1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustr
4 * This source code is part of
8 * GROningen MAchine for Chemical Simulations
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.
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.
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.
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.
30 * For more info, check our website at http://www.gromacs.org
33 * Gallium Rubidium Oxygen Manganese Argon Carbon Silicon
36 #ifndef _nbnxn_internal_h
37 #define _nsnxn_internal_h
41 #include "gmx_cyclecounter.h"
49 #define NBNXN_SEARCH_SSE
53 /* A pair-search grid struct for one domain decomposition zone */
55 rvec c0; /* The lower corner of the (local) grid */
56 rvec c1; /* The upper corner of the (local) grid */
57 real atom_density; /* The atom number density for the local grid */
59 gmx_bool bSimple; /* Is this grid simple or super/sub */
60 int na_c; /* Number of atoms per cluster */
61 int na_cj; /* Number of atoms for list j-clusters */
62 int na_sc; /* Number of atoms per super-cluster */
63 int na_c_2log; /* 2log of na_c */
65 int ncx; /* Number of (super-)cells along x */
66 int ncy; /* Number of (super-)cells along y */
67 int nc; /* Total number of (super-)cells */
69 real sx; /* x-size of a (super-)cell */
70 real sy; /* y-size of a (super-)cell */
71 real inv_sx; /* 1/sx */
72 real inv_sy; /* 1/sy */
74 int cell0; /* Index in nbs->cell corresponding to cell 0 */
76 int *cxy_na; /* The number of atoms for each column in x,y */
77 int *cxy_ind; /* Grid (super)cell index, offset from cell0 */
78 int cxy_nalloc; /* Allocation size for cxy_na and cxy_ind */
80 int *nsubc; /* The number of sub cells for each super cell */
81 float *bbcz; /* Bounding boxes in z for the super cells */
82 float *bb; /* 3D bounding boxes for the sub cells */
83 float *bbj; /* 3D j-b.boxes for SSE-double or AVX-single */
84 int *flags; /* Flag for the super cells */
85 int nc_nalloc; /* Allocation size for the pointers above */
87 float *bbcz_simple; /* bbcz for simple grid converted from super */
88 float *bb_simple; /* bb for simple grid converted from super */
89 int *flags_simple; /* flags for simple grid converted from super */
90 int nc_nalloc_simple; /* Allocation size for the pointers above */
92 int nsubc_tot; /* Total number of subcell, used for printing */
95 #ifdef NBNXN_SEARCH_SSE
96 #define GMX_MM128_HERE
97 #include "gmx_x86_simd_macros.h"
98 typedef struct nbnxn_x_ci_x86_simd128 {
99 /* The i-cluster coordinates for simple search */
100 gmx_mm_pr ix_SSE0,iy_SSE0,iz_SSE0;
101 gmx_mm_pr ix_SSE1,iy_SSE1,iz_SSE1;
102 gmx_mm_pr ix_SSE2,iy_SSE2,iz_SSE2;
103 gmx_mm_pr ix_SSE3,iy_SSE3,iz_SSE3;
104 } nbnxn_x_ci_x86_simd128_t;
105 #undef GMX_MM128_HERE
106 #ifdef GMX_X86_AVX_256
107 #define GMX_MM256_HERE
108 #include "gmx_x86_simd_macros.h"
109 typedef struct nbnxn_x_ci_x86_simd256 {
110 /* The i-cluster coordinates for simple search */
111 gmx_mm_pr ix_SSE0,iy_SSE0,iz_SSE0;
112 gmx_mm_pr ix_SSE1,iy_SSE1,iz_SSE1;
113 gmx_mm_pr ix_SSE2,iy_SSE2,iz_SSE2;
114 gmx_mm_pr ix_SSE3,iy_SSE3,iz_SSE3;
115 } nbnxn_x_ci_x86_simd256_t;
116 #undef GMX_MM256_HERE
120 /* Working data for the actual i-supercell during pair search */
121 typedef struct nbnxn_list_work {
122 gmx_cache_protect_t cp0; /* Protect cache between threads */
124 float *bb_ci; /* The bounding boxes, pbc shifted, for each cluster */
125 real *x_ci; /* The coordinates, pbc shifted, for each atom */
126 #ifdef NBNXN_SEARCH_SSE
127 nbnxn_x_ci_x86_simd128_t *x_ci_x86_simd128;
128 #ifdef GMX_X86_AVX_256
129 nbnxn_x_ci_x86_simd256_t *x_ci_x86_simd256;
132 int cj_ind; /* The current cj_ind index for the current list */
133 int cj4_init; /* The first unitialized cj4 block */
135 float *d2; /* Bounding box distance work array */
137 nbnxn_cj_t *cj; /* The j-cell list */
138 int cj_nalloc; /* Allocation size of cj */
140 int ncj_noq; /* Nr. of cluster pairs without Coul for flop count */
141 int ncj_hlj; /* Nr. of cluster pairs with 1/2 LJ for flop count */
143 gmx_cache_protect_t cp1; /* Protect cache between threads */
146 /* Function type for setting the i-atom coordinate working data */
148 gmx_icell_set_x_t(int ci,
149 real shx,real shy,real shz,
151 int stride,const real *x,
152 nbnxn_list_work_t *work);
154 static gmx_icell_set_x_t icell_set_x_simple;
155 #ifdef NBNXN_SEARCH_SSE
156 static gmx_icell_set_x_t icell_set_x_simple_x86_simd128;
157 #ifdef GMX_X86_AVX_256
158 static gmx_icell_set_x_t icell_set_x_simple_x86_simd256;
161 static gmx_icell_set_x_t icell_set_x_supersub;
162 #ifdef NBNXN_SEARCH_SSE
163 static gmx_icell_set_x_t icell_set_x_supersub_sse8;
166 /* Local cycle count struct for profiling */
173 /* Local cycle count enum for profiling */
174 enum { enbsCCgrid, enbsCCsearch, enbsCCcombine, enbsCCreducef, enbsCCnr };
176 /* Thread-local work struct, contains part of nbnxn_grid_t */
178 gmx_cache_protect_t cp0;
184 int sort_work_nalloc;
186 int ndistc; /* Number of distance checks for flop counting */
188 nbnxn_cycle_t cc[enbsCCnr];
190 gmx_cache_protect_t cp1;
191 } nbnxn_search_work_t;
193 /* Main pair-search struct, contains the grid(s), not the pair-list(s) */
194 typedef struct nbnxn_search {
195 int ePBC; /* PBC type enum */
196 matrix box; /* The periodic unit-cell */
198 gmx_bool DomDec; /* Are we doing domain decomposition? */
199 ivec dd_dim; /* Are we doing DD in x,y,z? */
200 gmx_domdec_zones_t *zones; /* The domain decomposition zones */
202 int ngrid; /* The number of grids, equal to #DD-zones */
203 nbnxn_grid_t *grid; /* Array of grids, size ngrid */
204 int *cell; /* Actual allocated cell array for all grids */
205 int cell_nalloc; /* Allocation size of cell */
206 int *a; /* Atom index for grid, the inverse of cell */
207 int a_nalloc; /* Allocation size of a */
209 int natoms_local; /* The local atoms run from 0 to natoms_local */
210 int natoms_nonlocal; /* The non-local atoms run from natoms_local
211 * to natoms_nonlocal */
213 gmx_bool print_cycles;
215 nbnxn_cycle_t cc[enbsCCnr];
217 gmx_icell_set_x_t *icell_set_x; /* Function for setting i-coords */
219 int nthread_max; /* Maximum number of threads for pair-search */
220 nbnxn_search_work_t *work; /* Work array, size nthread_max */
224 static void nbs_cycle_start(nbnxn_cycle_t *cc)
226 cc->start = gmx_cycles_read();
229 static void nbs_cycle_stop(nbnxn_cycle_t *cc)
231 cc->c += gmx_cycles_read() - cc->start;