(nthread) schedule(static)
1807 for(thread=0; thread<nthread; thread++)
1808 {
1809 calc_column_indices(grid,a0,a1,x,move,thread,nthread,
1810 nbs->cell,nbs->work[thread].cxy_na);
1811 }
1812
1813 /* Make the cell index as a function of x and y */
1814 ncz_max = 0;
1815 ncz = 0;
1816 grid->cxy_ind[0] = 0;
1817 for(i=0; i<grid->ncx*grid->ncy+1; i++)
1818 {
1819 /* We set ncz_max at the beginning of the loop iso at the end
1820 * to skip i=grid->ncx*grid->ncy which are moved particles
1821 * that do not need to be ordered on the grid.
1822 */
1823 if (ncz > ncz_max)
1824 {
1825 ncz_max = ncz;
1826 }
1827 cxy_na_i = nbs->work[0].cxy_na[i];
1828 for(thread=1; thread<nthread; thread++)
1829 {
1830 cxy_na_i += nbs->work[thread].cxy_na[i];
1831 }
1832 ncz = (cxy_na_i + grid->na_sc - 1)/grid->na_sc;
1833 if (nbat->XFormat == nbatX8)
1834 {
1835 /* Make the number of cell a multiple of 2 */
1836 ncz = (ncz + 1) & ~1;
1837 }
1838 grid->cxy_ind[i+1] = grid->cxy_ind[i] + ncz;
1839 /* Clear cxy_na, so we can reuse the array below */
1840 grid->cxy_na[i] = 0;
1841 }
1842 grid->nc = grid->cxy_ind[grid->ncx*grid->ncy] - grid->cxy_ind[0];
1843
1844 nbat->natoms = (grid->cell0 + grid->nc)*grid->na_sc;
1845
1846 if (debug)
1847 {
1848 fprintf(debug,"ns na_sc %d na_c %d super-cells: %d x %d y %d z %.1f maxz %d\n",
1849 grid->na_sc,grid->na_c,grid->nc,
1850 grid->ncx,grid->ncy,grid->nc/((double)(grid->ncx*grid->ncy)),
1851 ncz_max);
1852 if (gmx_debug_at)
1853 {
1854 i = 0;
1855 for(cy=0; cy<grid->ncy; cy++)
1856 {
1857 for(cx=0; cx<grid->ncx; cx++)
1858 {
1859 fprintf(debug," %2d",grid->cxy_ind[i+1]-grid->cxy_ind[i]);
1860 i++;
1861 }
1862 fprintf(debug,"\n");
1863 }
1864 }
1865 }
1866
1867 /* Make sure the work array for sorting is large enough */
1868 if (ncz_max*grid->na_sc*SGSF > nbs->work[0].sort_work_nalloc)
1869 {
1870 for(thread=0; thread<nbs->nthread_max; thread++)
1871 {
1872 nbs->work[thread].sort_work_nalloc =
1873 over_alloc_large(ncz_max*grid->na_sc*SGSF);
1874 srenew(nbs->work[thread].sort_work,
1875 nbs->work[thread].sort_work_nalloc);
1876 }
1877 }
1878
1879 /* Now we know the dimensions we can fill the grid.
1880 * This is the first, unsorted fill. We sort the columns after this.
1881 */
1882 for(i=a0; i<a1; i++)
1883 {
1884 /* At this point nbs->cell contains the local grid x,y indices */
1885 cxy = nbs->cell[i];
1886 nbs->a[(grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc + grid->cxy_na[cxy]++] = i;
1887 }
1888
1889 /* Set the cell indices for the moved particles */
1890 n0 = grid->nc*grid->na_sc;
1891 n1 = grid->nc*grid->na_sc+grid->cxy_na[grid->ncx*grid->ncy];
1892 for(i=n0; i<n1; i++)
1893 {
1894 nbs->cell[nbs->a[i]] = i;
1895 }
1896
1897 /* Sort the super-cell columns along z into the sub-cells. */
1898 #pragma omp parallel for num_threads(nbs->nthread_max) schedule(static)
1899 for(thread=0; thread<nbs->nthread_max; thread++)
1900 {
1901 if (grid->bSimple)
1902 {
1903 sort_columns_simple(nbs,dd_zone,grid,a0,a1,atinfo,x,nbat,
1904 ((thread+0)*grid->ncx*grid->ncy)/nthread,
1905 ((thread+1)*grid->ncx*grid->ncy)/nthread,
1906 nbs->work[thread].sort_work);
1907 }
1908 else
1909 {
1910 sort_columns_supersub(nbs,dd_zone,grid,a0,a1,atinfo,x,nbat,
1911 ((thread+0)*grid->ncx*grid->ncy)/nthread,
1912 ((thread+1)*grid->ncx*grid->ncy)/nthread,
1913 nbs->work[thread].sort_work);
1914 }
1915 }
1916
1917 #ifdef NBNXN_SEARCH_SSE
1918 if (grid->bSimple && nbat->XFormat == nbatX8)
1919 {
1920 combine_bounding_box_pairs(grid,grid->bb);
1921 }
1922 #endif
1923
1924 if (!grid->bSimple)
1925 {
1926 grid->nsubc_tot = 0;
1927 for(i=0; i<grid->nc; i++)
1928 {
1929 grid->nsubc_tot += grid->nsubc[i];
1930 }
1931 }
1932
1933 if (debug)
1934 {
1935 if (grid->bSimple)
1936 {
1937 print_bbsizes_simple(debug,nbs,grid);
1938 }
1939 else
1940 {
1941 fprintf(debug,"ns non-zero sub-cells: %d average atoms %.2f\n",
1942 grid->nsubc_tot,(a1-a0)/(double)grid->nsubc_tot);
1943
1944 print_bbsizes_supersub(debug,nbs,grid);
1945 }
1946 }
1947 }
1948
1949 /* Reallocation wrapper function for nbnxn data structures */
1950 static void nb_realloc_void(void **ptr,
1951 int nbytes_copy,int nbytes_new,
1952 gmx_nbat_alloc_t *ma,
1953 gmx_nbat_free_t *mf)
1954 {
1955 void *ptr_new;
1956
1957 ma(&ptr_new,nbytes_new);
1958
1959 if (nbytes_new > 0 && ptr_new == NULL)
1960 {
1961 gmx_fatal(FARGS, "Allocation of %d bytes failed", nbytes_new);
1962 }
1963
1964 if (nbytes_copy > 0)
1965 {
1966 if (nbytes_new < nbytes_copy)
1967 {
1968 gmx_incons("In nb_realloc_void: new size less than copy size");
1969 }
1970 memcpy(ptr_new,*ptr,nbytes_copy);
1971 }
1972 if (*ptr != NULL)
1973 {
1974 mf(*ptr);
1975 }
1976 *ptr = ptr_new;
1977 }
1978
1979 /* NOTE: does not preserve the contents! */
1980 static void nb_realloc_int(int **ptr,int n,
1981 gmx_nbat_alloc_t *ma,
1982 gmx_nbat_free_t *mf)
1983 {
1984 if (*ptr != NULL)
1985 {
1986 mf(*ptr);
1987 }
1988 ma((void **)ptr,n*sizeof(**ptr));
1989 }
1990
1991 /* NOTE: does not preserve the contents! */
1992 static void nb_realloc_real(real **ptr,int n,
1993 gmx_nbat_alloc_t *ma,
1994 gmx_nbat_free_t *mf)
1995 {
1996 if (*ptr != NULL)
1997 {
1998 mf(*ptr);
1999 }
2000 ma((void **)ptr,n*sizeof(**ptr));
2001 }
2002
2003 /* Reallocate the nbnxn_atomdata_t for a size of n atoms */
2004 static void nbnxn_atomdata_realloc(nbnxn_atomdata_t *nbat,int n)
2005 {
2006 int t;
2007
2008 nb_realloc_void((void **)&nbat->type,
2009 nbat->natoms*sizeof(*nbat->type),
2010 n*sizeof(*nbat->type),
2011 nbat->alloc,nbat->free);
2012 nb_realloc_void((void **)&nbat->lj_comb,
2013 nbat->natoms*2*sizeof(*nbat->lj_comb),
2014 n*2*sizeof(*nbat->lj_comb),
2015 nbat->alloc,nbat->free);
2016 if (nbat->XFormat != nbatXYZQ)
2017 {
2018 nb_realloc_void((void **)&nbat->q,
2019 nbat->natoms*sizeof(*nbat->q),
2020 n*sizeof(*nbat->q),
2021 nbat->alloc,nbat->free);
2022 }
2023 if (nbat->nenergrp > 1)
2024 {
2025 nb_realloc_void((void **)&nbat->energrp,
2026 nbat->natoms/nbat->na_c*sizeof(*nbat->energrp),
2027 n/nbat->na_c*sizeof(*nbat->energrp),
2028 nbat->alloc,nbat->free);
2029 }
2030 nb_realloc_void((void **)&nbat->x,
2031 nbat->natoms*nbat->xstride*sizeof(*nbat->x),
2032 n*nbat->xstride*sizeof(*nbat->x),
2033 nbat->alloc,nbat->free);
2034 for(t=0; t<nbat->nout; t++)
2035 {
2036 /* Allocate one element extra for possible signaling with CUDA */
2037 nb_realloc_void((void **)&nbat->out[t].f,
2038 nbat->natoms*nbat->fstride*sizeof(*nbat->out[t].f),
2039 n*nbat->fstride*sizeof(*nbat->out[t].f),
2040 nbat->alloc,nbat->free);
2041 }
2042 nbat->nalloc = n;
2043 }
2044
2045 /* Sets up a grid and puts the atoms on the grid.
2046 * This function only operates on one domain of the domain decompostion.
2047 * Note that without domain decomposition there is only one domain.
2048 */
2049 void nbnxn_put_on_grid(nbnxn_search_t nbs,
2050 int ePBC,matrix box,
2051 int dd_zone,
2052 rvec corner0,rvec corner1,
2053 int a0,int a1,
2054 real atom_density,
2055 const int *atinfo,
2056 rvec *x,
2057 int nmoved,int *move,
2058 int nb_kernel_type,
2059 nbnxn_atomdata_t *nbat)
2060 {
2061 nbnxn_grid_t *grid;
2062 int n;
2063 int nc_max_grid,nc_max;
2064
2065 grid = &nbs->grid[dd_zone];
2066
2067 nbs_cycle_start(&nbs->cc[enbsCCgrid]);
2068
2069 grid->bSimple = nbnxn_kernel_pairlist_simple(nb_kernel_type);
2070
2071 grid->na_c = kernel_to_ci_size(nb_kernel_type);
2072 grid->na_cj = kernel_to_cj_size(nb_kernel_type);
2073 grid->na_sc = (grid->bSimple ? 1 : GPU_NSUBCELL)*grid->na_c;
2074 grid->na_c_2log = get_2log(grid->na_c);
2075
2076 nbat->na_c = grid->na_c;
2077
2078 if (dd_zone == 0)
2079 {
2080 grid->cell0 = 0;
2081 }
2082 else
2083 {
2084 grid->cell0 =
2085 (nbs->grid[dd_zone-1].cell0 + nbs->grid[dd_zone-1].nc)*
2086 nbs->grid[dd_zone-1].na_sc/grid->na_sc;
2087 }
2088
2089 n = a1 - a0;
2090
2091 if (dd_zone == 0)
2092 {
2093 nbs->ePBC = ePBC;
2094 copy_mat(box,nbs->box);
2095
2096 if (atom_density >= 0)
2097 {
2098 grid->atom_density = atom_density;
2099 }
2100 else
2101 {
2102 grid->atom_density = grid_atom_density(n-nmoved,corner0,corner1);
2103 }
2104
2105 grid->cell0 = 0;
2106
2107 nbs->natoms_local = a1 - nmoved;
2108 /* We assume that nbnxn_put_on_grid is called first
2109 * for the local atoms (dd_zone=0).
2110 */
2111 nbs->natoms_nonlocal = a1 - nmoved;
2112 }
2113 else
2114 {
2115 nbs->natoms_nonlocal = max(nbs->natoms_nonlocal,a1);
2116 }
2117
2118 nc_max_grid = set_grid_size_xy(nbs,grid,n-nmoved,corner0,corner1,
2119 nbs->grid[0].atom_density,
2120 nbat->XFormat);
2121
2122 nc_max = grid->cell0 + nc_max_grid;
2123
2124 if (a1 > nbs->cell_nalloc)
2125 {
2126 nbs->cell_nalloc = over_alloc_large(a1);
2127 srenew(nbs->cell,nbs->cell_nalloc);
2128 }
2129
2130 /* To avoid conditionals we store the moved particles at the end of a,
2131 * make sure we have enough space.
2132 */
2133 if (nc_max*grid->na_sc + nmoved > nbs->a_nalloc)
2134 {
2135 nbs->a_nalloc = over_alloc_large(nc_max*grid->na_sc + nmoved);
2136 srenew(nbs->a,nbs->a_nalloc);
2137 }
2138
2139 if (nc_max*grid->na_sc > nbat->nalloc)
2140 {
2141 nbnxn_atomdata_realloc(nbat,nc_max*grid->na_sc);
2142 }
2143
2144 calc_cell_indices(nbs,dd_zone,grid,a0,a1,atinfo,x,move,nbat);
2145
2146 if (dd_zone == 0)
2147 {
2148 nbat->natoms_local = nbat->natoms;
2149 }
2150
2151 nbs_cycle_stop(&nbs->cc[enbsCCgrid]);
2152 }
2153
2154 /* Calls nbnxn_put_on_grid for all non-local domains */
2155 void nbnxn_put_on_grid_nonlocal(nbnxn_search_t nbs,
2156 const gmx_domdec_zones_t *zones,
2157 const int *atinfo,
2158 rvec *x,
2159 int nb_kernel_type,
2160 nbnxn_atomdata_t *nbat)
2161 {
2162 int zone,d;
2163 rvec c0,c1;
2164
2165 for(zone=1; zone<zones->n; zone++)
2166 {
2167 for(d=0; d<DIM; d++)
2168 {
2169 c0[d] = zones->size[zone].bb_x0[d];
2170 c1[d] = zones->size[zone].bb_x1[d];
2171 }
2172
2173 nbnxn_put_on_grid(nbs,nbs->ePBC,NULL,
2174 zone,c0,c1,
2175 zones->cg_range[zone],
2176 zones->cg_range[zone+1],
2177 -1,
2178 atinfo,
2179 x,
2180 0,NULL,
2181 nb_kernel_type,
2182 nbat);
2183 }
2184 }
2185
2186 /* Add simple grid type information to the local super/sub grid */
2187 void nbnxn_grid_add_simple(nbnxn_search_t nbs,
2188 nbnxn_atomdata_t *nbat)
2189 {
2190 nbnxn_grid_t *grid;
2191 float *bbcz,*bb;
2192 int ncd,sc;
2193
2194 grid = &nbs->grid[0];
2195
2196 if (grid->bSimple)
2197 {
2198 gmx_incons("nbnxn_grid_simple called with a simple grid");
2199 }
2200
2201 ncd = grid->na_sc/NBNXN_CPU_CLUSTER_I_SIZE;
2202
2203 if (grid->nc*ncd > grid->nc_nalloc_simple)
2204 {
2205 grid->nc_nalloc_simple = over_alloc_large(grid->nc*ncd);
2206 srenew(grid->bbcz_simple,grid->nc_nalloc_simple*NNBSBB_D);
2207 srenew(grid->bb_simple,grid->nc_nalloc_simple*NNBSBB_B);
2208 srenew(grid->flags_simple,grid->nc_nalloc_simple);
2209 if (nbat->XFormat)
2210 {
2211 sfree_aligned(grid->bbj);
2212 snew_aligned(grid->bbj,grid->nc_nalloc_simple/2,16);
2213 }
2214 }
2215
2216 bbcz = grid->bbcz_simple;
2217 bb = grid->bb_simple;
2218
2219 #pragma omp parallel for num_threads(gmx_omp_nthreads_get(emntPairsearch)) schedule(static)
2220 for(sc=0; sc<grid->nc; sc++)
2221 {
2222 int c,tx,na;
2223
2224 for(c=0; c<ncd; c++)
2225 {
2226 tx = sc*ncd + c;
2227
2228 na = NBNXN_CPU_CLUSTER_I_SIZE;
2229 while (na > 0 &&
2230 nbat->type[tx*NBNXN_CPU_CLUSTER_I_SIZE+na-1] == nbat->ntype-1)
2231 {
2232 na--;
2233 }
2234
2235 if (na > 0)
2236 {
2237 switch (nbat->XFormat)
2238 {
2239 case nbatX4:
2240 /* PACK_X4==NBNXN_CPU_CLUSTER_I_SIZE, so this is simple */
2241 calc_bounding_box_x_x4(na,nbat->x+tx*STRIDE_P4,
2242 bb+tx*NNBSBB_B);
2243 break;
2244 case nbatX8:
2245 /* PACK_X8>NBNXN_CPU_CLUSTER_I_SIZE, more complicated */
2246 calc_bounding_box_x_x8(na,nbat->x+X8_IND_A(tx*NBNXN_CPU_CLUSTER_I_SIZE),
2247 bb+tx*NNBSBB_B);
2248 break;
2249 default:
2250 calc_bounding_box(na,nbat->xstride,
2251 nbat->x+tx*NBNXN_CPU_CLUSTER_I_SIZE*nbat->xstride,
2252 bb+tx*NNBSBB_B);
2253 break;
2254 }
2255 bbcz[tx*NNBSBB_D+0] = bb[tx*NNBSBB_B +ZZ];
2256 bbcz[tx*NNBSBB_D+1] = bb[tx*NNBSBB_B+NNBSBB_C+ZZ];
2257
2258 /* No interaction optimization yet here */
2259 grid->flags_simple[tx] = NBNXN_CI_DO_LJ(0) | NBNXN_CI_DO_COUL(0);
2260 }
2261 else
2262 {
2263 grid->flags_simple[tx] = 0;
2264 }
2265 }
2266 }
2267
2268 #ifdef NBNXN_SEARCH_SSE
2269 if (grid->bSimple && nbat->XFormat == nbatX8)
2270 {
2271 combine_bounding_box_pairs(grid,grid->bb_simple);
2272 }
2273 #endif
2274 }
2275
2276 void nbnxn_get_ncells(nbnxn_search_t nbs,int *ncx,int *ncy)
2277 {
2278 *ncx = nbs->grid[0].ncx;
2279 *ncy = nbs->grid[0].ncy;
2280 }
2281
2282 void nbnxn_get_atomorder(nbnxn_search_t nbs,int **a,int *n)
2283 {
2284 const nbnxn_grid_t *grid;
2285
2286 grid = &nbs->grid[0];
2287
2288 /* Return the atom order for the home cell (index 0) */
2289 *a = nbs->a;
2290
2291 *n = grid->cxy_ind[grid->ncx*grid->ncy]*grid->na_sc;
2292 }
2293
2294 void nbnxn_set_atomorder(nbnxn_search_t nbs)
2295 {
2296 nbnxn_grid_t *grid;
2297 int ao,cx,cy,cxy,cz,j;
2298
2299 /* Set the atom order for the home cell (index 0) */
2300 grid = &nbs->grid[0];
2301
2302 ao = 0;
2303 for(cx=0; cx<grid->ncx; cx++)
2304 {
2305 for(cy=0; cy<grid->ncy; cy++)
2306 {
2307 cxy = cx*grid->ncy + cy;
2308 j = grid->cxy_ind[cxy]*grid->na_sc;
2309 for(cz=0; cz<grid->cxy_na[cxy]; cz++)
2310 {
2311 nbs->a[j] = ao;
2312 nbs->cell[ao] = j;
2313 ao++;
2314 j++;
2315 }
2316 }
2317 }
2318 }
2319
2320 /* Determines the cell range along one dimension that
2321 * the bounding box b0 - b1 sees.
2322 */
2323 static void get_cell_range(real b0,real b1,
2324 int nc,real c0,real s,real invs,
2325 real d2,real r2,int *cf,int *cl)
2326 {
2327 *cf = max((int)((b0 - c0)*invs),0);
2328
2329 while (*cf > 0 && d2 + sqr((b0 - c0) - (*cf-1+1)*s) < r2)
2330 {
2331 (*cf)--;
2332 }
2333
2334 *cl = min((int)((b1 - c0)*invs),nc-1);
2335 while (*cl < nc-1 && d2 + sqr((*cl+1)*s - (b1 - c0)) < r2)
2336 {
2337 (*cl)++;
2338 }
2339 }
2340
2341 /* Reference code calculating the distance^2 between two bounding boxes */
2342 static float box_dist2(float bx0,float bx1,float by0,
2343 float by1,float bz0,float bz1,
2344 const float *bb)
2345 {
2346 float d2;
2347 float dl,dh,dm,dm0;
2348
2349 d2 = 0;
2350
2351 dl = bx0 - bb[BBU_X];
2352 dh = bb[BBL_X] - bx1;
2353 dm = max(dl,dh);
2354 dm0 = max(dm,0);
2355 d2 += dm0*dm0;
2356
2357 dl = by0 - bb[BBU_Y];
2358 dh = bb[BBL_Y] - by1;
2359 dm = max(dl,dh);
2360 dm0 = max(dm,0);
2361 d2 += dm0*dm0;
2362
2363 dl = bz0 - bb[BBU_Z];
2364 dh = bb[BBL_Z] - bz1;
2365 dm = max(dl,dh);
2366 dm0 = max(dm,0);
2367 d2 += dm0*dm0;
2368
2369 return d2;
2370 }
2371
2372 /* Plain C code calculating the distance^2 between two bounding boxes */
2373 static float subc_bb_dist2(int si,const float *bb_i_ci,
2374 int csj,const float *bb_j_all)
2375 {
2376 const float *bb_i,*bb_j;
2377 float d2;
2378 float dl,dh,dm,dm0;
2379
2380 bb_i = bb_i_ci + si*NNBSBB_B;
2381 bb_j = bb_j_all + csj*NNBSBB_B;
2382
2383 d2 = 0;
2384
2385 dl = bb_i[BBL_X] - bb_j[BBU_X];
2386 dh = bb_j[BBL_X] - bb_i[BBU_X];
2387 dm = max(dl,dh);
2388 dm0 = max(dm,0);
2389 d2 += dm0*dm0;
2390
2391 dl = bb_i[BBL_Y] - bb_j[BBU_Y];
2392 dh = bb_j[BBL_Y] - bb_i[BBU_Y];
2393 dm = max(dl,dh);
2394 dm0 = max(dm,0);
2395 d2 += dm0*dm0;
2396
2397 dl = bb_i[BBL_Z] - bb_j[BBU_Z];
2398 dh = bb_j[BBL_Z] - bb_i[BBU_Z];
2399 dm = max(dl,dh);
2400 dm0 = max(dm,0);
2401 d2 += dm0*dm0;
2402
2403 return d2;
2404 }
2405
2406 #ifdef NBNXN_SEARCH_SSE
2407
2408 /* SSE code for bb distance for bb format xyz0 */
2409 static float subc_bb_dist2_sse(int na_c,
2410 int si,const float *bb_i_ci,
2411 int csj,const float *bb_j_all)
2412 {
2413 const float *bb_i,*bb_j;
2414
2415 __m128 bb_i_SSE0,bb_i_SSE1;
2416 __m128 bb_j_SSE0,bb_j_SSE1;
2417 __m128 dl_SSE;
2418 __m128 dh_SSE;
2419 __m128 dm_SSE;
2420 __m128 dm0_SSE;
2421 __m128 d2_SSE;
2422 #ifndef GMX_X86_SSE4_1
2423 float d2_array[7],*d2_align;
2424
2425 d2_align = (float *)(((size_t)(d2_array+3)) & (~((size_t)15)));
2426 #else
2427 float d2;
2428 #endif
2429
2430 bb_i = bb_i_ci + si*NNBSBB_B;
2431 bb_j = bb_j_all + csj*NNBSBB_B;
2432
2433 bb_i_SSE0 = _mm_load_ps(bb_i);
2434 bb_i_SSE1 = _mm_load_ps(bb_i+NNBSBB_C);
2435 bb_j_SSE0 = _mm_load_ps(bb_j);
2436 bb_j_SSE1 = _mm_load_ps(bb_j+NNBSBB_C);
2437
2438 dl_SSE = _mm_sub_ps(bb_i_SSE0,bb_j_SSE1);
2439 dh_SSE = _mm_sub_ps(bb_j_SSE0,bb_i_SSE1);
2440
2441 dm_SSE = _mm_max_ps(dl_SSE,dh_SSE);
2442 dm0_SSE = _mm_max_ps(dm_SSE,_mm_setzero_ps());
2443 #ifndef GMX_X86_SSE4_1
2444 d2_SSE = _mm_mul_ps(dm0_SSE,dm0_SSE);
2445
2446 _mm_store_ps(d2_align,d2_SSE);
2447
2448 return d2_align[0] + d2_align[1] + d2_align[2];
2449 #else
2450 /* SSE4.1 dot product of components 0,1,2 */
2451 d2_SSE = _mm_dp_ps(dm0_SSE,dm0_SSE,0x71);
2452
2453 _mm_store_ss(&d2,d2_SSE);
2454
2455 return d2;
2456 #endif
2457 }
2458
2459 /* Calculate bb bounding distances of bb_i[si,...,si+3] and store them in d2 */
2460 #define SUBC_BB_DIST2_SSE_XXXX_INNER(si,bb_i,d2) \
2461 { \
2462 int shi; \
2463 \
2464 __m128 dx_0,dy_0,dz_0; \
2465 __m128 dx_1,dy_1,dz_1; \
2466 \
2467 __m128 mx,my,mz; \
2468 __m128 m0x,m0y,m0z; \
2469 \
2470 __m128 d2x,d2y,d2z; \
2471 __m128 d2s,d2t; \
2472 \
2473 shi = si*NNBSBB_D*DIM; \
2474 \
2475 xi_l = _mm_load_ps(bb_i+shi+0*STRIDE_8BB); \
2476 yi_l = _mm_load_ps(bb_i+shi+1*STRIDE_8BB); \
2477 zi_l = _mm_load_ps(bb_i+shi+2*STRIDE_8BB); \
2478 xi_h = _mm_load_ps(bb_i+shi+3*STRIDE_8BB); \
2479 yi_h = _mm_load_ps(bb_i+shi+4*STRIDE_8BB); \
2480 zi_h = _mm_load_ps(bb_i+shi+5*STRIDE_8BB); \
2481 \
2482 dx_0 = _mm_sub_ps(xi_l,xj_h); \
2483 dy_0 = _mm_sub_ps(yi_l,yj_h); \
2484 dz_0 = _mm_sub_ps(zi_l,zj_h); \
2485 \
2486 dx_1 = _mm_sub_ps(xj_l,xi_h); \
2487 dy_1 = _mm_sub_ps(yj_l,yi_h); \
2488 dz_1 = _mm_sub_ps(zj_l,zi_h); \
2489 \
2490 mx = _mm_max_ps(dx_0,dx_1); \
2491 my = _mm_max_ps(dy_0,dy_1); \
2492 mz = _mm_max_ps(dz_0,dz_1); \
2493 \
2494 m0x = _mm_max_ps(mx,zero); \
2495 m0y = _mm_max_ps(my,zero); \
2496 m0z = _mm_max_ps(mz,zero); \
2497 \
2498 d2x = _mm_mul_ps(m0x,m0x); \
2499 d2y = _mm_mul_ps(m0y,m0y); \
2500 d2z = _mm_mul_ps(m0z,m0z); \
2501 \
2502 d2s = _mm_add_ps(d2x,d2y); \
2503 d2t = _mm_add_ps(d2s,d2z); \
2504 \
2505 _mm_store_ps(d2+si,d2t); \
2506 }
2507
2508 /* SSE code for nsi bb distances for bb format xxxxyyyyzzzz */
2509 static void subc_bb_dist2_sse_xxxx(const float *bb_j,
2510 int nsi,const float *bb_i,
2511 float *d2)
2512 {
2513 __m128 xj_l,yj_l,zj_l;
2514 __m128 xj_h,yj_h,zj_h;
2515 __m128 xi_l,yi_l,zi_l;
2516 __m128 xi_h,yi_h,zi_h;
2517
2518 __m128 zero;
2519
2520 zero = _mm_setzero_ps();
2521
2522 xj_l = _mm_set1_ps(bb_j[0*STRIDE_8BB]);
2523 yj_l = _mm_set1_ps(bb_j[1*STRIDE_8BB]);
2524 zj_l = _mm_set1_ps(bb_j[2*STRIDE_8BB]);
2525 xj_h = _mm_set1_ps(bb_j[3*STRIDE_8BB]);
2526 yj_h = _mm_set1_ps(bb_j[4*STRIDE_8BB]);
2527 zj_h = _mm_set1_ps(bb_j[5*STRIDE_8BB]);
2528
2529 /* Here we "loop" over si (0,STRIDE_8BB) from 0 to nsi with step STRIDE_8BB.
2530 * But as we know the number of iterations is 1 or 2, we unroll manually.
2531 */
2532 SUBC_BB_DIST2_SSE_XXXX_INNER(0,bb_i,d2);
2533 if (STRIDE_8BB < nsi)
2534 {
2535 SUBC_BB_DIST2_SSE_XXXX_INNER(STRIDE_8BB,bb_i,d2);
2536 }
2537 }
2538
2539 #endif /* NBNXN_SEARCH_SSE */
2540
2541 /* Plain C function which determines if any atom pair between two cells
2542 * is within distance sqrt(rl2).
2543 */
2544 static gmx_bool subc_in_range_x(int na_c,
2545 int si,const real *x_i,
2546 int csj,int stride,const real *x_j,
2547 real rl2)
2548 {
2549 int i,j,i0,j0;
2550 real d2;
2551
2552 for(i=0; i<na_c; i++)
2553 {
2554 i0 = (si*na_c + i)*DIM;
2555 for(j=0; j<na_c; j++)
2556 {
2557 j0 = (csj*na_c + j)*stride;
2558
2559 d2 = sqr(x_i[i0 ] - x_j[j0 ]) +
2560 sqr(x_i[i0+1] - x_j[j0+1]) +
2561 sqr(x_i[i0+2] - x_j[j0+2]);
2562
2563 if (d2 < rl2)
2564 {
2565 return TRUE;
2566 }
2567 }
2568 }
2569
2570 return FALSE;
2571 }
2572
2573 /* SSE function which determines if any atom pair between two cells,
2574 * both with 8 atoms, is within distance sqrt(rl2).
2575 */
2576 static gmx_bool subc_in_range_sse8(int na_c,
2577 int si,const real *x_i,
2578 int csj,int stride,const real *x_j,
2579 real rl2)
2580 {
2581 #ifdef NBNXN_SEARCH_SSE_SINGLE
2582 __m128 ix_SSE0,iy_SSE0,iz_SSE0;
2583 __m128 ix_SSE1,iy_SSE1,iz_SSE1;
2584
2585 __m128 rc2_SSE;
2586
2587 int na_c_sse;
2588 int j0,j1;
2589
2590 rc2_SSE = _mm_set1_ps(rl2);
2591
2592 na_c_sse = NBNXN_GPU_CLUSTER_SIZE/STRIDE_8BB;
2593 ix_SSE0 = _mm_load_ps(x_i+(si*na_c_sse*DIM+0)*STRIDE_8BB);
2594 iy_SSE0 = _mm_load_ps(x_i+(si*na_c_sse*DIM+1)*STRIDE_8BB);
2595 iz_SSE0 = _mm_load_ps(x_i+(si*na_c_sse*DIM+2)*STRIDE_8BB);
2596 ix_SSE1 = _mm_load_ps(x_i+(si*na_c_sse*DIM+3)*STRIDE_8BB);
2597 iy_SSE1 = _mm_load_ps(x_i+(si*na_c_sse*DIM+4)*STRIDE_8BB);
2598 iz_SSE1 = _mm_load_ps(x_i+(si*na_c_sse*DIM+5)*STRIDE_8BB);
2599
2600 /* We loop from the outer to the inner particles to maximize
2601 * the chance that we find a pair in range quickly and return.
2602 */
2603 j0 = csj*na_c;
2604 j1 = j0 + na_c - 1;
2605 while (j0 < j1)
2606 {
2607 __m128 jx0_SSE,jy0_SSE,jz0_SSE;
2608 __m128 jx1_SSE,jy1_SSE,jz1_SSE;
2609
2610 __m128 dx_SSE0,dy_SSE0,dz_SSE0;
2611 __m128 dx_SSE1,dy_SSE1,dz_SSE1;
2612 __m128 dx_SSE2,dy_SSE2,dz_SSE2;
2613 __m128 dx_SSE3,dy_SSE3,dz_SSE3;
2614
2615 __m128 rsq_SSE0;
2616 __m128 rsq_SSE1;
2617 __m128 rsq_SSE2;
2618 __m128 rsq_SSE3;
2619
2620 __m128 wco_SSE0;
2621 __m128 wco_SSE1;
2622 __m128 wco_SSE2;
2623 __m128 wco_SSE3;
2624 __m128 wco_any_SSE01,wco_any_SSE23,wco_any_SSE;
2625
2626 jx0_SSE = _mm_load1_ps(x_j+j0*stride+0);
2627 jy0_SSE = _mm_load1_ps(x_j+j0*stride+1);
2628 jz0_SSE = _mm_load1_ps(x_j+j0*stride+2);
2629
2630 jx1_SSE = _mm_load1_ps(x_j+j1*stride+0);
2631 jy1_SSE = _mm_load1_ps(x_j+j1*stride+1);
2632 jz1_SSE = _mm_load1_ps(x_j+j1*stride+2);
2633
2634 /* Calculate distance */
2635 dx_SSE0 = _mm_sub_ps(ix_SSE0,jx0_SSE);
2636 dy_SSE0 = _mm_sub_ps(iy_SSE0,jy0_SSE);
2637 dz_SSE0 = _mm_sub_ps(iz_SSE0,jz0_SSE);
2638 dx_SSE1 = _mm_sub_ps(ix_SSE1,jx0_SSE);
2639 dy_SSE1 = _mm_sub_ps(iy_SSE1,jy0_SSE);
2640 dz_SSE1 = _mm_sub_ps(iz_SSE1,jz0_SSE);
2641 dx_SSE2 = _mm_sub_ps(ix_SSE0,jx1_SSE);
2642 dy_SSE2 = _mm_sub_ps(iy_SSE0,jy1_SSE);
2643 dz_SSE2 = _mm_sub_ps(iz_SSE0,jz1_SSE);
2644 dx_SSE3 = _mm_sub_ps(ix_SSE1,jx1_SSE);
2645 dy_SSE3 = _mm_sub_ps(iy_SSE1,jy1_SSE);
2646 dz_SSE3 = _mm_sub_ps(iz_SSE1,jz1_SSE);
2647
2648 /* rsq = dx*dx+dy*dy+dz*dz */
2649 rsq_SSE0 = gmx_mm_calc_rsq_ps(dx_SSE0,dy_SSE0,dz_SSE0);
2650 rsq_SSE1 = gmx_mm_calc_rsq_ps(dx_SSE1,dy_SSE1,dz_SSE1);
2651 rsq_SSE2 = gmx_mm_calc_rsq_ps(dx_SSE2,dy_SSE2,dz_SSE2);
2652 rsq_SSE3 = gmx_mm_calc_rsq_ps(dx_SSE3,dy_SSE3,dz_SSE3);
2653
2654 wco_SSE0 = _mm_cmplt_ps(rsq_SSE0,rc2_SSE);
2655 wco_SSE1 = _mm_cmplt_ps(rsq_SSE1,rc2_SSE);
2656 wco_SSE2 = _mm_cmplt_ps(rsq_SSE2,rc2_SSE);
2657 wco_SSE3 = _mm_cmplt_ps(rsq_SSE3,rc2_SSE);
2658
2659 wco_any_SSE01 = _mm_or_ps(wco_SSE0,wco_SSE1);
2660 wco_any_SSE23 = _mm_or_ps(wco_SSE2,wco_SSE3);
2661 wco_any_SSE = _mm_or_ps(wco_any_SSE01,wco_any_SSE23);
2662
2663 if (_mm_movemask_ps(wco_any_SSE))
2664 {
2665 return TRUE;
2666 }
2667
2668 j0++;
2669 j1--;
2670 }
2671 return FALSE;
2672
2673 #else
2674 /* No SSE */
2675 gmx_incons("SSE function called without SSE support");
2676
2677 return TRUE;
2678 #endif
2679 }
2680
2681 /* Returns the j sub-cell for index cj_ind */
2682 static int nbl_cj(const nbnxn_pairlist_t *nbl,int cj_ind)
2683 {
2684 return nbl->cj4[cj_ind>>2].cj[cj_ind & 3];
2685 }
2686
2687 /* Returns the i-interaction mask of the j sub-cell for index cj_ind */
2688 static unsigned nbl_imask0(const nbnxn_pairlist_t *nbl,int cj_ind)
2689 {
2690 return nbl->cj4[cj_ind>>2].imei[0].imask;
2691 }
2692
2693 /* Ensures there is enough space for extra extra exclusion masks */
2694 static void check_excl_space(nbnxn_pairlist_t *nbl,int extra)
2695 {
2696 if (nbl->nexcl+extra > nbl->excl_nalloc)
2697 {
2698 nbl->excl_nalloc = over_alloc_small(nbl->nexcl+extra);
2699 nb_realloc_void((void **)&nbl->excl,
2700 nbl->nexcl*sizeof(*nbl->excl),
2701 nbl->excl_nalloc*sizeof(*nbl->excl),
2702 nbl->alloc,nbl->free);
2703 }
2704 }
2705
2706 /* Ensures there is enough space for ncell extra j-cells in the list */
2707 static void check_subcell_list_space_simple(nbnxn_pairlist_t *nbl,
2708 int ncell)
2709 {
2710 int cj_max;
2711
2712 cj_max = nbl->ncj + ncell;
2713
2714 if (cj_max > nbl->cj_nalloc)
2715 {
2716 nbl->cj_nalloc = over_alloc_small(cj_max);
2717 nb_realloc_void((void **)&nbl->cj,
2718 nbl->ncj*sizeof(*nbl->cj),
2719 nbl->cj_nalloc*sizeof(*nbl->cj),
2720 nbl->alloc,nbl->free);
2721 }
2722 }
2723
2724 /* Ensures there is enough space for ncell extra j-subcells in the list */
2725 static void check_subcell_list_space_supersub(nbnxn_pairlist_t *nbl,
2726 int nsupercell)
2727 {
2728 int ncj4_max,j4,j,w,t;
2729
2730 #define NWARP 2
2731 #define WARP_SIZE 32
2732
2733 /* We can have maximally nsupercell*GPU_NSUBCELL sj lists */
2734 /* We can store 4 j-subcell - i-supercell pairs in one struct.
2735 * since we round down, we need one extra entry.
2736 */
2737 ncj4_max = ((nbl->work->cj_ind + nsupercell*GPU_NSUBCELL + 4-1) >> 2);
2738
2739 if (ncj4_max > nbl->cj4_nalloc)
2740 {
2741 nbl->cj4_nalloc = over_alloc_small(ncj4_max);
2742 nb_realloc_void((void **)&nbl->cj4,
2743 nbl->work->cj4_init*sizeof(*nbl->cj4),
2744 nbl->cj4_nalloc*sizeof(*nbl->cj4),
2745 nbl->alloc,nbl->free);
2746 }
2747
2748 if (ncj4_max > nbl->work->cj4_init)
2749 {
2750 for(j4=nbl->work->cj4_init; j4<ncj4_max; j4++)
2751 {
2752 /* No i-subcells and no excl's in the list initially */
2753 for(w=0; w<NWARP; w++)
2754 {
2755 nbl->cj4[j4].imei[w].imask = 0U;
2756 nbl->cj4[j4].imei[w].excl_ind = 0;
2757
2758 }
2759 }
2760 nbl->work->cj4_init = ncj4_max;
2761 }
2762 }
2763
2764 /* Default nbnxn allocation routine, allocates 32 byte aligned,
2765 * which works for plain C and aligned SSE and AVX loads/stores.
2766 */
2767 static void nbnxn_alloc_aligned(void **ptr,size_t nbytes)
2768 {
2769 *ptr = save_malloc_aligned("ptr",__FILE__,__LINE__,nbytes,1,32);
2770 }
2771
2772 /* Free function for memory allocated with nbnxn_alloc_aligned */
2773 static void nbnxn_free_aligned(void *ptr)
2774 {
2775 sfree_aligned(ptr);
2776 }
2777
2778 /* Set all excl masks for one GPU warp no exclusions */
2779 static void set_no_excls(nbnxn_excl_t *excl)
2780 {
2781 int t;
2782
2783 for(t=0; t<WARP_SIZE; t++)
2784 {
2785 /* Turn all interaction bits on */
2786 excl->pair[t] = NBNXN_INT_MASK_ALL;
2787 }
2788 }
2789
2790 /* Initializes a single nbnxn_pairlist_t data structure */
2791 static void nbnxn_init_pairlist(nbnxn_pairlist_t *nbl,
2792 gmx_bool bSimple,
2793 gmx_nbat_alloc_t *alloc,
2794 gmx_nbat_free_t *free)
2795 {
2796 if (alloc == NULL)
2797 {
2798 nbl->alloc = nbnxn_alloc_aligned;
2799 }
2800 else
2801 {
2802 nbl->alloc = alloc;
2803 }
2804 if (free == NULL)
2805 {
2806 nbl->free = nbnxn_free_aligned;
2807 }
2808 else
2809 {
2810 nbl->free = free;
2811 }
2812
2813 nbl->bSimple = bSimple;
2814 nbl->na_sc = 0;
2815 nbl->na_ci = 0;
2816 nbl->na_cj = 0;
2817 nbl->nci = 0;
2818 nbl->ci = NULL;
2819 nbl->ci_nalloc = 0;
2820 nbl->ncj = 0;
2821 nbl->cj = NULL;
2822 nbl->cj_nalloc = 0;
2823 nbl->ncj4 = 0;
2824 /* We need one element extra in sj, so alloc initially with 1 */
2825 nbl->cj4_nalloc = 0;
2826 nbl->cj4 = NULL;
2827 nbl->nci_tot = 0;
2828
2829 if (!nbl->bSimple)
2830 {
2831 nbl->excl = NULL;
2832 nbl->excl_nalloc = 0;
2833 nbl->nexcl = 0;
2834 check_excl_space(nbl,1);
2835 nbl->nexcl = 1;
2836 set_no_excls(&nbl->excl[0]);
2837 }
2838
2839 snew(nbl->work,1);
2840 #ifdef NBNXN_BBXXXX
2841 snew_aligned(nbl->work->bb_ci,GPU_NSUBCELL/STRIDE_8BB*NNBSBB_XXXX,16);
2842 #else
2843 snew_aligned(nbl->work->bb_ci,GPU_NSUBCELL*NNBSBB_B,16);
2844 #endif
2845 snew_aligned(nbl->work->x_ci,NBNXN_NA_SC_MAX*DIM,16);
2846 #ifdef NBNXN_SEARCH_SSE
2847 snew_aligned(nbl->work->x_ci_x86_simd128,1,16);
2848 #ifdef GMX_X86_AVX_256
2849 snew_aligned(nbl->work->x_ci_x86_simd256,1,32);
2850 #endif
2851 #endif
2852 snew_aligned(nbl->work->d2,GPU_NSUBCELL,16);
2853 }
2854
2855 void nbnxn_init_pairlist_set(nbnxn_pairlist_set_t *nbl_list,
2856 gmx_bool bSimple, gmx_bool bCombined,
2857 gmx_nbat_alloc_t *alloc,
2858 gmx_nbat_free_t *free)
2859 {
2860 int i;
2861
2862 nbl_list->bSimple = bSimple;
2863 nbl_list->bCombined = bCombined;
2864
2865 nbl_list->nnbl = gmx_omp_nthreads_get(emntNonbonded);
2866
2867 snew(nbl_list->nbl,nbl_list->nnbl);
2868 /* Execute in order to avoid memory interleaving between threads */
2869 #pragma omp parallel for num_threads(nbl_list->nnbl) schedule(static)
2870 for(i=0; i<nbl_list->nnbl; i++)
2871 {
2872 /* Allocate the nblist data structure locally on each thread
2873 * to optimize memory access for NUMA architectures.
2874 */
2875 snew(nbl_list->nbl[i],1);
2876
2877 /* Only list 0 is used on the GPU, use normal allocation for i>0 */
2878 if (i == 0)
2879 {
2880 nbnxn_init_pairlist(nbl_list->nbl[i],nbl_list->bSimple,alloc,free);
2881 }
2882 else
2883 {
2884 nbnxn_init_pairlist(nbl_list->nbl[i],nbl_list->bSimple,NULL,NULL);
2885 }
2886 }
2887 }
2888
2889 /* Print statistics of a pair list, used for debug output */
2890 static void print_nblist_statistics_simple(FILE *fp,const nbnxn_pairlist_t *nbl,
2891 const nbnxn_search_t nbs,real rl)
2892 {
2893 const nbnxn_grid_t *grid;
2894 int cs[SHIFTS];
2895 int s,i,j;
2896 int npexcl;
2897
2898 /* This code only produces correct statistics with domain decomposition */
2899 grid = &nbs->grid[0];
2900
2901 fprintf(fp,"nbl nci %d ncj %d\n",
2902 nbl->nci,nbl->ncj);
2903 fprintf(fp,"nbl na_sc %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
2904 nbl->na_sc,rl,nbl->ncj,nbl->ncj/(double)grid->nc,
2905 nbl->ncj/(double)grid->nc*grid->na_sc,
2906 nbl->ncj/(double)grid->nc*grid->na_sc/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid->nc*grid->na_sc/det(nbs->box)));
2907
2908 fprintf(fp,"nbl average j cell list length %.1f\n",
2909 0.25*nbl->ncj/(double)nbl->nci);
2910
2911 for(s=0; s<SHIFTS; s++)
2912 {
2913 cs[s] = 0;
2914 }
2915 npexcl = 0;
2916 for(i=0; i<nbl->nci; i++)
2917 {
2918 cs[nbl->ci[i].shift & NBNXN_CI_SHIFT] +=
2919 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start;
2920
2921 j = nbl->ci[i].cj_ind_start;
2922 while (j < nbl->ci[i].cj_ind_end &&
2923 nbl->cj[j].excl != NBNXN_INT_MASK_ALL)
2924 {
2925 npexcl++;
2926 j++;
2927 }
2928 }
2929 fprintf(fp,"nbl cell pairs, total: %d excl: %d %.1f%%\n",
2930 nbl->ncj,npexcl,100*npexcl/(double)nbl->ncj);
2931 for(s=0; s<SHIFTS; s++)
2932 {
2933 if (cs[s] > 0)
2934 {
2935 fprintf(fp,"nbl shift %2d ncj %3d\n",s,cs[s]);
2936 }
2937 }
2938 }
2939
2940 /* Print statistics of a pair lists, used for debug output */
2941 static void print_nblist_statistics_supersub(FILE *fp,const nbnxn_pairlist_t *nbl,
2942 const nbnxn_search_t nbs,real rl)
2943 {
2944 const nbnxn_grid_t *grid;
2945 int i,j4,j,si,b;
2946 int c[GPU_NSUBCELL+1];
2947
2948 /* This code only produces correct statistics with domain decomposition */
2949 grid = &nbs->grid[0];
2950
2951 fprintf(fp,"nbl nsci %d ncj4 %d nsi %d excl4 %d\n",
2952 nbl->nsci,nbl->ncj4,nbl->nci_tot,nbl->nexcl);
2953 fprintf(fp,"nbl na_c %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
2954 nbl->na_ci,rl,nbl->nci_tot,nbl->nci_tot/(double)grid->nsubc_tot,
2955 nbl->nci_tot/(double)grid->nsubc_tot*grid->na_c,
2956 nbl->nci_tot/(double)grid->nsubc_tot*grid->na_c/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid->nsubc_tot*grid->na_c/det(nbs->box)));
2957
2958 fprintf(fp,"nbl average j super cell list length %.1f\n",
2959 0.25*nbl->ncj4/(double)nbl->nsci);
2960 fprintf(fp,"nbl average i sub cell list length %.1f\n",
2961 nbl->nci_tot/(0.25*nbl->ncj4));
2962
2963 for(si=0; si<=GPU_NSUBCELL; si++)
2964 {
2965 c[si] = 0;
2966 }
2967 for(i=0; i<nbl->nsci; i++)
2968 {
2969 for(j4=nbl->sci[i].cj4_ind_start; j4<nbl->sci[i].cj4_ind_end; j4++)
2970 {
2971 for(j=0; j<4; j++)
2972 {
2973 b = 0;
2974 for(si=0; si<GPU_NSUBCELL; si++)
2975 {
2976 if (nbl->cj4[j4].imei[0].imask & (1U << (j*GPU_NSUBCELL + si)))
2977 {
2978 b++;
2979 }
2980 }
2981 c[b]++;
2982 }
2983 }
2984 }
2985 for(b=0; b<=GPU_NSUBCELL; b++)
2986 {
2987 fprintf(fp,"nbl j-list #i-subcell %d %7d %4.1f\n",
2988 b,c[b],100.0*c[b]/(double)(nbl->ncj4*NBNXN_GPU_JGROUP_SIZE));
2989 }
2990 }
2991
2992 /* Print the full pair list, used for debug output */
2993 static void print_supersub_nsp(const char *fn,
2994 const nbnxn_pairlist_t *nbl,
2995 int iloc)
2996 {
2997 char buf[STRLEN];
2998 FILE *fp;
2999 int i,nsp,j4,p;
3000
3001 sprintf(buf,"%s_%s.xvg",fn,NONLOCAL_I(iloc) ? "nl" : "l");
3002 fp = ffopen(buf,"w");
3003
3004 for(i=0; i<nbl->nci; i++)
3005 {
3006 nsp = 0;
3007 for(j4=nbl->sci[i].cj4_ind_start; j4<nbl->sci[i].cj4_ind_end; j4++)
3008 {
3009 for(p=0; p<NBNXN_GPU_JGROUP_SIZE*GPU_NSUBCELL; p++)
3010 {
3011 nsp += (nbl->cj4[j4].imei[0].imask >> p) & 1;
3012 }
3013 }
3014 fprintf(fp,"%4d %3d %3d\n",
3015 i,
3016 nsp,
3017 nbl->sci[i].cj4_ind_end-nbl->sci[i].cj4_ind_start);
3018 }
3019
3020 fclose(fp);
3021 }
3022
3023 /* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp */
3024 static void low_get_nbl_exclusions(nbnxn_pairlist_t *nbl,int cj4,
3025 int warp,nbnxn_excl_t **excl)
3026 {
3027 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
3028 {
3029 /* No exclusions set, make a new list entry */
3030 nbl->cj4[cj4].imei[warp].excl_ind = nbl->nexcl;
3031 nbl->nexcl++;
3032 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
3033 set_no_excls(*excl);
3034 }
3035 else
3036 {
3037 /* We already have some exclusions, new ones can be added to the list */
3038 *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
3039 }
3040 }
3041
3042 /* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp,
3043 * allocates extra memory, if necessary.
3044 */
3045 static void get_nbl_exclusions_1(nbnxn_pairlist_t *nbl,int cj4,
3046 int warp,nbnxn_excl_t **excl)
3047 {
3048 if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
3049 {
3050 /* We need to make a new list entry, check if we have space */
3051 check_excl_space(nbl,1);
3052 }
3053 low_get_nbl_exclusions(nbl,cj4,warp,excl);
3054 }
3055
3056 /* Returns pointers to the exclusion mask for cj4-unit cj4 for both warps,
3057 * allocates extra memory, if necessary.
3058 */
3059 static void get_nbl_exclusions_2(nbnxn_pairlist_t *nbl,int cj4,
3060 nbnxn_excl_t **excl_w0,
3061 nbnxn_excl_t **excl_w1)
3062 {
3063 /* Check for space we might need */
3064 check_excl_space(nbl,2);
3065
3066 low_get_nbl_exclusions(nbl,cj4,0,excl_w0);
3067 low_get_nbl_exclusions(nbl,cj4,1,excl_w1);
3068 }
3069
3070 /* Sets the self exclusions i=j and pair exclusions i>j */
3071 static void set_self_and_newton_excls_supersub(nbnxn_pairlist_t *nbl,
3072 int cj4_ind,int sj_offset,
3073 int si)
3074 {
3075 nbnxn_excl_t *excl[2];
3076 int ei,ej,w;
3077
3078 /* Here we only set the set self and double pair exclusions */
3079
3080 get_nbl_exclusions_2(nbl,cj4_ind,&excl[0],&excl[1]);
3081
3082 /* Only minor < major bits set */
3083 for(ej=0; ej<nbl->na_ci; ej++)
3084 {
3085 w = (ej>>2);
3086 for(ei=ej; ei<nbl->na_ci; ei++)
3087 {
3088 excl[w]->pair[(ej&(4-1))*nbl->na_ci+ei] &=
3089 ~(1U << (sj_offset*GPU_NSUBCELL+si));
3090 }
3091 }
3092 }
3093
3094 /* Returns a diagonal or off-diagonal interaction mask for plain C lists */
3095 static unsigned int get_imask(gmx_bool rdiag,int ci,int cj)
3096 {
3097 return (rdiag && ci == cj ? NBNXN_INT_MASK_DIAG : NBNXN_INT_MASK_ALL);
3098 }
3099
3100 #ifdef NBNXN_SEARCH_SSE
3101 /* Returns a diagonal or off-diagonal interaction mask for SIMD128 lists */
3102 static unsigned int get_imask_x86_simd128(gmx_bool rdiag,int ci,int cj)
3103 {
3104 #ifndef GMX_DOUBLE /* cj-size = 4 */
3105 return (rdiag && ci == cj ? NBNXN_INT_MASK_DIAG : NBNXN_INT_MASK_ALL);
3106 #else /* cj-size = 2 */
3107 return (rdiag && ci*2 == cj ? NBNXN_INT_MASK_DIAG_J2_0 :
3108 (rdiag && ci*2+1 == cj ? NBNXN_INT_MASK_DIAG_J2_1 :
3109 NBNXN_INT_MASK_ALL));
3110 #endif
3111 }
3112
3113 #ifdef GMX_X86_AVX_256
3114 /* Returns a diagonal or off-diagonal interaction mask for SIMD256 lists */
3115 static unsigned int get_imask_x86_simd256(gmx_bool rdiag,int ci,int cj)
3116 {
3117 #ifndef GMX_DOUBLE /* cj-size = 8 */
3118 return (rdiag && ci == cj*2 ? NBNXN_INT_MASK_DIAG_J8_0 :
3119 (rdiag && ci == cj*2+1 ? NBNXN_INT_MASK_DIAG_J8_1 :
3120 NBNXN_INT_MASK_ALL));
3121 #else /* cj-size = 2 */
3122 return (rdiag && ci == cj ? NBNXN_INT_MASK_DIAG : NBNXN_INT_MASK_ALL);
3123 #endif
3124 }
3125 #endif
3126 #endif /* NBNXN_SEARCH_SSE */
3127
3128 /* Plain C code for making a pair list of cell ci vs cell cjf-cjl.
3129 * Checks bounding box distances and possibly atom pair distances.
3130 */
3131 static void make_cluster_list_simple(const nbnxn_grid_t *gridj,
3132 nbnxn_pairlist_t *nbl,
3133 int ci,int cjf,int cjl,
3134 gmx_bool remove_sub_diag,
3135 const real *x_j,
3136 real rl2,float rbb2,
3137 int *ndistc)
3138 {
3139 const nbnxn_list_work_t *work;
3140
3141 const float *bb_ci;
3142 const real *x_ci;
3143
3144 gmx_bool InRange;
3145 real d2;
3146 int cjf_gl,cjl_gl,cj;
3147
3148 work = nbl->work;
3149
3150 bb_ci = nbl->work->bb_ci;
3151 x_ci = nbl->work->x_ci;
3152
3153 InRange = FALSE;
3154 while (!InRange && cjf <= cjl)
3155 {
3156 d2 = subc_bb_dist2(0,bb_ci,cjf,gridj->bb);
3157 *ndistc += 2;
3158
3159 /* Check if the distance is within the distance where
3160 * we use only the bounding box distance rbb,
3161 * or within the cut-off and there is at least one atom pair
3162 * within the cut-off.
3163 */
3164 if (d2 < rbb2)
3165 {
3166 InRange = TRUE;
3167 }
3168 else if (d2 < rl2)
3169 {
3170 int i,j;
3171
3172 cjf_gl = gridj->cell0 + cjf;
3173 for(i=0; i<NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
3174 {
3175 for(j=0; j<NBNXN_CPU_CLUSTER_I_SIZE; j++)
3176 {
3177 InRange = InRange ||
3178 (sqr(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
3179 sqr(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
3180 sqr(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rl2);
3181 }
3182 }
3183 *ndistc += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
3184 }
3185 if (!InRange)
3186 {
3187 cjf++;
3188 }
3189 }
3190 if (!InRange)
3191 {
3192 return;
3193 }
3194
3195 InRange = FALSE;
3196 while (!InRange && cjl > cjf)
3197 {
3198 d2 = subc_bb_dist2(0,bb_ci,cjl,gridj->bb);
3199 *ndistc += 2;
3200
3201 /* Check if the distance is within the distance where
3202 * we use only the bounding box distance rbb,
3203 * or within the cut-off and there is at least one atom pair
3204 * within the cut-off.
3205 */
3206 if (d2 < rbb2)
3207 {
3208 InRange = TRUE;
3209 }
3210 else if (d2 < rl2)
3211 {
3212 int i,j;
3213
3214 cjl_gl = gridj->cell0 + cjl;
3215 for(i=0; i<NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
3216 {
3217 for(j=0; j<NBNXN_CPU_CLUSTER_I_SIZE; j++)
3218 {
3219 InRange = InRange ||
3220 (sqr(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
3221 sqr(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
3222 sqr(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rl2);
3223 }
3224 }
3225 *ndistc += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
3226 }
3227 if (!InRange)
3228 {
3229 cjl--;
3230 }
3231 }
3232
3233 if (cjf <= cjl)
3234 {
3235 for(cj=cjf; cj<=cjl; cj++)
3236 {
3237 /* Store cj and the interaction mask */
3238 nbl->cj[nbl->ncj].cj = gridj->cell0 + cj;
3239 nbl->cj[nbl->ncj].excl = get_imask(remove_sub_diag,ci,cj);
3240 nbl->ncj++;
3241 }
3242 /* Increase the closing index in i super-cell list */
3243 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
3244 }
3245 }
3246
3247 #ifdef NBNXN_SEARCH_SSE
3248 /* Include make_cluster_list_x86_simd128/256 */
3249 #define GMX_MM128_HERE
3250 #include "gmx_x86_simd_macros.h"
3251 #define STRIDE_S PACK_X4
3252 #include "nbnxn_search_x86_simd.h"
3253 #undef STRIDE_S
3254 #undef GMX_MM128_HERE
3255 #ifdef GMX_X86_AVX_256
3256 /* Include make_cluster_list_x86_simd128/256 */
3257 #define GMX_MM256_HERE
3258 #include "gmx_x86_simd_macros.h"
3259 #define STRIDE_S GMX_X86_SIMD_WIDTH_HERE
3260 #include "nbnxn_search_x86_simd.h"
3261 #undef STRIDE_S
3262 #undef GMX_MM256_HERE
3263 #endif
3264 #endif
3265
3266 /* Plain C or SSE code for making a pair list of super-cell sci vs scj.
3267 * Checks bounding box distances and possibly atom pair distances.
3268 */
3269 static void make_cluster_list_supersub(const nbnxn_search_t nbs,
3270 const nbnxn_grid_t *gridi,
3271 const nbnxn_grid_t *gridj,
3272 nbnxn_pairlist_t *nbl,
3273 int sci,int scj,
3274 gmx_bool sci_equals_scj,
3275 int stride,const real *x,
3276 real rl2,float rbb2,
3277 int *ndistc)
3278 {
3279 int na_c;
3280 int npair;
3281 int cjo,ci1,ci,cj,cj_gl;
3282 int cj4_ind,cj_offset;
3283 unsigned imask;
3284 nbnxn_cj4_t *cj4;
3285 const float *bb_ci;
3286 const real *x_ci;
3287 float *d2l,d2;
3288 int w;
3289 #define PRUNE_LIST_CPU_ONE
3290 #ifdef PRUNE_LIST_CPU_ONE
3291 int ci_last=-1;
3292 #endif
3293
3294 d2l = nbl->work->d2;
3295
3296 bb_ci = nbl->work->bb_ci;
3297 x_ci = nbl->work->x_ci;
3298
3299 na_c = gridj->na_c;
3300
3301 for(cjo=0; cjo<gridj->nsubc[scj]; cjo++)
3302 {
3303 cj4_ind = (nbl->work->cj_ind >> 2);
3304 cj_offset = nbl->work->cj_ind - cj4_ind*NBNXN_GPU_JGROUP_SIZE;
3305 cj4 = &nbl->cj4[cj4_ind];
3306
3307 cj = scj*GPU_NSUBCELL + cjo;
3308
3309 cj_gl = gridj->cell0*GPU_NSUBCELL + cj;
3310
3311 /* Initialize this j-subcell i-subcell list */
3312 cj4->cj[cj_offset] = cj_gl;
3313 imask = 0;
3314
3315 if (sci_equals_scj)
3316 {
3317 ci1 = cjo + 1;
3318 }
3319 else
3320 {
3321 ci1 = gridi->nsubc[sci];
3322 }
3323
3324 #ifdef NBNXN_BBXXXX
3325 /* Determine all ci1 bb distances in one call with SSE */
3326 subc_bb_dist2_sse_xxxx(gridj->bb+(cj>>STRIDE_8BB_2LOG)*NNBSBB_XXXX+(cj & (STRIDE_8BB-1)),
3327 ci1,bb_ci,d2l);
3328 *ndistc += na_c*2;
3329 #endif
3330
3331 npair = 0;
3332 /* We use a fixed upper-bound instead of ci1 to help optimization */
3333 for(ci=0; ci<GPU_NSUBCELL; ci++)
3334 {
3335 if (ci == ci1)
3336 {
3337 break;
3338 }
3339
3340 #ifndef NBNXN_BBXXXX
3341 /* Determine the bb distance between ci and cj */
3342 d2l[ci] = subc_bb_dist2(ci,bb_ci,cj,gridj->bb);
3343 *ndistc += 2;
3344 #endif
3345 d2 = d2l[ci];
3346
3347 #ifdef PRUNE_LIST_CPU_ALL
3348 /* Check if the distance is within the distance where
3349 * we use only the bounding box distance rbb,
3350 * or within the cut-off and there is at least one atom pair
3351 * within the cut-off. This check is very costly.
3352 */
3353 *ndistc += na_c*na_c;
3354 if (d2 < rbb2 ||
3355 (d2 < rl2 && subc_in_range_x(na_c,ci,x_ci,cj_gl,stride,x,rl2)))
3356 #else
3357 /* Check if the distance between the two bounding boxes
3358 * in within the pair-list cut-off.
3359 */
3360 if (d2 < rl2)
3361 #endif
3362 {
3363 /* Flag this i-subcell to be taken into account */
3364 imask |= (1U << (cj_offset*GPU_NSUBCELL+ci));
3365
3366 #ifdef PRUNE_LIST_CPU_ONE
3367 ci_last = ci;
3368 #endif
3369
3370 npair++;
3371 }
3372 }
3373
3374 #ifdef PRUNE_LIST_CPU_ONE
3375 /* If we only found 1 pair, check if any atoms are actually
3376 * within the cut-off, so we could get rid of it.
3377 */
3378 if (npair == 1 && d2l[ci_last] >= rbb2)
3379 {
3380 /* Avoid using function pointers here, as it's slower */
3381 if (
3382 #ifdef NBNXN_8BB_SSE
3383 !subc_in_range_sse8
3384 #else
3385 !subc_in_range_x
3386 #endif
3387 (na_c,ci_last,x_ci,cj_gl,stride,x,rl2))
3388 {
3389 imask &= ~(1U << (cj_offset*GPU_NSUBCELL+ci_last));
3390 npair--;
3391 }
3392 }
3393 #endif
3394
3395 if (npair > 0)
3396 {
3397 /* We have a useful sj entry, close it now */
3398
3399 /* Set the exclucions for the ci== sj entry.
3400 * Here we don't bother to check if this entry is actually flagged,
3401 * as it will nearly always be in the list.
3402 */
3403 if (sci_equals_scj)
3404 {
3405 set_self_and_newton_excls_supersub(nbl,cj4_ind,cj_offset,cjo);
3406 }
3407
3408 /* Copy the cluster interaction mask to the list */
3409 for(w=0; w<NWARP; w++)
3410 {
3411 cj4->imei[w].imask |= imask;
3412 }
3413
3414 nbl->work->cj_ind++;
3415
3416 /* Keep the count */
3417 nbl->nci_tot += npair;
3418
3419 /* Increase the closing index in i super-cell list */
3420 nbl->sci[nbl->nsci].cj4_ind_end = ((nbl->work->cj_ind+4-1)>>2);
3421 }
3422 }
3423 }
3424
3425 /* Set all atom-pair exclusions from the topology stored in excl
3426 * as masks in the pair-list for simple list i-entry nbl_ci
3427 */
3428 static void set_ci_top_excls(const nbnxn_search_t nbs,
3429 nbnxn_pairlist_t *nbl,
3430 gmx_bool diagRemoved,
3431 int na_ci_2log,
3432 int na_cj_2log,
3433 const nbnxn_ci_t *nbl_ci,
3434 const t_blocka *excl)
3435 {
3436 const int *cell;
3437 int ci;
3438 int cj_ind_first,cj_ind_last;
3439 int cj_first,cj_last;
3440 int ndirect;
3441 int i,ai,aj,si,eind,ge,se;
3442 int found,cj_ind_0,cj_ind_1,cj_ind_m;
3443 int cj_m;
3444 gmx_bool Found_si;
3445 int si_ind;
3446 nbnxn_excl_t *nbl_excl;
3447 int inner_i,inner_e;
3448
3449 cell = nbs->cell;
3450
3451 if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
3452 {
3453 /* Empty list */
3454 return;
3455 }
3456
3457 ci = nbl_ci->ci;
3458
3459 cj_ind_first = nbl_ci->cj_ind_start;
3460 cj_ind_last = nbl->ncj - 1;
3461
3462 cj_first = nbl->cj[cj_ind_first].cj;
3463 cj_last = nbl->cj[cj_ind_last].cj;
3464
3465 /* Determine how many contiguous j-cells we have starting
3466 * from the first i-cell. This number can be used to directly
3467 * calculate j-cell indices for excluded atoms.
3468 */
3469 ndirect = 0;
3470 if (na_ci_2log == na_cj_2log)
3471 {
3472 while (cj_ind_first + ndirect <= cj_ind_last &&
3473 nbl->cj[cj_ind_first+ndirect].cj == ci + ndirect)
3474 {
3475 ndirect++;
3476 }
3477 }
3478 #ifdef NBNXN_SEARCH_SSE
3479 else
3480 {
3481 while (cj_ind_first + ndirect <= cj_ind_last &&
3482 nbl->cj[cj_ind_first+ndirect].cj == ci_to_cj(na_cj_2log,ci) + ndirect)
3483 {
3484 ndirect++;
3485 }
3486 }
3487 #endif
3488
3489 /* Loop over the atoms in the i super-cell */
3490 for(i=0; i<nbl->na_sc; i++)
3491 {
3492 ai = nbs->a[ci*nbl->na_sc+i];
3493 if (ai >= 0)
3494 {
3495 si = (i>>na_ci_2log);
3496
3497 /* Loop over the topology-based exclusions for this i-atom */
3498 for(eind=excl->index[ai]; eind<excl->index[ai+1]; eind++)
3499 {
3500 aj = excl->a[eind];
3501
3502 if (aj == ai)
3503 {
3504 /* The self exclusion are already set, save some time */
3505 continue;
3506 }
3507
3508 ge = cell[aj];
3509
3510 /* Without shifts we only calculate interactions j>i
3511 * for one-way pair-lists.
3512 */
3513 if (diagRemoved && ge <= ci*nbl->na_sc + i)
3514 {
3515 continue;
3516 }
3517
3518 se = (ge >> na_cj_2log);
3519
3520 /* Could the cluster se be in our list? */
3521 if (se >= cj_first && se <= cj_last)
3522 {
3523 if (se < cj_first + ndirect)
3524 {
3525 /* We can calculate cj_ind directly from se */
3526 found = cj_ind_first + se - cj_first;
3527 }
3528 else
3529 {
3530 /* Search for se using bisection */
3531 found = -1;
3532 cj_ind_0 = cj_ind_first + ndirect;
3533 cj_ind_1 = cj_ind_last + 1;
3534 while (found == -1 && cj_ind_0 < cj_ind_1)
3535 {
3536 cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
3537
3538 cj_m = nbl->cj[cj_ind_m].cj;
3539
3540 if (se == cj_m)
3541 {
3542 found = cj_ind_m;
3543 }
3544 else if (se < cj_m)
3545 {
3546 cj_ind_1 = cj_ind_m;
3547 }
3548 else
3549 {
3550 cj_ind_0 = cj_ind_m + 1;
3551 }
3552 }
3553 }
3554
3555 if (found >= 0)
3556 {
3557 inner_i = i - (si << na_ci_2log);
3558 inner_e = ge - (se << na_cj_2log);
3559
3560 nbl->cj[found].excl &= ~(1U<<((inner_i<<na_cj_2log) + inner_e));
3561 }
3562 }
3563 }
3564 }
3565 }
3566 }
3567
3568 /* Set all atom-pair exclusions from the topology stored in excl
3569 * as masks in the pair-list for i-super-cell entry nbl_sci
3570 */
3571 static void set_sci_top_excls(const nbnxn_search_t nbs,
3572 nbnxn_pairlist_t *nbl,
3573 gmx_bool diagRemoved,
3574 int na_c_2log,
3575 const nbnxn_sci_t *nbl_sci,
3576 const t_blocka *excl)
3577 {
3578 const int *cell;
3579 int na_c;
3580 int sci;
3581 int cj_ind_first,cj_ind_last;
3582 int cj_first,cj_last;
3583 int ndirect;
3584 int i,ai,aj,si,eind,ge,se;
3585 int found,cj_ind_0,cj_ind_1,cj_ind_m;
3586 int cj_m;
3587 gmx_bool Found_si;
3588 int si_ind;
3589 nbnxn_excl_t *nbl_excl;
3590 int inner_i,inner_e,w;
3591
3592 cell = nbs->cell;
3593
3594 na_c = nbl->na_ci;
3595
3596 if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
3597 {
3598 /* Empty list */
3599 return;
3600 }
3601
3602 sci = nbl_sci->sci;
3603
3604 cj_ind_first = nbl_sci->cj4_ind_start*NBNXN_GPU_JGROUP_SIZE;
3605 cj_ind_last = nbl->work->cj_ind - 1;
3606
3607 cj_first = nbl->cj4[nbl_sci->cj4_ind_start].cj[0];
3608 cj_last = nbl_cj(nbl,cj_ind_last);
3609
3610 /* Determine how many contiguous j-clusters we have starting
3611 * from the first i-cluster. This number can be used to directly
3612 * calculate j-cluster indices for excluded atoms.
3613 */
3614 ndirect = 0;
3615 while (cj_ind_first + ndirect <= cj_ind_last &&
3616 nbl_cj(nbl,cj_ind_first+ndirect) == sci*GPU_NSUBCELL + ndirect)
3617 {
3618 ndirect++;
3619 }
3620
3621 /* Loop over the atoms in the i super-cell */
3622 for(i=0; i<nbl->na_sc; i++)
3623 {
3624 ai = nbs->a[sci*nbl->na_sc+i];
3625 if (ai >= 0)
3626 {
3627 si = (i>>na_c_2log);
3628
3629 /* Loop over the topology-based exclusions for this i-atom */
3630 for(eind=excl->index[ai]; eind<excl->index[ai+1]; eind++)
3631 {
3632 aj = excl->a[eind];
3633
3634 if (aj == ai)
3635 {
3636 /* The self exclusion are already set, save some time */
3637 continue;
3638 }
3639
3640 ge = cell[aj];
3641
3642 /* Without shifts we only calculate interactions j>i
3643 * for one-way pair-lists.
3644 */
3645 if (diagRemoved && ge <= sci*nbl->na_sc + i)
3646 {
3647 continue;
3648 }
3649
3650 se = ge>>na_c_2log;
3651 /* Could the cluster se be in our list? */
3652 if (se >= cj_first && se <= cj_last)
3653 {
3654 if (se < cj_first + ndirect)
3655 {
3656 /* We can calculate cj_ind directly from se */
3657 found = cj_ind_first + se - cj_first;
3658 }
3659 else
3660 {
3661 /* Search for se using bisection */
3662 found = -1;
3663 cj_ind_0 = cj_ind_first + ndirect;
3664 cj_ind_1 = cj_ind_last + 1;
3665 while (found == -1 && cj_ind_0 < cj_ind_1)
3666 {
3667 cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
3668
3669 cj_m = nbl_cj(nbl,cj_ind_m);
3670
3671 if (se == cj_m)
3672 {
3673 found = cj_ind_m;
3674 }
3675 else if (se < cj_m)
3676 {
3677 cj_ind_1 = cj_ind_m;
3678 }
3679 else
3680 {
3681 cj_ind_0 = cj_ind_m + 1;
3682 }
3683 }
3684 }
3685
3686 if (found >= 0)
3687 {
3688 inner_i = i - si*na_c;
3689 inner_e = ge - se*na_c;
3690
3691 /* Macro for getting the index of atom a within a cluster */
3692 #define AMODI(a) ((a) & (NBNXN_CPU_CLUSTER_I_SIZE - 1))
3693 /* Macro for converting an atom number to a cluster number */
3694 #define A2CI(a) ((a) >> NBNXN_CPU_CLUSTER_I_SIZE_2LOG)
3695
3696 if (nbl_imask0(nbl,found) & (1U << (AMODI(found)*GPU_NSUBCELL + si)))
3697 {
3698 w = (inner_e >> 2);
3699
3700 get_nbl_exclusions_1(nbl,A2CI(found),w,&nbl_excl);
3701
3702 nbl_excl->pair[AMODI(inner_e)*nbl->na_ci+inner_i] &=
3703 ~(1U << (AMODI(found)*GPU_NSUBCELL + si));
3704 }
3705
3706 #undef AMODI
3707 #undef A2CI
3708 }
3709 }
3710 }
3711 }
3712 }
3713 }
3714
3715 /* Reallocate the simple ci list for at least n entries */
3716 static void nb_realloc_ci(nbnxn_pairlist_t *nbl,int n)
3717 {
3718 nbl->ci_nalloc = over_alloc_small(n);
3719 nb_realloc_void((void **)&nbl->ci,
3720 nbl->nci*sizeof(*nbl->ci),
3721 nbl->ci_nalloc*sizeof(*nbl->ci),
3722 nbl->alloc,nbl->free);
3723 }
3724
3725 /* Reallocate the super-cell sci list for at least n entries */
3726 static void nb_realloc_sci(nbnxn_pairlist_t *nbl,int n)
3727 {
3728 nbl->sci_nalloc = over_alloc_small(n);
3729 nb_realloc_void((void **)&nbl->sci,
3730 nbl->nsci*sizeof(*nbl->sci),
3731 nbl->sci_nalloc*sizeof(*nbl->sci),
3732 nbl->alloc,nbl->free);
3733 }
3734
3735 /* Make a new ci entry at index nbl->nci */
3736 static void new_ci_entry(nbnxn_pairlist_t *nbl,int ci,int shift,int flags,
3737 nbnxn_list_work_t *work)
3738 {
3739 if (nbl->nci + 1 > nbl->ci_nalloc)
3740 {
3741 nb_realloc_ci(nbl,nbl->nci+1);
3742 }
3743 nbl->ci[nbl->nci].ci = ci;
3744 nbl->ci[nbl->nci].shift = shift;
3745 /* Store the interaction flags along with the shift */
3746 nbl->ci[nbl->nci].shift |= flags;
3747 nbl->ci[nbl->nci].cj_ind_start = nbl->ncj;
3748 nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
3749 }
3750
3751 /* Make a new sci entry at index nbl->nsci */
3752 static void new_sci_entry(nbnxn_pairlist_t *nbl,int sci,int shift,int flags,
3753 nbnxn_list_work_t *work)
3754 {
3755 if (nbl->nsci + 1 > nbl->sci_nalloc)
3756 {
3757 nb_realloc_sci(nbl,nbl->nsci+1);
3758 }
3759 nbl->sci[nbl->nsci].sci = sci;
3760 nbl->sci[nbl->nsci].shift = shift;
3761 nbl->sci[nbl->nsci].cj4_ind_start = nbl->ncj4;
3762 nbl->sci[nbl->nsci].cj4_ind_end = nbl->ncj4;
3763 }
3764
3765 /* Sort the simple j-list cj on exclusions.
3766 * Entries with exclusions will all be sorted to the beginning of the list.
3767 */
3768 static void sort_cj_excl(nbnxn_cj_t *cj,int ncj,
3769 nbnxn_list_work_t *work)
3770 {
3771 int jnew,j;
3772
3773 if (ncj > work->cj_nalloc)
3774 {
3775 work->cj_nalloc = over_alloc_large(ncj);
3776 srenew(work->cj,work->cj_nalloc);
3777 }
3778
3779 /* Make a list of the j-cells involving exclusions */
3780 jnew = 0;
3781 for(j=0; j<ncj; j++)
3782 {
3783 if (cj[j].excl != NBNXN_INT_MASK_ALL)
3784 {
3785 work->cj[jnew++] = cj[j];
3786 }
3787 }
3788 /* Check if there are exclusions at all or not just the first entry */
3789 if (!((jnew == 0) ||
3790 (jnew == 1 && cj[0].excl != NBNXN_INT_MASK_ALL)))
3791 {
3792 for(j=0; j<ncj; j++)
3793 {
3794 if (cj[j].excl == NBNXN_INT_MASK_ALL)
3795 {
3796 work->cj[jnew++] = cj[j];
3797 }
3798 }
3799 for(j=0; j<ncj; j++)
3800 {
3801 cj[j] = work->cj[j];
3802 }
3803 }
3804 }
3805
3806 /* Close this simple list i entry */
3807 static void close_ci_entry_simple(nbnxn_pairlist_t *nbl)
3808 {
3809 int jlen;
3810
3811 /* All content of the new ci entry have already been filled correctly,
3812 * we only need to increase the count here (for non empty lists).
3813 */
3814 jlen = nbl->ci[nbl->nci].cj_ind_end - nbl->ci[nbl->nci].cj_ind_start;
3815 if (jlen > 0)
3816 {
3817 sort_cj_excl(nbl->cj+nbl->ci[nbl->nci].cj_ind_start,jlen,nbl->work);
3818
3819 if (nbl->ci[nbl->nci].shift & NBNXN_CI_HALF_LJ(0))
3820 {
3821 nbl->work->ncj_hlj += jlen;
3822 }
3823 else if (!(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_COUL(0)))
3824 {
3825 nbl->work->ncj_noq += jlen;
3826 }
3827
3828 nbl->nci++;
3829 }
3830 }
3831
3832 /* Split sci entry for load balancing on the GPU.
3833 * As we only now the current count on our own thread,
3834 * we will need to estimate the current total amount of i-entries.
3835 * As the lists get concatenated later, this estimate depends
3836 * both on nthread and our own thread index thread.
3837 */
3838 static void split_sci_entry(nbnxn_pairlist_t *nbl,
3839 int nsp_max_av,gmx_bool progBal,int nc_bal,
3840 int thread,int nthread)
3841 {
3842 int nsci_est;
3843 int nsp_max;
3844 int cj4_start,cj4_end,j4len,cj4;
3845 int sci;
3846 int nsp,nsp_sci,nsp_cj4,nsp_cj4_e,nsp_cj4_p;
3847 int p;
3848
3849 /* Estimate the total numbers of ci's of the nblist combined
3850 * over all threads using the target number of ci's.
3851 */
3852 nsci_est = nc_bal*thread/nthread + nbl->nsci;
3853 if (progBal)
3854 {
3855 /* The first ci blocks should be larger, to avoid overhead.
3856 * The last ci blocks should be smaller, to improve load balancing.
3857 */
3858 nsp_max = max(1,
3859 nsp_max_av*nc_bal*3/(2*(nsci_est - 1 + nc_bal)));
3860 }
3861 else
3862 {
3863 nsp_max = nsp_max_av;
3864 }
3865
3866 cj4_start = nbl->sci[nbl->nsci-1].cj4_ind_start;
3867 cj4_end = nbl->sci[nbl->nsci-1].cj4_ind_end;
3868 j4len = cj4_end - cj4_start;
3869
3870 if (j4len > 1 && j4len*GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE > nsp_max)
3871 {
3872 /* Remove the last ci entry and process the cj4's again */
3873 nbl->nsci -= 1;
3874
3875 sci = nbl->nsci;
3876 cj4 = cj4_start;
3877 nsp = 0;
3878 nsp_sci = 0;
3879 nsp_cj4_e = 0;
3880 nsp_cj4 = 0;
3881 while (cj4 < cj4_end)
3882 {
3883 nsp_cj4_p = nsp_cj4;
3884 nsp_cj4 = 0;
3885 for(p=0; p<GPU_NSUBCELL*NBNXN_GPU_JGROUP_SIZE; p++)
3886 {
3887 nsp_cj4 += (nbl->cj4[cj4].imei[0].imask >> p) & 1;
3888 }
3889 nsp += nsp_cj4;
3890
3891 if (nsp > nsp_max && nsp > nsp_cj4)
3892 {
3893 nbl->sci[sci].cj4_ind_end = cj4;
3894 sci++;
3895 nbl->nsci++;
3896 if (nbl->nsci+1 > nbl->sci_nalloc)
3897 {
3898 nb_realloc_sci(nbl,nbl->nsci+1);
3899 }
3900 nbl->sci[sci].sci = nbl->sci[nbl->nsci-1].sci;
3901 nbl->sci[sci].shift = nbl->sci[nbl->nsci-1].shift;
3902 nbl->sci[sci].cj4_ind_start = cj4;
3903 nsp_sci = nsp - nsp_cj4;
3904 nsp_cj4_e = nsp_cj4_p;
3905 nsp = nsp_cj4;
3906 }
3907
3908 cj4++;
3909 }
3910
3911 /* Put the remaining cj4's in a new ci entry */
3912 nbl->sci[sci].cj4_ind_end = cj4_end;
3913
3914 /* Possibly balance out the last two ci's
3915 * by moving the last cj4 of the second last ci.
3916 */
3917 if (nsp_sci - nsp_cj4_e >= nsp + nsp_cj4_e)
3918 {
3919 nbl->sci[sci-1].cj4_ind_end--;
3920 nbl->sci[sci].cj4_ind_start--;
3921 }
3922
3923 sci++;
3924 nbl->nsci++;
3925 }
3926 }
3927
3928 /* Clost this super/sub list i entry */
3929 static void close_ci_entry_supersub(nbnxn_pairlist_t *nbl,
3930 int nsp_max_av,
3931 gmx_bool progBal,int nc_bal,
3932 int thread,int nthread)
3933 {
3934 int j4len,tlen;
3935 int nb,b;
3936
3937 /* All content of the new ci entry have already been filled correctly,
3938 * we only need to increase the count here (for non empty lists).
3939 */
3940 j4len = nbl->sci[nbl->nsci].cj4_ind_end - nbl->sci[nbl->nsci].cj4_ind_start;
3941 if (j4len > 0)
3942 {
3943 /* We can only have complete blocks of 4 j-entries in a list,
3944 * so round the count up before closing.
3945 */
3946 nbl->ncj4 = ((nbl->work->cj_ind + 4-1) >> 2);
3947 nbl->work->cj_ind = nbl->ncj4*NBNXN_GPU_JGROUP_SIZE;
3948
3949 nbl->nsci++;
3950
3951 if (nsp_max_av > 0)
3952 {
3953 split_sci_entry(nbl,nsp_max_av,progBal,nc_bal,thread,nthread);
3954 }
3955 }
3956 }
3957
3958 /* Syncs the working array before adding another grid pair to the list */
3959 static void sync_work(nbnxn_pairlist_t *nbl)
3960 {
3961 if (!nbl->bSimple)
3962 {
3963 nbl->work->cj_ind = nbl->ncj4*NBNXN_GPU_JGROUP_SIZE;
3964 nbl->work->cj4_init = nbl->ncj4;
3965 }
3966 }
3967
3968 /* Clears an nbnxn_pairlist_t data structure */
3969 static void clear_pairlist(nbnxn_pairlist_t *nbl)
3970 {
3971 nbl->nci = 0;
3972 nbl->nsci = 0;
3973 nbl->ncj = 0;
3974 nbl->ncj4 = 0;
3975 nbl->nci_tot = 0;
3976 nbl->nexcl = 1;
3977
3978 nbl->work->ncj_noq = 0;
3979 nbl->work->ncj_hlj = 0;
3980 }
3981
3982 /* Sets a simple list i-cell bounding box, including PBC shift */
3983 static void set_icell_bb_simple(const float *bb,int ci,
3984 real shx,real shy,real shz,
3985 float *bb_ci)
3986 {
3987 int ia;
3988
3989 ia = ci*NNBSBB_B;
3990 bb_ci[BBL_X] = bb[ia+BBL_X] + shx;
3991 bb_ci[BBL_Y] = bb[ia+BBL_Y] + shy;
3992 bb_ci[BBL_Z] = bb[ia+BBL_Z] + shz;
3993 bb_ci[BBU_X] = bb[ia+BBU_X] + shx;
3994 bb_ci[BBU_Y] = bb[ia+BBU_Y] + shy;
3995 bb_ci[BBU_Z] = bb[ia+BBU_Z] + shz;
3996 }
3997
3998 /* Sets a super-cell and sub cell bounding boxes, including PBC shift */
3999 static void set_icell_bb_supersub(const float *bb,int ci,
4000 real shx,real shy,real shz,
4001 float *bb_ci)
4002 {
4003 int ia,m,i;
4004
4005 #ifdef NBNXN_BBXXXX
4006 ia = ci*(GPU_NSUBCELL>>STRIDE_8BB_2LOG)*NNBSBB_XXXX;
4007 for(m=0; m<(GPU_NSUBCELL>>STRIDE_8BB_2LOG)*NNBSBB_XXXX; m+=NNBSBB_XXXX)
4008 {
4009 for(i=0; i<STRIDE_8BB; i++)
4010 {
4011 bb_ci[m+0*STRIDE_8BB+i] = bb[ia+m+0*STRIDE_8BB+i] + shx;
4012 bb_ci[m+1*STRIDE_8BB+i] = bb[ia+m+1*STRIDE_8BB+i] + shy;
4013 bb_ci[m+2*STRIDE_8BB+i] = bb[ia+m+2*STRIDE_8BB+i] + shz;
4014 bb_ci[m+3*STRIDE_8BB+i] = bb[ia+m+3*STRIDE_8BB+i] + shx;
4015 bb_ci[m+4*STRIDE_8BB+i] = bb[ia+m+4*STRIDE_8BB+i] + shy;
4016 bb_ci[m+5*STRIDE_8BB+i] = bb[ia+m+5*STRIDE_8BB+i] + shz;
4017 }
4018 }
4019 #else
4020 ia = ci*GPU_NSUBCELL*NNBSBB_B;
4021 for(i=0; i<GPU_NSUBCELL*NNBSBB_B; i+=NNBSBB_B)
4022 {
4023 bb_ci[BBL_X] = bb[ia+BBL_X] + shx;
4024 bb_ci[BBL_Y] = bb[ia+BBL_Y] + shy;
4025 bb_ci[BBL_Z] = bb[ia+BBL_Z] + shz;
4026 bb_ci[BBU_X] = bb[ia+BBU_X] + shx;
4027 bb_ci[BBU_Y] = bb[ia+BBU_Y] + shy;
4028 bb_ci[BBU_Z] = bb[ia+BBU_Z] + shz;
4029 }
4030 #endif
4031 }
4032
4033 /* Copies PBC shifted i-cell atom coordinates x,y,z to working array */
4034 static void icell_set_x_simple(int ci,
4035 real shx,real shy,real shz,
4036 int na_c,
4037 int stride,const real *x,
4038 nbnxn_list_work_t *work)
4039 {
4040 int ia,i;
4041
4042 ia = ci*NBNXN_CPU_CLUSTER_I_SIZE;
4043
4044 for(i=0; i<NBNXN_CPU_CLUSTER_I_SIZE; i++)
4045 {
4046 work->x_ci[i*STRIDE_XYZ+XX] = x[(ia+i)*stride+XX] + shx;
4047 work->x_ci[i*STRIDE_XYZ+YY] = x[(ia+i)*stride+YY] + shy;
4048 work->x_ci[i*STRIDE_XYZ+ZZ] = x[(ia+i)*stride+ZZ] + shz;
4049 }
4050 }
4051
4052 /* Copies PBC shifted super-cell atom coordinates x,y,z to working array */
4053 static void icell_set_x_supersub(int ci,
4054 real shx,real shy,real shz,
4055 int na_c,
4056 int stride,const real *x,
4057 nbnxn_list_work_t *work)
4058 {
4059 int ia,i;
4060 real *x_ci;
4061
4062 x_ci = work->x_ci;
4063
4064 ia = ci*GPU_NSUBCELL*na_c;
4065 for(i=0; i<GPU_NSUBCELL*na_c; i++)
4066 {
4067 x_ci[i*DIM + XX] = x[(ia+i)*stride + XX] + shx;
4068 x_ci[i*DIM + YY] = x[(ia+i)*stride + YY] + shy;
4069 x_ci[i*DIM + ZZ] = x[(ia+i)*stride + ZZ] + shz;
4070 }
4071 }
4072
4073 #ifdef NBNXN_SEARCH_SSE
4074 /* Copies PBC shifted super-cell packed atom coordinates to working array */
4075 static void icell_set_x_supersub_sse8(int ci,
4076 real shx,real shy,real shz,
4077 int na_c,
4078 int stride,const real *x,
4079 nbnxn_list_work_t *work)
4080 {
4081 int si,io,ia,i,j;
4082 real *x_ci;
4083
4084 x_ci = work->x_ci;
4085
4086 for(si=0; si<GPU_NSUBCELL; si++)
4087 {
4088 for(i=0; i<na_c; i+=STRIDE_8BB)
4089 {
4090 io = si*na_c + i;
4091 ia = ci*GPU_NSUBCELL*na_c + io;
4092 for(j=0; j<STRIDE_8BB; j++)
4093 {
4094 x_ci[io*DIM + j + XX*STRIDE_8BB] = x[(ia+j)*stride+XX] + shx;
4095 x_ci[io*DIM + j + YY*STRIDE_8BB] = x[(ia+j)*stride+YY] + shy;
4096 x_ci[io*DIM + j + ZZ*STRIDE_8BB] = x[(ia+j)*stride+ZZ] + shz;
4097 }
4098 }
4099 }
4100 }
4101 #endif
4102
4103 static real nbnxn_rlist_inc_nonloc_fac = 0.6;
4104
4105 /* Due to the cluster size the effective pair-list is longer than
4106 * that of a simple atom pair-list. This function gives the extra distance.
4107 */
4108 real nbnxn_get_rlist_effective_inc(int cluster_size,real atom_density)
4109 {
4110 return ((0.5 + nbnxn_rlist_inc_nonloc_fac)*sqr(((cluster_size) - 1.0)/(cluster_size))*pow((cluster_size)/(atom_density),1.0/3.0));
4111 }
4112
4113 /* Estimates the interaction volume^2 for non-local interactions */
4114 static real nonlocal_vol2(const gmx_domdec_zones_t *zones,rvec ls,real r)
4115 {
4116 int z,d;
4117 real cl,ca,za;
4118 real vold_est;
4119 real vol2_est_tot;
4120
4121 vol2_est_tot = 0;
4122
4123 /* Here we simply add up the volumes of 1, 2 or 3 1D decomposition
4124 * not home interaction volume^2. As these volumes are not additive,
4125 * this is an overestimate, but it would only be significant in the limit
4126 * of small cells, where we anyhow need to split the lists into
4127 * as small parts as possible.
4128 */
4129
4130 for(z=0; z<zones->n; z++)
4131 {
4132 if (zones->shift[z][XX] + zones->shift[z][YY] + zones->shift[z][ZZ] == 1)
4133 {
4134 cl = 0;
4135 ca = 1;
4136 za = 1;
4137 for(d=0; d<DIM; d++)
4138 {
4139 if (zones->shift[z][d] == 0)
4140 {
4141 cl += 0.5*ls[d];
4142 ca *= ls[d];
4143 za *= zones->size[z].x1[d] - zones->size[z].x0[d];
4144 }
4145 }
4146
4147 /* 4 octants of a sphere */
4148 vold_est = 0.25*M_PI*r*r*r*r;
4149 /* 4 quarter pie slices on the edges */
4150 vold_est += 4*cl*M_PI/6.0*r*r*r;
4151 /* One rectangular volume on a face */
4152 vold_est += ca*0.5*r*r;
4153
4154 vol2_est_tot += vold_est*za;
4155 }
4156 }
4157
4158 return vol2_est_tot;
4159 }
4160
4161 /* Estimates the average size of a full j-list for super/sub setup */
4162 static int get_nsubpair_max(const nbnxn_search_t nbs,
4163 int iloc,
4164 real rlist,
4165 int min_ci_balanced)
4166 {
4167 const nbnxn_grid_t *grid;
4168 rvec ls;
4169 real xy_diag2,r_eff_sup,vol_est,nsp_est,nsp_est_nl;
4170 int nsubpair_max;
4171
4172 grid = &nbs->grid[0];
4173
4174 ls[XX] = (grid->c1[XX] - grid->c0[XX])/(grid->ncx*GPU_NSUBCELL_X);
4175 ls[YY] = (grid->c1[YY] - grid->c0[YY])/(grid->ncy*GPU_NSUBCELL_Y);
4176 ls[ZZ] = (grid->c1[ZZ] - grid->c0[ZZ])*grid->ncx*grid->ncy/(grid->nc*GPU_NSUBCELL_Z);
4177
4178 /* The average squared length of the diagonal of a sub cell */
4179 xy_diag2 = ls[XX]*ls[XX] + ls[YY]*ls[YY] + ls[ZZ]*ls[ZZ];
4180
4181 /* The formulas below are a heuristic estimate of the average nsj per si*/
4182 r_eff_sup = rlist + nbnxn_rlist_inc_nonloc_fac*sqr((grid->na_c - 1.0)/grid->na_c)*sqrt(xy_diag2/3);
4183
4184 if (!nbs->DomDec || nbs->zones->n == 1)
4185 {
4186 nsp_est_nl = 0;
4187 }
4188 else
4189 {
4190 nsp_est_nl =
4191 sqr(grid->atom_density/grid->na_c)*
4192 nonlocal_vol2(nbs->zones,ls,r_eff_sup);
4193 }
4194
4195 if (LOCAL_I(iloc))
4196 {
4197 /* Sub-cell interacts with itself */
4198 vol_est = ls[XX]*ls[YY]*ls[ZZ];
4199 /* 6/2 rectangular volume on the faces */
4200 vol_est += (ls[XX]*ls[YY] + ls[XX]*ls[ZZ] + ls[YY]*ls[ZZ])*r_eff_sup;
4201 /* 12/2 quarter pie slices on the edges */
4202 vol_est += 2*(ls[XX] + ls[YY] + ls[ZZ])*0.25*M_PI*sqr(r_eff_sup);
4203 /* 4 octants of a sphere */
4204 vol_est += 0.5*4.0/3.0*M_PI*pow(r_eff_sup,3);
4205
4206 nsp_est = grid->nsubc_tot*vol_est*grid->atom_density/grid->na_c;
4207
4208 /* Subtract the non-local pair count */
4209 nsp_est -= nsp_est_nl;
4210
4211 if (debug)
4212 {
4213 fprintf(debug,"nsp_est local %5.1f non-local %5.1f\n",
4214 nsp_est,nsp_est_nl);
4215 }
4216 }
4217 else
4218 {
4219 nsp_est = nsp_est_nl;
4220 }
4221
4222 if (min_ci_balanced <= 0 || grid->nc >= min_ci_balanced || grid->nc == 0)
4223 {
4224 /* We don't need to worry */
4225 nsubpair_max = -1;
4226 }
4227 else
4228 {
4229 /* Thus the (average) maximum j-list size should be as follows */
4230 nsubpair_max = max(1,(int)(nsp_est/min_ci_balanced+0.5));
4231
4232 /* Since the target value is a maximum (this avoid high outliers,
4233 * which lead to load imbalance), not average, we get more lists
4234 * than we ask for (to compensate we need to add GPU_NSUBCELL*4/4).
4235 * But more importantly, the optimal GPU performance moves
4236 * to lower number of block for very small blocks.
4237 * To compensate we add the maximum pair count per cj4.
4238 */
4239 nsubpair_max += GPU_NSUBCELL*NBNXN_CPU_CLUSTER_I_SIZE;
4240 }
4241
4242 if (debug)
4243 {
4244 fprintf(debug,"nbl nsp estimate %.1f, nsubpair_max %d\n",
4245 nsp_est,nsubpair_max);
4246 }
4247
4248 return nsubpair_max;
4249 }
4250
4251 /* Debug list print function */
4252 static void print_nblist_ci_cj(FILE *fp,const nbnxn_pairlist_t *nbl)
4253 {
4254 int i,j;
4255
4256 for(i=0; i<nbl->nci; i++)
4257 {
4258 fprintf(fp,"ci %4d shift %2d ncj %3d\n",
4259 nbl->ci[i].ci,nbl->ci[i].shift,
4260 nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start);
4261
4262 for(j=nbl->ci[i].cj_ind_start; j<nbl->ci[i].cj_ind_end; j++)
4263 {
4264 fprintf(fp," cj %5d imask %x\n",
4265 nbl->cj[j].cj,
4266 nbl->cj[j].excl);
4267 }
4268 }
4269 }
4270
4271 /* Debug list print function */
4272 static void print_nblist_sci_cj(FILE *fp,const nbnxn_pairlist_t *nbl)
4273 {
4274 int i,j4,j;
4275
4276 for(i=0; i<nbl->nsci; i++)
4277 {
4278 fprintf(fp,"ci %4d shift %2d ncj4 %2d\n",
4279 nbl->sci[i].sci,nbl->sci[i].shift,
4280 nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start);
4281
4282 for(j4=nbl->sci[i].cj4_ind_start; j4<nbl->sci[i].cj4_ind_end; j4++)
4283 {
4284 for(j=0; j<4; j++)
4285 {
4286 fprintf(fp," sj %5d imask %x\n",
4287 nbl->cj4[j4].cj[j],
4288 nbl->cj4[j4].imei[0].imask);
4289 }
4290 }
4291 }
4292 }
4293
4294 /* Combine pair lists *nbl generated on multiple threads nblc */
4295 static void combine_nblists(int nnbl,nbnxn_pairlist_t **nbl,
4296 nbnxn_pairlist_t *nblc)
4297 {
4298 int nsci,ncj4,nexcl;
4299 int n,i;
4300
4301 if (nblc->bSimple)
4302 {
4303 gmx_incons("combine_nblists does not support simple lists");
4304 }
4305
4306 nsci = nblc->nsci;
4307 ncj4 = nblc->ncj4;
4308 nexcl = nblc->nexcl;
4309 for(i=0; i<nnbl; i++)
4310 {
4311 nsci += nbl[i]->nsci;
4312 ncj4 += nbl[i]->ncj4;
4313 nexcl += nbl[i]->nexcl;
4314 }
4315
4316 if (nsci > nblc->sci_nalloc)
4317 {
4318 nb_realloc_sci(nblc,nsci);
4319 }
4320 if (ncj4 > nblc->cj4_nalloc)
4321 {
4322 nblc->cj4_nalloc = over_alloc_small(ncj4);
4323 nb_realloc_void((void **)&nblc->cj4,
4324 nblc->ncj4*sizeof(*nblc->cj4),
4325 nblc->cj4_nalloc*sizeof(*nblc->cj4),
4326 nblc->alloc,nblc->free);
4327 }
4328 if (nexcl > nblc->excl_nalloc)
4329 {
4330 nblc->excl_nalloc = over_alloc_small(nexcl);
4331 nb_realloc_void((void **)&nblc->excl,
4332 nblc->nexcl*sizeof(*nblc->excl),
4333 nblc->excl_nalloc*sizeof(*nblc->excl),
4334 nblc->alloc,nblc->free);
4335 }
4336
4337 /* Each thread should copy its own data to the combined arrays,
4338 * as otherwise data will go back and forth between different caches.
4339 */
4340 #pragma omp parallel for num_threads(gmx_omp_nthreads_get(emntPairsearch)) schedule(static)
4341 for(n=0; n<nnbl; n++)
4342 {
4343 int sci_offset;
4344 int cj4_offset;
4345 int ci_offset;
4346 int excl_offset;
4347 int i,j4;
4348 const nbnxn_pairlist_t *nbli;
4349
4350 /* Determine the offset in the combined data for our thread */
4351 sci_offset = nblc->nsci;
4352 cj4_offset = nblc->ncj4;
4353 ci_offset = nblc->nci_tot;
4354 excl_offset = nblc->nexcl;
4355
4356 for(i=0; i<n; i++)
4357 {
4358 sci_offset += nbl[i]->nsci;
4359 cj4_offset += nbl[i]->ncj4;
4360 ci_offset += nbl[i]->nci_tot;
4361 excl_offset += nbl[i]->nexcl;
4362 }
4363
4364 nbli = nbl[n];
4365
4366 for(i=0; i<nbli->nsci; i++)
4367 {
4368 nblc->sci[sci_offset+i] = nbli->sci[i];
4369 nblc->sci[sci_offset+i].cj4_ind_start += cj4_offset;
4370 nblc->sci[sci_offset+i].cj4_ind_end += cj4_offset;
4371 }
4372
4373 for(j4=0; j4<nbli->ncj4; j4++)
4374 {
4375 nblc->cj4[cj4_offset+j4] = nbli->cj4[j4];
4376 nblc->cj4[cj4_offset+j4].imei[0].excl_ind += excl_offset;
4377 nblc->cj4[cj4_offset+j4].imei[1].excl_ind += excl_offset;
4378 }
4379
4380 for(j4=0; j4<nbli->nexcl; j4++)
4381 {
4382 nblc->excl[excl_offset+j4] = nbli->excl[j4];
4383 }
4384 }
4385
4386 for(n=0; n<nnbl; n++)
4387 {
4388 nblc->nsci += nbl[n]->nsci;
4389 nblc->ncj4 += nbl[n]->ncj4;
4390 nblc->nci_tot += nbl[n]->nci_tot;
4391 nblc->nexcl += nbl[n]->nexcl;
4392 }
4393 }
4394
4395 /* Returns the next ci to be processes by our thread */
4396 static gmx_bool next_ci(const nbnxn_grid_t *grid,
4397 int conv,
4398 int nth,int ci_block,
4399 int *ci_x,int *ci_y,
4400 int *ci_b,int *ci)
4401 {
4402 (*ci_b)++;
4403 (*ci)++;
4404
4405 if (*ci_b == ci_block)
4406 {
4407 /* Jump to the next block assigned to this task */
4408 *ci += (nth - 1)*ci_block;
4409 *ci_b = 0;
4410 }
4411
4412 if (*ci >= grid->nc*conv)
4413 {
4414 return FALSE;
4415 }
4416
4417 while (*ci >= grid->cxy_ind[*ci_x*grid->ncy + *ci_y + 1]*conv)
4418 {
4419 *ci_y += 1;
4420 if (*ci_y == grid->ncy)
4421 {
4422 *ci_x += 1;
4423 *ci_y = 0;
4424 }
4425 }
4426
4427 return TRUE;
4428 }
4429
4430 /* Returns the distance^2 for which we put cell pairs in the list
4431 * without checking atom pair distances. This is usually < rlist^2.
4432 */
4433 static float boundingbox_only_distance2(const nbnxn_grid_t *gridi,
4434 const nbnxn_grid_t *gridj,
4435 real rlist,
4436 gmx_bool simple)
4437 {
4438 /* If the distance between two sub-cell bounding boxes is less
4439 * than this distance, do not check the distance between
4440 * all particle pairs in the sub-cell, since then it is likely
4441 * that the box pair has atom pairs within the cut-off.
4442 * We use the nblist cut-off minus 0.5 times the average x/y diagonal
4443 * spacing of the sub-cells. Around 40% of the checked pairs are pruned.
4444 * Using more than 0.5 gains at most 0.5%.
4445 * If forces are calculated more than twice, the performance gain
4446 * in the force calculation outweighs the cost of checking.
4447 * Note that with subcell lists, the atom-pair distance check
4448 * is only performed when only 1 out of 8 sub-cells in within range,
4449 * this is because the GPU is much faster than the cpu.
4450 */
4451 real bbx,bby;
4452 real rbb2;
4453
4454 bbx = 0.5*(gridi->sx + gridj->sx);
4455 bby = 0.5*(gridi->sy + gridj->sy);
4456 if (!simple)
4457 {
4458 bbx /= GPU_NSUBCELL_X;
4459 bby /= GPU_NSUBCELL_Y;
4460 }
4461
4462 rbb2 = sqr(max(0,rlist - 0.5*sqrt(bbx*bbx + bby*bby)));
4463
4464 #ifndef GMX_DOUBLE
4465 return rbb2;
4466 #else
4467 return (float)((1+GMX_FLOAT_EPS)*rbb2);
4468 #endif
4469 }
4470
4471 /* Generates the part of pair-list nbl assigned to our thread */
4472 static void nbnxn_make_pairlist_part(const nbnxn_search_t nbs,
4473 const nbnxn_grid_t *gridi,
4474 const nbnxn_grid_t *gridj,
4475 nbnxn_search_work_t *work,
4476 const nbnxn_atomdata_t *nbat,
4477 const t_blocka *excl,
4478 real rlist,
4479 int nb_kernel_type,
4480 int nsubpair_max,
4481 gmx_bool progBal,
4482 int min_ci_balanced,
4483 int th,int nth,
4484 nbnxn_pairlist_t *nbl)
4485 {
4486 int na_cj_2log;
4487 matrix box;
4488 real rl2;
4489 float rbb2;
4490 int d;
4491 int ci_block,ci_b,ci,ci_x,ci_y,ci_xy,cj;
4492 ivec shp;
4493 int tx,ty,tz;
4494 int shift;
4495 gmx_bool bMakeList;
4496 real shx,shy,shz;
4497 int conv_i,cell0_i;
4498 const float *bb_i,*bbcz_i,*bbcz_j;
4499 const int *flags_i;
4500 real bx0,bx1,by0,by1,bz0,bz1;
4501 real bz1_frac;
4502 real d2cx,d2z,d2z_cx,d2z_cy,d2zx,d2zxy,d2xy;
4503 int cxf,cxl,cyf,cyf_x,cyl;
4504 int cx,cy;
4505 int c0,c1,cs,cf,cl;
4506 int ndistc;
4507 int ncpcheck;
4508
4509 nbs_cycle_start(&work->cc[enbsCCsearch]);
4510
4511 if (gridj->bSimple != nbl->bSimple)
4512 {
4513 gmx_incons("Grid incompatible with pair-list");
4514 }
4515
4516 sync_work(nbl);
4517
4518 nbl->na_sc = gridj->na_sc;
4519 nbl->na_ci = gridj->na_c;
4520 nbl->na_cj = kernel_to_cj_size(nb_kernel_type);
4521 na_cj_2log = get_2log(nbl->na_cj);
4522
4523 nbl->rlist = rlist;
4524
4525 copy_mat(nbs->box,box);
4526
4527 rl2 = nbl->rlist*nbl->rlist;
4528
4529 rbb2 = boundingbox_only_distance2(gridi,gridj,nbl->rlist,nbl->bSimple);
4530
4531 if (debug)
4532 {
4533 fprintf(debug,"nbl bounding box only distance %f\n",sqrt(rbb2));
4534 }
4535
4536 /* Set the shift range */
4537 for(d=0; d<DIM; d++)
4538 {
4539 /* Check if we need periodicity shifts.
4540 * Without PBC or with domain decomposition we don't need them.
4541 */
4542 if (d >= ePBC2npbcdim(nbs->ePBC) || nbs->dd_dim[d])
4543 {
4544 shp[d] = 0;
4545 }
4546 else
4547 {
4548 if (d == XX &&
4549 box[XX][XX] - fabs(box[YY][XX]) - fabs(box[ZZ][XX]) < sqrt(rl2))
4550 {
4551 shp[d] = 2;
4552 }
4553 else
4554 {
4555 shp[d] = 1;
4556 }
4557 }
4558 }
4559
4560 if (nbl->bSimple && !gridi->bSimple)
4561 {
4562 conv_i = gridi->na_sc/gridj->na_sc;
4563 bb_i = gridi->bb_simple;
4564 bbcz_i = gridi->bbcz_simple;
4565 flags_i = gridi->flags_simple;
4566 }
4567 else
4568 {
4569 conv_i = 1;
4570 bb_i = gridi->bb;
4571 bbcz_i = gridi->bbcz;
4572 flags_i = gridi->flags;
4573 }
4574 cell0_i = gridi->cell0*conv_i;
4575
4576 bbcz_j = gridj->bbcz;
4577
4578 if (conv_i == 1)
4579 {
4580 #define CI_BLOCK_ENUM 5
4581 #define CI_BLOCK_DENOM 11
4582 /* Here we decide how to distribute the blocks over the threads.
4583 * We use prime numbers to try to avoid that the grid size becomes
4584 * a multiple of the number of threads, which would lead to some
4585 * threads getting "inner" pairs and others getting boundary pairs,
4586 * which in turns will lead to load imbalance between threads.
4587 * Set the block size as 5/11/ntask times the average number of cells
4588 * in a y,z slab. This should ensure a quite uniform distribution
4589 * of the grid parts of the different thread along all three grid
4590 * zone boundaries with 3D domain decomposition. At the same time
4591 * the blocks will not become too small.
4592 */
4593 ci_block = (gridi->nc*CI_BLOCK_ENUM)/(CI_BLOCK_DENOM*gridi->ncx*nth);
4594
4595 /* Ensure the blocks are not too small: avoids cache invalidation */
4596 if (ci_block*gridi->na_sc < 16)
4597 {
4598 ci_block = (16 + gridi->na_sc - 1)/gridi->na_sc;
4599 }
4600
4601 /* Without domain decomposition
4602 * or with less than 3 blocks per task, divide in nth blocks.
4603 */
4604 if (!nbs->DomDec || ci_block*3*nth > gridi->nc)
4605 {
4606 ci_block = (gridi->nc + nth - 1)/nth;
4607 }
4608 }
4609 else
4610 {
4611 /* Blocks of the conversion factor - 1 give a large repeat count
4612 * combined with a small block size. This should result in good
4613 * load balancing for both small and large domains.
4614 */
4615 ci_block = conv_i - 1;
4616 }
4617 if (debug)
4618 {
4619 fprintf(debug,"nbl nc_i %d col.av. %.1f ci_block %d\n",
4620 gridi->nc,gridi->nc/(double)(gridi->ncx*gridi->ncy),ci_block);
4621 }
4622
4623 ndistc = 0;
4624 ncpcheck = 0;
4625
4626 ci_b = -1;
4627 ci = th*ci_block - 1;
4628 ci_x = 0;
4629 ci_y = 0;
4630 while (next_ci(gridi,conv_i,nth,ci_block,&ci_x,&ci_y,&ci_b,&ci))
4631 {
4632 if (nbl->bSimple && flags_i[ci] == 0)
4633 {
4634 continue;
4635 }
4636
4637 d2cx = 0;
4638 if (gridj != gridi && shp[XX] == 0)
4639 {
4640 if (nbl->bSimple)
4641 {
4642 bx1 = bb_i[ci*NNBSBB_B+NNBSBB_C+XX];
4643 }
4644 else
4645 {
4646 bx1 = gridi->c0[XX] + (ci_x+1)*gridi->sx;
4647 }
4648 if (bx1 < gridj->c0[XX])
4649 {
4650 d2cx = sqr(gridj->c0[XX] - bx1);
4651
4652 if (d2cx >= rl2)
4653 {
4654 continue;
4655 }
4656 }
4657 }
4658
4659 ci_xy = ci_x*gridi->ncy + ci_y;
4660
4661 /* Loop over shift vectors in three dimensions */
4662 for (tz=-shp[ZZ]; tz<=shp[ZZ]; tz++)
4663 {
4664 shz = tz*box[ZZ][ZZ];
4665
4666 bz0 = bbcz_i[ci*NNBSBB_D ] + shz;
4667 bz1 = bbcz_i[ci*NNBSBB_D+1] + shz;
4668
4669 if (tz == 0)
4670 {
4671 d2z = 0;
4672 }
4673 else if (tz < 0)
4674 {
4675 d2z = sqr(bz1);
4676 }
4677 else
4678 {
4679 d2z = sqr(bz0 - box[ZZ][ZZ]);
4680 }
4681
4682 d2z_cx = d2z + d2cx;
4683
4684 if (d2z_cx >= rl2)
4685 {
4686 continue;
4687 }
4688
4689 bz1_frac =
4690 bz1/((real)(gridi->cxy_ind[ci_xy+1] - gridi->cxy_ind[ci_xy]));
4691 if (bz1_frac < 0)
4692 {
4693 bz1_frac = 0;
4694 }
4695 /* The check with bz1_frac close to or larger than 1 comes later */
4696
4697 for (ty=-shp[YY]; ty<=shp[YY]; ty++)
4698 {
4699 shy = ty*box[YY][YY] + tz*box[ZZ][YY];
4700
4701 if (nbl->bSimple)
4702 {
4703 by0 = bb_i[ci*NNBSBB_B +YY] + shy;
4704 by1 = bb_i[ci*NNBSBB_B+NNBSBB_C+YY] + shy;
4705 }
4706 else
4707 {
4708 by0 = gridi->c0[YY] + (ci_y )*gridi->sy + shy;
4709 by1 = gridi->c0[YY] + (ci_y+1)*gridi->sy + shy;
4710 }
4711
4712 get_cell_range(by0,by1,
4713 gridj->ncy,gridj->c0[YY],gridj->sy,gridj->inv_sy,
4714 d2z_cx,rl2,
4715 &cyf,&cyl);
4716
4717 if (cyf > cyl)
4718 {
4719 continue;
4720 }
4721
4722 d2z_cy = d2z;
4723 if (by1 < gridj->c0[YY])
4724 {
4725 d2z_cy += sqr(gridj->c0[YY] - by1);
4726 }
4727 else if (by0 > gridj->c1[YY])
4728 {
4729 d2z_cy += sqr(by0 - gridj->c1[YY]);
4730 }
4731
4732 for (tx=-shp[XX]; tx<=shp[XX]; tx++)
4733 {
4734 shift = XYZ2IS(tx,ty,tz);
4735
4736 #ifdef NBNXN_SHIFT_BACKWARD
4737 if (gridi == gridj && shift > CENTRAL)
4738 {
4739 continue;
4740 }
4741 #endif
4742
4743 shx = tx*box[XX][XX] + ty*box[YY][XX] + tz*box[ZZ][XX];
4744
4745 if (nbl->bSimple)
4746 {
4747 bx0 = bb_i[ci*NNBSBB_B +XX] + shx;
4748 bx1 = bb_i[ci*NNBSBB_B+NNBSBB_C+XX] + shx;
4749 }
4750 else
4751 {
4752 bx0 = gridi->c0[XX] + (ci_x )*gridi->sx + shx;
4753 bx1 = gridi->c0[XX] + (ci_x+1)*gridi->sx + shx;
4754 }
4755
4756 get_cell_range(bx0,bx1,
4757 gridj->ncx,gridj->c0[XX],gridj->sx,gridj->inv_sx,
4758 d2z_cy,rl2,
4759 &cxf,&cxl);
4760
4761 if (cxf > cxl)
4762 {
4763 continue;
4764 }
4765
4766 if (nbl->bSimple)
4767 {
4768 new_ci_entry(nbl,cell0_i+ci,shift,flags_i[ci],
4769 nbl->work);
4770 }
4771 else
4772 {
4773 new_sci_entry(nbl,cell0_i+ci,shift,flags_i[ci],
4774 nbl->work);
4775 }
4776
4777 #ifndef NBNXN_SHIFT_BACKWARD
4778 if (cxf < ci_x)
4779 #else
4780 if (shift == CENTRAL && gridi == gridj &&
4781 cxf < ci_x)
4782 #endif
4783 {
4784 /* Leave the pairs with i > j.
4785 * x is the major index, so skip half of it.
4786 */
4787 cxf = ci_x;
4788 }
4789
4790 if (nbl->bSimple)
4791 {
4792 set_icell_bb_simple(bb_i,ci,shx,shy,shz,
4793 nbl->work->bb_ci);
4794 }
4795 else
4796 {
4797 set_icell_bb_supersub(bb_i,ci,shx,shy,shz,
4798 nbl->work->bb_ci);
4799 }
4800
4801 nbs->icell_set_x(cell0_i+ci,shx,shy,shz,
4802 gridi->na_c,nbat->xstride,nbat->x,
4803 nbl->work);
4804
4805 for(cx=cxf; cx<=cxl; cx++)
4806 {
4807 d2zx = d2z;
4808 if (gridj->c0[XX] + cx*gridj->sx > bx1)
4809 {
4810 d2zx += sqr(gridj->c0[XX] + cx*gridj->sx - bx1);
4811 }
4812 else if (gridj->c0[XX] + (cx+1)*gridj->sx < bx0)
4813 {
4814 d2zx += sqr(gridj->c0[XX] + (cx+1)*gridj->sx - bx0);
4815 }
4816
4817 #ifndef NBNXN_SHIFT_BACKWARD
4818 if (gridi == gridj &&
4819 cx == 0 && cyf < ci_y)
4820 #else
4821 if (gridi == gridj &&
4822 cx == 0 && shift == CENTRAL && cyf < ci_y)
4823 #endif
4824 {
4825 /* Leave the pairs with i > j.
4826 * Skip half of y when i and j have the same x.
4827 */
4828 cyf_x = ci_y;
4829 }
4830 else
4831 {
4832 cyf_x = cyf;
4833 }
4834
4835 for(cy=cyf_x; cy<=cyl; cy++)
4836 {
4837 c0 = gridj->cxy_ind[cx*gridj->ncy+cy];
4838 c1 = gridj->cxy_ind[cx*gridj->ncy+cy+1];
4839 #ifdef NBNXN_SHIFT_BACKWARD
4840 if (gridi == gridj &&
4841 shift == CENTRAL && c0 < ci)
4842 {
4843 c0 = ci;
4844 }
4845 #endif
4846
4847 d2zxy = d2zx;
4848 if (gridj->c0[YY] + cy*gridj->sy > by1)
4849 {
4850 d2zxy += sqr(gridj->c0[YY] + cy*gridj->sy - by1);
4851 }
4852 else if (gridj->c0[YY] + (cy+1)*gridj->sy < by0)
4853 {
4854 d2zxy += sqr(gridj->c0[YY] + (cy+1)*gridj->sy - by0);
4855 }
4856 if (c1 > c0 && d2zxy < rl2)
4857 {
4858 cs = c0 + (int)(bz1_frac*(c1 - c0));
4859 if (cs >= c1)
4860 {
4861 cs = c1 - 1;
4862 }
4863
4864 d2xy = d2zxy - d2z;
4865
4866 /* Find the lowest cell that can possibly
4867 * be within range.
4868 */
4869 cf = cs;
4870 while(cf > c0 &&
4871 (bbcz_j[cf*NNBSBB_D+1] >= bz0 ||
4872 d2xy + sqr(bbcz_j[cf*NNBSBB_D+1] - bz0) < rl2))
4873 {
4874 cf--;
4875 }
4876
4877 /* Find the highest cell that can possibly
4878 * be within range.
4879 */
4880 cl = cs;
4881 while(cl < c1-1 &&
4882 (bbcz_j[cl*NNBSBB_D] <= bz1 ||
4883 d2xy + sqr(bbcz_j[cl*NNBSBB_D] - bz1) < rl2))
4884 {
4885 cl++;
4886 }
4887
4888 #ifdef NBNXN_REFCODE
4889 {
4890 /* Simple reference code, for debugging,
4891 * overrides the more complex code above.
4892 */
4893 int k;
4894 cf = c1;
4895 cl = -1;
4896 for(k=c0; k<c1; k++)
4897 {
4898 if (box_dist2(bx0,bx1,by0,by1,bz0,bz1,
4899 bb+k*NNBSBB_B) < rl2 &&
4900 k < cf)
4901 {
4902 cf = k;
4903 }
4904 if (box_dist2(bx0,bx1,by0,by1,bz0,bz1,
4905 bb+k*NNBSBB_B) < rl2 &&
4906 k > cl)
4907 {
4908 cl = k;
4909 }
4910 }
4911 }
4912 #endif
4913
4914 if (gridi == gridj)
4915 {
4916 /* We want each atom/cell pair only once,
4917 * only use cj >= ci.
4918 */
4919 #ifndef NBNXN_SHIFT_BACKWARD
4920 cf = max(cf,ci);
4921 #else
4922 if (shift == CENTRAL)
4923 {
4924 cf = max(cf,ci);
4925 }
4926 #endif
4927 }
4928
4929 if (cf <= cl)
4930 {
4931 switch (nb_kernel_type)
4932 {
4933 case nbk4x4_PlainC:
4934 check_subcell_list_space_simple(nbl,cl-cf+1);
4935
4936 make_cluster_list_simple(gridj,
4937 nbl,ci,cf,cl,
4938 (gridi == gridj && shift == CENTRAL),
4939 nbat->x,
4940 rl2,rbb2,
4941 &ndistc);
4942 break;
4943 #ifdef NBNXN_SEARCH_SSE
4944 case nbk4xN_X86_SIMD128:
4945 check_subcell_list_space_simple(nbl,ci_to_cj(na_cj_2log,cl-cf)+2);
4946 make_cluster_list_x86_simd128(gridj,
4947 nbl,ci,cf,cl,
4948 (gridi == gridj && shift == CENTRAL),
4949 nbat->x,
4950 rl2,rbb2,
4951 &ndistc);
4952 break;
4953 #ifdef GMX_X86_AVX_256
4954 case nbk4xN_X86_SIMD256:
4955 check_subcell_list_space_simple(nbl,ci_to_cj(na_cj_2log,cl-cf)+2);
4956 make_cluster_list_x86_simd256(gridj,
4957 nbl,ci,cf,cl,
4958 (gridi == gridj && shift == CENTRAL),
4959 nbat->x,
4960 rl2,rbb2,
4961 &ndistc);
4962 break;
4963 #endif
4964 #endif
4965 case nbk8x8x8_PlainC:
4966 case nbk8x8x8_CUDA:
4967 check_subcell_list_space_supersub(nbl,cl-cf+1);
4968 for(cj=cf; cj<=cl; cj++)
4969 {
4970 make_cluster_list_supersub(nbs,gridi,gridj,
4971 nbl,ci,cj,
4972 (gridi == gridj && shift == CENTRAL && ci == cj),
4973 nbat->xstride,nbat->x,
4974 rl2,rbb2,
4975 &ndistc);
4976 }
4977 break;
4978 }
4979 ncpcheck += cl - cf + 1;
4980 }
4981 }
4982 }
4983 }
4984
4985 /* Set the exclusions for this ci list */
4986 if (nbl->bSimple)
4987 {
4988 set_ci_top_excls(nbs,
4989 nbl,
4990 shift == CENTRAL && gridi == gridj,
4991 gridj->na_c_2log,
4992 na_cj_2log,
4993 &(nbl->ci[nbl->nci]),
4994 excl);
4995 }
4996 else
4997 {
4998 set_sci_top_excls(nbs,
4999 nbl,
5000 shift == CENTRAL && gridi == gridj,
5001 gridj->na_c_2log,
5002 &(nbl->sci[nbl->nsci]),
5003 excl);
5004 }
5005
5006 /* Close this ci list */
5007 if (nbl->bSimple)
5008 {
5009 close_ci_entry_simple(nbl);
5010 }
5011 else
5012 {
5013 close_ci_entry_supersub(nbl,
5014 nsubpair_max,
5015 progBal,min_ci_balanced,
5016 th,nth);
5017 }
5018 }
5019 }
5020 }
5021 }
5022
5023 work->ndistc = ndistc;
5024
5025 nbs_cycle_stop(&work->cc[enbsCCsearch]);
5026
5027 if (debug)
5028 {
5029 fprintf(debug,"number of distance checks %d\n",ndistc);
5030 fprintf(debug,"ncpcheck %s %d\n",gridi==gridj ? "local" : "non-local",
5031 ncpcheck);
5032
5033 if (nbl->bSimple)
5034 {
5035 print_nblist_statistics_simple(debug,nbl,nbs,rlist);
5036 }
5037 else
5038 {
5039 print_nblist_statistics_supersub(debug,nbl,nbs,rlist);
5040 }
5041
5042 }
5043 }
5044
5045 /* Make a local or non-local pair-list, depending on iloc */
5046 void nbnxn_make_pairlist(const nbnxn_search_t nbs,
5047 const nbnxn_atomdata_t *nbat,
5048 const t_blocka *excl,
5049 real rlist,
5050 int min_ci_balanced,
5051 nbnxn_pairlist_set_t *nbl_list,
5052 int iloc,
5053 int nb_kernel_type,
5054 t_nrnb *nrnb)
5055 {
5056 const nbnxn_grid_t *gridi,*gridj;
5057 int nzi,zi,zj0,zj1,zj;
5058 int nsubpair_max;
5059 int nth,th;
5060 int nnbl;
5061 nbnxn_pairlist_t **nbl;
5062 gmx_bool CombineNBLists;
5063 int np_tot,np_noq,np_hlj,nap;
5064
5065 nnbl = nbl_list->nnbl;
5066 nbl = nbl_list->nbl;
5067 CombineNBLists = nbl_list->bCombined;
5068
5069 if (debug)
5070 {
5071 fprintf(debug,"ns making %d nblists\n", nnbl);
5072 }
5073
5074 if (nbl_list->bSimple)
5075 {
5076 switch (nb_kernel_type)
5077 {
5078 #ifdef NBNXN_SEARCH_SSE
5079 case nbk4xN_X86_SIMD128:
5080 nbs->icell_set_x = icell_set_x_x86_simd128;
5081 break;
5082 #ifdef GMX_X86_AVX_256
5083 case nbk4xN_X86_SIMD256:
5084 nbs->icell_set_x = icell_set_x_x86_simd256;
5085 break;
5086 #endif
5087 #endif
5088 default:
5089 nbs->icell_set_x = icell_set_x_simple;
5090 break;
5091 }
5092 }
5093 else
5094 {
5095 #ifdef NBNXN_SEARCH_SSE
5096 nbs->icell_set_x = icell_set_x_supersub_sse8;
5097 #else
5098 nbs->icell_set_x = icell_set_x_supersub;
5099 #endif
5100 }
5101
5102 if (LOCAL_I(iloc))
5103 {
5104 /* Only zone (grid) 0 vs 0 */
5105 nzi = 1;
5106 zj0 = 0;
5107 zj1 = 1;
5108 }
5109 else
5110 {
5111 nzi = nbs->zones->nizone;
5112 }
5113
5114 if (!nbl_list->bSimple && min_ci_balanced > 0)
5115 {
5116 nsubpair_max = get_nsubpair_max(nbs,iloc,rlist,min_ci_balanced);
5117 }
5118 else
5119 {
5120 nsubpair_max = 0;
5121 }
5122
5123 /* Clear all pair-lists */
5124 for(th=0; th<nnbl; th++)
5125 {
5126 clear_pairlist(nbl[th]);
5127 }
5128
5129 for(zi=0; zi<nzi; zi++)
5130 {
5131 gridi = &nbs->grid[zi];
5132
5133 if (NONLOCAL_I(iloc))
5134 {
5135 zj0 = nbs->zones->izone[zi].j0;
5136 zj1 = nbs->zones->izone[zi].j1;
5137 if (zi == 0)
5138 {
5139 zj0++;
5140 }
5141 }
5142 for(zj=zj0; zj<zj1; zj++)
5143 {
5144 gridj = &nbs->grid[zj];
5145
5146 if (debug)
5147 {
5148 fprintf(debug,"ns search grid %d vs %d\n",zi,zj);
5149 }
5150
5151 nbs_cycle_start(&nbs->cc[enbsCCsearch]);
5152
5153 #pragma omp parallel for num_threads(nnbl) schedule(static)
5154 for(th=0; th<nnbl; th++)
5155 {
5156 if (CombineNBLists && th > 0)
5157 {
5158 clear_pairlist(nbl[th]);
5159 }
5160
5161 /* Divide the i super cell equally over the nblists */
5162 nbnxn_make_pairlist_part(nbs,gridi,gridj,
5163 &nbs->work[th],nbat,excl,
5164 rlist,
5165 nb_kernel_type,
5166 nsubpair_max,
5167 (LOCAL_I(iloc) || nbs->zones->n <= 2),
5168 min_ci_balanced,
5169 th,nnbl,
5170 nbl[th]);
5171 }
5172 nbs_cycle_stop(&nbs->cc[enbsCCsearch]);
5173
5174 np_tot = 0;
5175 np_noq = 0;
5176 np_hlj = 0;
5177 for(th=0; th<nnbl; th++)
5178 {
5179 inc_nrnb(nrnb,eNR_NBNXN_DIST2,nbs->work[th].ndistc);
5180
5181 if (nbl_list->bSimple)
5182 {
5183 np_tot += nbl[th]->ncj;
5184 np_noq += nbl[th]->work->ncj_noq;
5185 np_hlj += nbl[th]->work->ncj_hlj;
5186 }
5187 else
5188 {
5189 /* This count ignores potential subsequent pair pruning */
5190 np_tot += nbl[th]->nci_tot;
5191 }
5192 }
5193 nap = nbl[0]->na_ci*nbl[0]->na_cj;
5194 nbl_list->natpair_ljq = (np_tot - np_noq)*nap - np_hlj*nap/2;
5195 nbl_list->natpair_lj = np_noq*nap;
5196 nbl_list->natpair_q = np_hlj*nap/2;
5197
5198 if (CombineNBLists && nnbl > 1)
5199 {
5200 nbs_cycle_start(&nbs->cc[enbsCCcombine]);
5201
5202 combine_nblists(nnbl-1,nbl+1,nbl[0]);
5203
5204 nbs_cycle_stop(&nbs->cc[enbsCCcombine]);
5205 }
5206
5207 }
5208 }
5209
5210 /*
5211 print_supersub_nsp("nsubpair",nbl[0],iloc);
5212 */
5213
5214 /* Special performance logging stuff (env.var. GMX_NBNXN_CYCLE) */
5215 if (LOCAL_I(iloc))
5216 {
5217 nbs->search_count++;
5218 }
5219 if (nbs->print_cycles &&
5220 (!nbs->DomDec || (nbs->DomDec && !LOCAL_I(iloc))) &&
5221 nbs->search_count % 100 == 0)
5222 {
5223 nbs_cycle_print(stderr,nbs);
5224 }
5225
5226 if (debug && (CombineNBLists && nnbl > 1))
5227 {
5228 if (nbl[0]->bSimple)
5229 {
5230 print_nblist_statistics_simple(debug,nbl[0],nbs,rlist);
5231 }
5232 else
5233 {
5234 print_nblist_statistics_supersub(debug,nbl[0],nbs,rlist);
5235 }
5236 }
5237
5238 if (gmx_debug_at)
5239 {
5240 if (nbl[0]->bSimple)
5241 {
5242 print_nblist_ci_cj(debug,nbl[0]);
5243 }
5244 else
5245 {
5246 print_nblist_sci_cj(debug,nbl[0]);
5247 }
5248 }
5249 }
5250
5251 /* Initializes an nbnxn_atomdata_output_t data structure */
5252 static void nbnxn_atomdata_output_init(nbnxn_atomdata_output_t *out,
5253 int nb_kernel_type,
5254 int nenergrp,int stride,
5255 gmx_nbat_alloc_t *ma)
5256 {
5257 int cj_size;
5258
5259 out->f = NULL;
5260 ma((void **)&out->fshift,SHIFTS*DIM*sizeof(*out->fshift));
5261 out->nV = nenergrp*nenergrp;
5262 ma((void **)&out->Vvdw,out->nV*sizeof(*out->Vvdw));
5263 ma((void **)&out->Vc ,out->nV*sizeof(*out->Vc ));
5264
5265 if (nb_kernel_type == nbk4xN_X86_SIMD128 ||
5266 nb_kernel_type == nbk4xN_X86_SIMD256)
5267 {
5268 cj_size = kernel_to_cj_size(nb_kernel_type);
5269 out->nVS = nenergrp*nenergrp*stride*(cj_size>>1)*cj_size;
5270 ma((void **)&out->VSvdw,out->nVS*sizeof(*out->VSvdw));
5271 ma((void **)&out->VSc ,out->nVS*sizeof(*out->VSc ));
5272 }
5273 else
5274 {
5275 out->nVS = 0;
5276 }
5277 }
5278
5279 /* Determines the combination rule (or none) to be used, stores it,
5280 * and sets the LJ parameters required with the rule.
5281 */
5282 static void set_combination_rule_data(nbnxn_atomdata_t *nbat)
5283 {
5284 int nt,i,j;
5285 real c6,c12;
5286
5287 nt = nbat->ntype;
5288
5289 switch (nbat->comb_rule)
5290 {
5291 case ljcrGEOM:
5292 nbat->comb_rule = ljcrGEOM;
5293
5294 for(i=0; i<nt; i++)
5295 {
5296 /* Copy the diagonal from the nbfp matrix */
5297 nbat->nbfp_comb[i*2 ] = sqrt(nbat->nbfp[(i*nt+i)*2 ]);
5298 nbat->nbfp_comb[i*2+1] = sqrt(nbat->nbfp[(i*nt+i)*2+1]);
5299 }
5300 break;
5301 case ljcrLB:
5302 for(i=0; i<nt; i++)
5303 {
5304 /* Get 6*C6 and 12*C12 from the diagonal of the nbfp matrix */
5305 c6 = nbat->nbfp[(i*nt+i)*2 ];
5306 c12 = nbat->nbfp[(i*nt+i)*2+1];
5307 if (c6 > 0 && c12 > 0)
5308 {
5309 /* We store 0.5*2^1/6*sigma and sqrt(4*3*eps),
5310 * so we get 6*C6 and 12*C12 after combining.
5311 */
5312 nbat->nbfp_comb[i*2 ] = 0.5*pow(c12/c6,1.0/6.0);
5313 nbat->nbfp_comb[i*2+1] = sqrt(c6*c6/c12);
5314 }
5315 else
5316 {
5317 nbat->nbfp_comb[i*2 ] = 0;
5318 nbat->nbfp_comb[i*2+1] = 0;
5319 }
5320 }
5321 break;
5322 case ljcrNONE:
5323 /* In nbfp_s4 we use a stride of 4 for storing two parameters */
5324 nbat->alloc((void **)&nbat->nbfp_s4,nt*nt*4*sizeof(*nbat->nbfp_s4));
5325 for(i=0; i<nt; i++)
5326 {
5327 for(j=0; j<nt; j++)
5328 {
5329 nbat->nbfp_s4[(i*nt+j)*4+0] = nbat->nbfp[(i*nt+j)*2+0];
5330 nbat->nbfp_s4[(i*nt+j)*4+1] = nbat->nbfp[(i*nt+j)*2+1];
5331 nbat->nbfp_s4[(i*nt+j)*4+2] = 0;
5332 nbat->nbfp_s4[(i*nt+j)*4+3] = 0;
5333 }
5334 }
5335 break;
5336 default:
5337 gmx_incons("Unknown combination rule");
5338 break;
5339 }
5340 }
5341
5342 /* Initializes an nbnxn_atomdata_t data structure */
5343 void nbnxn_atomdata_init(FILE *fp,
5344 nbnxn_atomdata_t *nbat,
5345 int nb_kernel_type,
5346 int ntype,const real *nbfp,
5347 int n_energygroups,
5348 int nout,
5349 gmx_nbat_alloc_t *alloc,
5350 gmx_nbat_free_t *free)
5351 {
5352 int i,j;
5353 real c6,c12,tol;
5354 char *ptr;
5355 gmx_bool simple,bCombGeom,bCombLB;
5356
5357 if (alloc == NULL)
5358 {
5359 nbat->alloc = nbnxn_alloc_aligned;
5360 }
5361 else
5362 {
5363 nbat->alloc = alloc;
5364 }
5365 if (free == NULL)
5366 {
5367 nbat->free = nbnxn_free_aligned;
5368 }
5369 else
5370 {
5371 nbat->free = free;
5372 }
5373
5374 if (debug)
5375 {
5376 fprintf(debug,"There are %d atom types in the system, adding one for nbnxn_atomdata_t\n",ntype);
5377 }
5378 nbat->ntype = ntype + 1;
5379 nbat->alloc((void **)&nbat->nbfp,
5380 nbat->ntype*nbat->ntype*2*sizeof(*nbat->nbfp));
5381 nbat->alloc((void **)&nbat->nbfp_comb,nbat->ntype*2*sizeof(*nbat->nbfp_comb));
5382
5383 /* A tolerance of 1e-5 seems reasonable for (possibly hand-typed)
5384 * force-field floating point parameters.
5385 */
5386 tol = 1e-5;
5387 ptr = getenv("GMX_LJCOMB_TOL");
5388 if (ptr != NULL)
5389 {
5390 double dbl;
5391
5392 sscanf(ptr,"%lf",&dbl);
5393 tol = dbl;
5394 }
5395 bCombGeom = TRUE;
5396 bCombLB = TRUE;
5397
5398 /* Temporarily fill nbat->nbfp_comb with sigma and epsilon
5399 * to check for the LB rule.
5400 */
5401 for(i=0; i<ntype; i++)
5402 {
5403 c6 = nbfp[(i*ntype+i)*2 ];
5404 c12 = nbfp[(i*ntype+i)*2+1];
5405 if (c6 > 0 && c12 > 0)
5406 {
5407 nbat->nbfp_comb[i*2 ] = pow(c12/c6,1.0/6.0);
5408 nbat->nbfp_comb[i*2+1] = 0.25*c6*c6/c12;
5409 }
5410 else if (c6 == 0 && c12 == 0)
5411 {
5412 nbat->nbfp_comb[i*2 ] = 0;
5413 nbat->nbfp_comb[i*2+1] = 0;
5414 }
5415 else
5416 {
5417 /* Can not use LB rule with only dispersion or repulsion */
5418 bCombLB = FALSE;
5419 }
5420 }
5421
5422 for(i=0; i<nbat->ntype; i++)
5423 {
5424 for(j=0; j<nbat->ntype; j++)
5425 {
5426 if (i < ntype && j < ntype)
5427 {
5428 /* We store the prefactor in the derivative of the potential
5429 * in the parameter to avoid multiplications in the inner loop.
5430 */
5431 c6 = nbfp[(i*ntype+j)*2 ];
5432 c12 = nbfp[(i*ntype+j)*2+1];
5433 nbat->nbfp[(i*nbat->ntype+j)*2 ] = 6.0*c6;
5434 nbat->nbfp[(i*nbat->ntype+j)*2+1] = 12.0*c12;
5435
5436 bCombGeom = bCombGeom &&
5437 gmx_within_tol(c6*c6 ,nbfp[(i*ntype+i)*2 ]*nbfp[(j*ntype+j)*2 ],tol) &&
5438 gmx_within_tol(c12*c12,nbfp[(i*ntype+i)*2+1]*nbfp[(j*ntype+j)*2+1],tol);
5439
5440 bCombLB = bCombLB &&
5441 ((c6 == 0 && c12 == 0 &&
5442 (nbat->nbfp_comb[i*2+1] == 0 || nbat->nbfp_comb[j*2+1] == 0)) ||
5443 (c6 > 0 && c12 > 0 &&
5444 gmx_within_tol(pow(c12/c6,1.0/6.0),0.5*(nbat->nbfp_comb[i*2]+nbat->nbfp_comb[j*2]),tol) &&
5445 gmx_within_tol(0.25*c6*c6/c12,sqrt(nbat->nbfp_comb[i*2+1]*nbat->nbfp_comb[j*2+1]),tol)));
5446 }
5447 else
5448 {
5449 /* Add zero parameters for the additional dummy atom type */
5450 nbat->nbfp[(i*nbat->ntype+j)*2 ] = 0;
5451 nbat->nbfp[(i*nbat->ntype+j)*2+1] = 0;
5452 }
5453 }
5454 }
5455 if (debug)
5456 {
5457 fprintf(debug,"Combination rules: geometric %d Lorentz-Berthelot %d\n",
5458 bCombGeom,bCombLB);
5459 }
5460
5461 simple = nbnxn_kernel_pairlist_simple(nb_kernel_type);
5462
5463 if (simple)
5464 {
5465 /* We prefer the geometic combination rule,
5466 * as that gives a slightly faster kernel than the LB rule.
5467 */
5468 if (bCombGeom)
5469 {
5470 nbat->comb_rule = ljcrGEOM;
5471 }
5472 else if (bCombLB)
5473 {
5474 nbat->comb_rule = ljcrLB;
5475 }
5476 else
5477 {
5478 nbat->comb_rule = ljcrNONE;
5479
5480 nbat->free(nbat->nbfp_comb);
5481 }
5482
5483 if (fp)
5484 {
5485 if (nbat->comb_rule == ljcrNONE)
5486 {
5487 fprintf(fp,"Using full Lennard-Jones parameter combination matrix\n\n");
5488 }
5489 else
5490 {
5491 fprintf(fp,"Using %s Lennard-Jones combination rule\n\n",
5492 nbat->comb_rule==ljcrGEOM ? "geometric" : "Lorentz-Berthelot");
5493 }
5494 }
5495
5496 set_combination_rule_data(nbat);
5497 }
5498 else
5499 {
5500 nbat->comb_rule = ljcrNONE;
5501
5502 nbat->free(nbat->nbfp_comb);
5503 }
5504
5505 nbat->natoms = 0;
5506 nbat->type = NULL;
5507 nbat->lj_comb = NULL;
5508 if (simple)
5509 {
5510 switch (nb_kernel_type)
5511 {
5512 case nbk4xN_X86_SIMD128:
5513 nbat->XFormat = nbatX4;
5514 break;
5515 case nbk4xN_X86_SIMD256:
5516 #ifndef GMX_DOUBLE
5517 nbat->XFormat = nbatX8;
5518 #else
5519 nbat->XFormat = nbatX4;
5520 #endif
5521 break;
5522 default:
5523 nbat->XFormat = nbatXYZ;
5524 break;
5525 }
5526
5527 nbat->FFormat = nbat->XFormat;
5528 }
5529 else
5530 {
5531 nbat->XFormat = nbatXYZQ;
5532 nbat->FFormat = nbatXYZ;
5533 }
5534 nbat->q = NULL;
5535 nbat->nenergrp = n_energygroups;
5536 if (!simple)
5537 {
5538 /* Energy groups not supported yet for super-sub lists */
5539 nbat->nenergrp = 1;
5540 }
5541 /* Temporary storage goes is #grp^3*8 real, so limit to 64 */
5542 if (nbat->nenergrp > 64)
5543 {
5544 gmx_fatal(FARGS,"With NxN kernels not more than 64 energy groups are supported\n");
5545 }
5546 nbat->neg_2log = 1;
5547 while (nbat->nenergrp > (1<<nbat->neg_2log))
5548 {
5549 nbat->neg_2log++;
5550 }
5551 nbat->energrp = NULL;
5552 nbat->alloc((void **)&nbat->shift_vec,SHIFTS*sizeof(*nbat->shift_vec));
5553 nbat->xstride = (nbat->XFormat == nbatXYZQ ? STRIDE_XYZQ : DIM);
5554 nbat->fstride = (nbat->FFormat == nbatXYZQ ? STRIDE_XYZQ : DIM);
5555 nbat->x = NULL;
5556 nbat->nout = nout;
5557 snew(nbat->out,nbat->nout);
5558 nbat->nalloc = 0;
5559 for(i=0; i<nbat->nout; i++)
5560 {
5561 nbnxn_atomdata_output_init(&nbat->out[i],
5562 nb_kernel_type,
5563 nbat->nenergrp,1<<nbat->neg_2log,
5564 nbat->alloc);
5565 }
5566 }
5567
5568 static void copy_lj_to_nbat_lj_comb_x4(const real *ljparam_type,
5569 const int *type,int na,
5570 real *ljparam_at)
5571 {
5572 int is,k,i;
5573
5574 /* The LJ params follow the combination rule:
5575 * copy the params for the type array to the atom array.
5576 */
5577 for(is=0; is<na; is+=PACK_X4)
5578 {
5579 for(k=0; k<PACK_X4; k++)
5580 {
5581 i = is + k;
5582 ljparam_at[is*2 +k] = ljparam_type[type[i]*2 ];
5583 ljparam_at[is*2+PACK_X4+k] = ljparam_type[type[i]*2+1];
5584 }
5585 }
5586 }
5587
5588 static void copy_lj_to_nbat_lj_comb_x8(const real *ljparam_type,
5589 const int *type,int na,
5590 real *ljparam_at)
5591 {
5592 int is,k,i;
5593
5594 /* The LJ params follow the combination rule:
5595 * copy the params for the type array to the atom array.
5596 */
5597 for(is=0; is<na; is+=PACK_X8)
5598 {
5599 for(k=0; k<PACK_X8; k++)
5600 {
5601 i = is + k;
5602 ljparam_at[is*2 +k] = ljparam_type[type[i]*2 ];
5603 ljparam_at[is*2+PACK_X8+k] = ljparam_type[type[i]*2+1];
5604 }
5605 }
5606 }
5607
5608 /* Sets the atom type and LJ data in nbnxn_atomdata_t */
5609 static void nbnxn_atomdata_set_atomtypes(nbnxn_atomdata_t *nbat,
5610 int ngrid,
5611 const nbnxn_search_t nbs,
5612 const int *type)
5613 {
5614 int g,i,ncz,ash;
5615 const nbnxn_grid_t *grid;
5616
5617 for(g=0; g<ngrid; g++)
5618 {
5619 grid = &nbs->grid[g];
5620
5621 /* Loop over all columns and copy and fill */
5622 for(i=0; i<grid->ncx*grid->ncy; i++)
5623 {
5624 ncz = grid->cxy_ind[i+1] - grid->cxy_ind[i];
5625 ash = (grid->cell0 + grid->cxy_ind[i])*grid->na_sc;
5626
5627 copy_int_to_nbat_int(nbs->a+ash,grid->cxy_na[i],ncz*grid->na_sc,
5628 type,nbat->ntype-1,nbat->type+ash);
5629
5630 if (nbat->comb_rule != ljcrNONE)
5631 {
5632 if (nbat->XFormat == nbatX4)
5633 {
5634 copy_lj_to_nbat_lj_comb_x4(nbat->nbfp_comb,
5635 nbat->type+ash,ncz*grid->na_sc,
5636 nbat->lj_comb+ash*2);
5637 }
5638 else if (nbat->XFormat == nbatX8)
5639 {
5640 copy_lj_to_nbat_lj_comb_x8(nbat->nbfp_comb,
5641 nbat->type+ash,ncz*grid->na_sc,
5642 nbat->lj_comb+ash*2);
5643 }
5644 }
5645 }
5646 }
5647 }
5648
5649 /* Sets the charges in nbnxn_atomdata_t *nbat */
5650 static void nbnxn_atomdata_set_charges(nbnxn_atomdata_t *nbat,
5651 int ngrid,
5652 const nbnxn_search_t nbs,
5653 const real *charge)
5654 {
5655 int g,cxy,ncz,ash,na,na_round,i,j;
5656 real *q;
5657 const nbnxn_grid_t *grid;
5658
5659 for(g=0; g<ngrid; g++)
5660 {
5661 grid = &nbs->grid[g];
5662
5663 /* Loop over all columns and copy and fill */
5664 for(cxy=0; cxy<grid->ncx*grid->ncy; cxy++)
5665 {
5666 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
5667 na = grid->cxy_na[cxy];
5668 na_round = (grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy])*grid->na_sc;
5669
5670 if (nbat->XFormat == nbatXYZQ)
5671 {
5672 q = nbat->x + ash*STRIDE_XYZQ + ZZ + 1;
5673 for(i=0; i<na; i++)
5674 {
5675 *q = charge[nbs->a[ash+i]];
5676 q += STRIDE_XYZQ;
5677 }
5678 /* Complete the partially filled last cell with zeros */
5679 for(; i<na_round; i++)
5680 {
5681 *q = 0;
5682 q += STRIDE_XYZQ;
5683 }
5684 }
5685 else
5686 {
5687 q = nbat->q + ash;
5688 for(i=0; i<na; i++)
5689 {
5690 *q = charge[nbs->a[ash+i]];
5691 q++;
5692 }
5693 /* Complete the partially filled last cell with zeros */
5694 for(; i<na_round; i++)
5695 {
5696 *q = 0;
5697 q++;
5698 }
5699 }
5700 }
5701 }
5702 }
5703
5704 /* Copies the energy group indices to a reordered and packed array */
5705 static void copy_egp_to_nbat_egps(const int *a,int na,int na_round,
5706 int na_c,int bit_shift,
5707 const int *in,int *innb)
5708 {
5709 int i,j,sa,at;
5710 int comb;
5711
5712 j = 0;
5713 for(i=0; i<na; i+=na_c)
5714 {
5715 /* Store na_c energy groups number into one int */
5716 comb = 0;
5717 for(sa=0; sa<na_c; sa++)
5718 {
5719 at = a[i+sa];
5720 if (at >= 0)
5721 {
5722 comb |= (GET_CGINFO_GID(in[at]) << (sa*bit_shift));
5723 }
5724 }
5725 innb[j++] = comb;
5726 }
5727 /* Complete the partially filled last cell with fill */
5728 for(; i<na_round; i+=na_c)
5729 {
5730 innb[j++] = 0;
5731 }
5732 }
5733
5734 /* Set the energy group indices for atoms in nbnxn_atomdata_t */
5735 static void nbnxn_atomdata_set_energygroups(nbnxn_atomdata_t *nbat,
5736 int ngrid,
5737 const nbnxn_search_t nbs,
5738 const int *atinfo)
5739 {
5740 int g,i,ncz,ash;
5741 const nbnxn_grid_t *grid;
5742
5743 for(g=0; g<ngrid; g++)
5744 {
5745 grid = &nbs->grid[g];
5746
5747 /* Loop over all columns and copy and fill */
5748 for(i=0; i<grid->ncx*grid->ncy; i++)
5749 {
5750 ncz = grid->cxy_ind[i+1] - grid->cxy_ind[i];
5751 ash = (grid->cell0 + grid->cxy_ind[i])*grid->na_sc;
5752
5753 copy_egp_to_nbat_egps(nbs->a+ash,grid->cxy_na[i],ncz*grid->na_sc,
5754 nbat->na_c,nbat->neg_2log,
5755 atinfo,nbat->energrp+(ash>>grid->na_c_2log));
5756 }
5757 }
5758 }
5759
5760 /* Sets all required atom parameter data in nbnxn_atomdata_t */
5761 void nbnxn_atomdata_set(nbnxn_atomdata_t *nbat,
5762 int locality,
5763 const nbnxn_search_t nbs,
5764 const t_mdatoms *mdatoms,
5765 const int *atinfo)
5766 {
5767 int ngrid;
5768
5769 if (locality == eatLocal)
5770 {
5771 ngrid = 1;
5772 }
5773 else
5774 {
5775 ngrid = nbs->ngrid;
5776 }
5777
5778 nbnxn_atomdata_set_atomtypes(nbat,ngrid,nbs,mdatoms->typeA);
5779
5780 nbnxn_atomdata_set_charges(nbat,ngrid,nbs,mdatoms->chargeA);
5781
5782 if (nbat->nenergrp > 1)
5783 {
5784 nbnxn_atomdata_set_energygroups(nbat,ngrid,nbs,atinfo);
5785 }
5786 }
5787
5788 /* Copies the shift vector array to nbnxn_atomdata_t */
5789 void nbnxn_atomdata_copy_shiftvec(gmx_bool bDynamicBox,
5790 rvec *shift_vec,
5791 nbnxn_atomdata_t *nbat)
5792 {
5793 int i;
5794
5795 nbat->bDynamicBox = bDynamicBox;
5796 for(i=0; i<SHIFTS; i++)
5797 {
5798 copy_rvec(shift_vec[i],nbat->shift_vec[i]);
5799 }
5800 }
5801
5802 /* Copies (and reorders) the coordinates to nbnxn_atomdata_t */
5803 void nbnxn_atomdata_copy_x_to_nbat_x(const nbnxn_search_t nbs,
5804 int locality,
5805 gmx_bool FillLocal,
5806 rvec *x,
5807 nbnxn_atomdata_t *nbat)
5808 {
5809 int g0=0,g1=0;
5810 int nth,th;
5811
5812 switch (locality)
5813 {
5814 case eatAll:
5815 g0 = 0;
5816 g1 = nbs->ngrid;
5817 break;
5818 case eatLocal:
5819 g0 = 0;
5820 g1 = 1;
5821 break;
5822 case eatNonlocal:
5823 g0 = 1;
5824 g1 = nbs->ngrid;
5825 break;
5826 }
5827
5828 if (FillLocal)
5829 {
5830 nbat->natoms_local = nbs->grid[0].nc*nbs->grid[0].na_sc;
5831 }
5832
5833 nth = gmx_omp_nthreads_get(emntPairsearch);
5834
5835 #pragma omp parallel for num_threads(nth) schedule(static)
5836 for(th=0; th<nth; th++)
5837 {
5838 int g;
5839
5840 for(g=g0; g<g1; g++)
5841 {
5842 const nbnxn_grid_t *grid;
5843 int cxy0,cxy1,cxy;
5844
5845 grid = &nbs->grid[g];
5846
5847 cxy0 = (grid->ncx*grid->ncy* th +nth-1)/nth;
5848 cxy1 = (grid->ncx*grid->ncy*(th+1)+nth-1)/nth;
5849
5850 for(cxy=cxy0; cxy<cxy1; cxy++)
5851 {
5852 int na,ash,na_fill;
5853
5854 na = grid->cxy_na[cxy];
5855 ash = (grid->cell0 + grid->cxy_ind[cxy])*grid->na_sc;
5856
5857 if (g == 0 && FillLocal)
5858 {
5859 na_fill =
5860 (grid->cxy_ind[cxy+1] - grid->cxy_ind[cxy])*grid->na_sc;
5861 }
5862 else
5863 {
5864 /* We fill only the real particle locations.
5865 * We assume the filling entries at the end have been
5866 * properly set before during ns.
5867 */
5868 na_fill = na;
5869 }
5870 copy_rvec_to_nbat_real(nbs->a+ash,na,na_fill,x,
5871 nbat->XFormat,nbat->x,ash,
5872 0,0,0);
5873 }
5874 }
5875 }
5876 }
5877
5878 /* Add part of the force array(s) from nbnxn_atomdata_t to f */
5879 static void
5880 nbnxn_atomdata_add_nbat_f_to_f_part(const nbnxn_search_t nbs,
5881 const nbnxn_atomdata_t *nbat,
5882 nbnxn_atomdata_output_t *out,
5883 int nfa,
5884 int a0,int a1,
5885 rvec *f)
5886 {
5887 int a,i,fa;
5888 const int *cell;
5889 const real *fnb;
5890
5891 cell = nbs->cell;
5892
5893 /* Loop over all columns and copy and fill */
5894 switch (nbat->FFormat)
5895 {
5896 case nbatXYZ:
5897 case nbatXYZQ:
5898 if (nfa == 1)
5899 {
5900 fnb = out[0].f;
5901
5902 for(a=a0; a<a1; a++)
5903 {
5904 i = cell[a]*nbat->fstride;
5905
5906 f[a][XX] += fnb[i];
5907 f[a][YY] += fnb[i+1];
5908 f[a][ZZ] += fnb[i+2];
5909 }
5910 }
5911 else
5912 {
5913 for(a=a0; a<a1; a++)
5914 {
5915 i = cell[a]*nbat->fstride;
5916
5917 for(fa=0; fa<nfa; fa++)
5918 {
5919 f[a][XX] += out[fa].f[i];
5920 f[a][YY] += out[fa].f[i+1];
5921 f[a][ZZ] += out[fa].f[i+2];
5922 }
5923 }
5924 }
5925 break;
5926 case nbatX4:
5927 if (nfa == 1)
5928 {
5929 fnb = out[0].f;
5930
5931 for(a=a0; a<a1; a++)
5932 {
5933 i = X4_IND_A(cell[a]);
5934
5935 f[a][XX] += fnb[i+XX*PACK_X4];
5936 f[a][YY] += fnb[i+YY*PACK_X4];
5937 f[a][ZZ] += fnb[i+ZZ*PACK_X4];
5938 }
5939 }
5940 else
5941 {
5942 for(a=a0; a<a1; a++)
5943 {
5944 i = X4_IND_A(cell[a]);
5945
5946 for(fa=0; fa<nfa; fa++)
5947 {
5948 f[a][XX] += out[fa].f[i+XX*PACK_X4];
5949 f[a][YY] += out[fa].f[i+YY*PACK_X4];
5950 f[a][ZZ] += out[fa].f[i+ZZ*PACK_X4];
5951 }
5952 }
5953 }
5954 break;
5955 case nbatX8:
5956 if (nfa == 1)
5957 {
5958 fnb = out[0].f;
5959
5960 for(a=a0; a<a1; a++)
5961 {
5962 i = X8_IND_A(cell[a]);
5963
5964 f[a][XX] += fnb[i+XX*PACK_X8];
5965 f[a][YY] += fnb[i+YY*PACK_X8];
5966 f[a][ZZ] += fnb[i+ZZ*PACK_X8];
5967 }
5968 }
5969 else
5970 {
5971 for(a=a0; a<a1; a++)
5972 {
5973 i = X8_IND_A(cell[a]);
5974
5975 for(fa=0; fa<nfa; fa++)
5976 {
5977 f[a][XX] += out[fa].f[i+XX*PACK_X8];
5978 f[a][YY] += out[fa].f[i+YY*PACK_X8];
5979 f[a][ZZ] += out[fa].f[i+ZZ*PACK_X8];
5980 }
5981 }
5982 }
5983 break;
5984 }
5985 }
5986
5987 /* Add the force array(s) from nbnxn_atomdata_t to f */
5988 void nbnxn_atomdata_add_nbat_f_to_f(const nbnxn_search_t nbs,
5989 int locality,
5990 const nbnxn_atomdata_t *nbat,
5991 rvec *f)
5992 {
5993 int a0=0,na=0;
5994 int nth,th;
5995
5996 nbs_cycle_start(&nbs->cc[enbsCCreducef]);
5997
5998 switch (locality)
5999 {
6000 case eatAll:
6001 a0 = 0;
6002 na = nbs->natoms_nonlocal;
6003 break;
6004 case eatLocal:
6005 a0 = 0;
6006 na = nbs->natoms_local;
6007 break;
6008 case eatNonlocal:
6009 a0 = nbs->natoms_local;
6010 na = nbs->natoms_nonlocal - nbs->natoms_local;
6011 break;
6012 }
6013
6014 nth = gmx_omp_nthreads_get(emntNonbonded);
6015 #pragma omp parallel for num_threads(nth) schedule(static)
6016 for(th=0; th<nth; th++)
6017 {
6018 nbnxn_atomdata_add_nbat_f_to_f_part(nbs,nbat,
6019 nbat->out,
6020 nbat->nout,
6021 a0+((th+0)*na)/nth,
6022 a0+((th+1)*na)/nth,
6023 f);
6024 }
6025
6026 nbs_cycle_stop(&nbs->cc[enbsCCreducef]);
6027 }
6028
6029 /* Adds the shift forces from nbnxn_atomdata_t to fshift */
6030 void nbnxn_atomdata_add_nbat_fshift_to_fshift(const nbnxn_atomdata_t *nbat,
6031 rvec *fshift)
6032 {
6033 const nbnxn_atomdata_output_t *out;
6034 int th;
6035 int s;
6036 rvec sum;
6037
6038 out = nbat->out;
6039
6040 for(s=0; s<SHIFTS; s++)
6041 {
6042 clear_rvec(sum);
6043 for(th=0; th<nbat->nout; th++)
6044 {
6045 sum[XX] += out[th].fshift[s*DIM+XX];
6046 sum[YY] += out[th].fshift[s*DIM+YY];
6047 sum[ZZ] += out[th].fshift[s*DIM+ZZ];
6048 }
6049 rvec_inc(fshift[s],sum);
6050 }
6051 }
The ultimate molecular dynamics simulation package