src/gromacs/gmxlib/nonbonded/nonbonded.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013,2014,2015, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #include "gmxpre.h"
  38
  39 #include "nonbonded.h"
  40
  41 #include "config.h"
  42
  43 #include <cassert>
  44 #include <cstdio>
  45 #include <cstdlib>
  46
  47 #include "thread_mpi/threads.h"
  48
  49 #include "gromacs/fileio/txtdump.h"
  50 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  51 #include "gromacs/gmxlib/nonbonded/nb_generic.h"
  52 #include "gromacs/gmxlib/nonbonded/nb_generic_cg.h"
  53 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  54 #include "gromacs/legacyheaders/names.h"
  55 #include "gromacs/legacyheaders/nrnb.h"
  56 #include "gromacs/legacyheaders/types/forcerec.h"
  57 #include "gromacs/legacyheaders/types/nblist.h"
  58 #include "gromacs/legacyheaders/types/nrnb.h"
  59 #include "gromacs/listed-forces/bonded.h"
  60 #include "gromacs/math/utilities.h"
  61 #include "gromacs/math/vec.h"
  62 #include "gromacs/mdtypes/mdatom.h"
  63 #include "gromacs/pbcutil/ishift.h"
  64 #include "gromacs/pbcutil/mshift.h"
  65 #include "gromacs/pbcutil/pbc.h"
  66 #include "gromacs/simd/simd.h"
  67 #include "gromacs/tables/forcetable.h"
  68 #include "gromacs/utility/arraysize.h"
  69 #include "gromacs/utility/basedefinitions.h"
  70 #include "gromacs/utility/cstringutil.h"
  71 #include "gromacs/utility/fatalerror.h"
  72 #include "gromacs/utility/smalloc.h"
  73
  74 /* Different default (c) and SIMD instructions interaction-specific kernels */
  75 #include "gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_c.h"
  76
  77 #if GMX_SIMD_X86_SSE2 && !(defined GMX_DOUBLE)
  78 #    include "gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_sse2_single.h"
  79 #endif
  80 #if GMX_SIMD_X86_SSE4_1 && !(defined GMX_DOUBLE)
  81 #    include "gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_sse4_1_single.h"
  82 #endif
  83 #if GMX_SIMD_X86_AVX_128_FMA && !(defined GMX_DOUBLE)
  84 #    include "gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_avx_128_fma_single.h"
  85 #endif
  86 #if GMX_SIMD_X86_AVX_256_OR_HIGHER && !(defined GMX_DOUBLE)
  87 #    include "gromacs/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_avx_256_single.h"
  88 #endif
  89 #if GMX_SIMD_X86_SSE2 && defined GMX_DOUBLE
  90 #    include "gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_sse2_double.h"
  91 #endif
  92 #if GMX_SIMD_X86_SSE4_1 && defined GMX_DOUBLE
  93 #    include "gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_sse4_1_double.h"
  94 #endif
  95 #if GMX_SIMD_X86_AVX_128_FMA && defined GMX_DOUBLE
  96 #    include "gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_avx_128_fma_double.h"
  97 #endif
  98 #if GMX_SIMD_X86_AVX_256_OR_HIGHER && defined GMX_DOUBLE
  99 #    include "gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_avx_256_double.h"
 100 #endif
 101 #if GMX_SIMD_SPARC64_HPC_ACE && defined GMX_DOUBLE
 102 #    include "gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_sparc64_hpc_ace_double.h"
 103 #endif
 104
 105
 106 static tMPI_Thread_mutex_t nonbonded_setup_mutex = TMPI_THREAD_MUTEX_INITIALIZER;
 107 static gmx_bool            nonbonded_setup_done  = FALSE;
 108
 109
 110 void
 111 gmx_nonbonded_setup(t_forcerec *   fr,
 112                     gmx_bool       bGenericKernelOnly)
 113 {
 114     tMPI_Thread_mutex_lock(&nonbonded_setup_mutex);
 115     /* Here we are guaranteed only one thread made it. */
 116     if (nonbonded_setup_done == FALSE)
 117     {
 118         if (bGenericKernelOnly == FALSE)
 119         {
 120             /* Add the generic kernels to the structure stored statically in nb_kernel.c */
 121             nb_kernel_list_add_kernels(kernellist_c, kernellist_c_size);
 122
 123             if (!(fr != NULL && fr->use_simd_kernels == FALSE))
 124             {
 125                 /* Add interaction-specific kernels for different architectures */
 126                 /* Single precision */
 127 #if GMX_SIMD_X86_SSE2 && !(defined GMX_DOUBLE)
 128                 nb_kernel_list_add_kernels(kernellist_sse2_single, kernellist_sse2_single_size);
 129 #endif
 130 #if GMX_SIMD_X86_SSE4_1 && !(defined GMX_DOUBLE)
 131                 nb_kernel_list_add_kernels(kernellist_sse4_1_single, kernellist_sse4_1_single_size);
 132 #endif
 133 #if GMX_SIMD_X86_AVX_128_FMA && !(defined GMX_DOUBLE)
 134                 nb_kernel_list_add_kernels(kernellist_avx_128_fma_single, kernellist_avx_128_fma_single_size);
 135 #endif
 136 #if GMX_SIMD_X86_AVX_256_OR_HIGHER && !(defined GMX_DOUBLE)
 137                 nb_kernel_list_add_kernels(kernellist_avx_256_single, kernellist_avx_256_single_size);
 138 #endif
 139                 /* Double precision */
 140 #if GMX_SIMD_X86_SSE2 && defined GMX_DOUBLE
 141                 nb_kernel_list_add_kernels(kernellist_sse2_double, kernellist_sse2_double_size);
 142 #endif
 143 #if GMX_SIMD_X86_SSE4_1 && defined GMX_DOUBLE
 144                 nb_kernel_list_add_kernels(kernellist_sse4_1_double, kernellist_sse4_1_double_size);
 145 #endif
 146 #if GMX_SIMD_X86_AVX_128_FMA && defined GMX_DOUBLE
 147                 nb_kernel_list_add_kernels(kernellist_avx_128_fma_double, kernellist_avx_128_fma_double_size);
 148 #endif
 149 #if GMX_SIMD_X86_AVX_256_OR_HIGHER && defined GMX_DOUBLE
 150                 nb_kernel_list_add_kernels(kernellist_avx_256_double, kernellist_avx_256_double_size);
 151 #endif
 152 #if GMX_SIMD_SPARC64_HPC_ACE && defined GMX_DOUBLE
 153                 nb_kernel_list_add_kernels(kernellist_sparc64_hpc_ace_double, kernellist_sparc64_hpc_ace_double_size);
 154 #endif
 155                 ; /* empty statement to avoid a completely empty block */
 156             }
 157         }
 158         /* Create a hash for faster lookups */
 159         nb_kernel_list_hash_init();
 160
 161         nonbonded_setup_done = TRUE;
 162     }
 163     tMPI_Thread_mutex_unlock(&nonbonded_setup_mutex);
 164 }
 165
 166
 167
 168 void
 169 gmx_nonbonded_set_kernel_pointers(FILE *log, t_nblist *nl, gmx_bool bElecAndVdwSwitchDiffers)
 170 {
 171     const char *     elec;
 172     const char *     elec_mod;
 173     const char *     vdw;
 174     const char *     vdw_mod;
 175     const char *     geom;
 176     const char *     other;
 177
 178     struct
 179     {
 180         const char *  arch;
 181         int           simd_padding_width;
 182     }
 183     arch_and_padding[] =
 184     {
 185         /* Single precision */
 186 #if GMX_SIMD_X86_AVX_256_OR_HIGHER && !(defined GMX_DOUBLE)
 187         { "avx_256_single", 8 },
 188 #endif
 189 #if GMX_SIMD_X86_AVX_128_FMA && !(defined GMX_DOUBLE)
 190         { "avx_128_fma_single", 4 },
 191 #endif
 192 #if GMX_SIMD_X86_SSE4_1 && !(defined GMX_DOUBLE)
 193         { "sse4_1_single", 4 },
 194 #endif
 195 #if GMX_SIMD_X86_SSE2 && !(defined GMX_DOUBLE)
 196         { "sse2_single", 4 },
 197 #endif
 198         /* Double precision */
 199 #if GMX_SIMD_X86_AVX_256_OR_HIGHER && defined GMX_DOUBLE
 200         { "avx_256_double", 4 },
 201 #endif
 202 #if GMX_SIMD_X86_AVX_128_FMA && defined GMX_DOUBLE
 203         /* Sic. Double precision 2-way SIMD does not require neighbor list padding,
 204          * since the kernels execute a loop unrolled a factor 2, followed by
 205          * a possible single odd-element epilogue.
 206          */
 207         { "avx_128_fma_double", 1 },
 208 #endif
 209 #if GMX_SIMD_X86_SSE2 && defined GMX_DOUBLE
 210         /* No padding - see comment above */
 211         { "sse2_double", 1 },
 212 #endif
 213 #if GMX_SIMD_X86_SSE4_1 && defined GMX_DOUBLE
 214         /* No padding - see comment above */
 215         { "sse4_1_double", 1 },
 216 #endif
 217 #if GMX_SIMD_SPARC64_HPC_ACE && defined GMX_DOUBLE
 218         /* No padding - see comment above */
 219         { "sparc64_hpc_ace_double", 1 },
 220 #endif
 221         { "c", 1 },
 222     };
 223     int              narch = asize(arch_and_padding);
 224     int              i;
 225
 226     if (nonbonded_setup_done == FALSE)
 227     {
 228         /* We typically call this setup routine before starting timers,
 229          * but if that has not been done for whatever reason we do it now.
 230          */
 231         gmx_nonbonded_setup(NULL, FALSE);
 232     }
 233
 234     /* Not used yet */
 235     other = "";
 236
 237     nl->kernelptr_vf = NULL;
 238     nl->kernelptr_v  = NULL;
 239     nl->kernelptr_f  = NULL;
 240
 241     elec     = gmx_nbkernel_elec_names[nl->ielec];
 242     elec_mod = eintmod_names[nl->ielecmod];
 243     vdw      = gmx_nbkernel_vdw_names[nl->ivdw];
 244     vdw_mod  = eintmod_names[nl->ivdwmod];
 245     geom     = gmx_nblist_geometry_names[nl->igeometry];
 246
 247     if (nl->type == GMX_NBLIST_INTERACTION_FREE_ENERGY)
 248     {
 249         nl->kernelptr_vf       = (void *) gmx_nb_free_energy_kernel;
 250         nl->kernelptr_f        = (void *) gmx_nb_free_energy_kernel;
 251         nl->simd_padding_width = 1;
 252     }
 253     else if (!gmx_strcasecmp_min(geom, "CG-CG"))
 254     {
 255         nl->kernelptr_vf       = (void *) gmx_nb_generic_cg_kernel;
 256         nl->kernelptr_f        = (void *) gmx_nb_generic_cg_kernel;
 257         nl->simd_padding_width = 1;
 258     }
 259     else
 260     {
 261         /* Try to find a specific kernel first */
 262
 263         for (i = 0; i < narch && nl->kernelptr_vf == NULL; i++)
 264         {
 265             nl->kernelptr_vf       = (void *) nb_kernel_list_findkernel(log, arch_and_padding[i].arch, elec, elec_mod, vdw, vdw_mod, geom, other, "PotentialAndForce");
 266             nl->simd_padding_width = arch_and_padding[i].simd_padding_width;
 267         }
 268         for (i = 0; i < narch && nl->kernelptr_f == NULL; i++)
 269         {
 270             nl->kernelptr_f        = (void *) nb_kernel_list_findkernel(log, arch_and_padding[i].arch, elec, elec_mod, vdw, vdw_mod, geom, other, "Force");
 271             nl->simd_padding_width = arch_and_padding[i].simd_padding_width;
 272
 273             /* If there is not force-only optimized kernel, is there a potential & force one? */
 274             if (nl->kernelptr_f == NULL)
 275             {
 276                 nl->kernelptr_f        = (void *) nb_kernel_list_findkernel(NULL, arch_and_padding[i].arch, elec, elec_mod, vdw, vdw_mod, geom, other, "PotentialAndForce");
 277                 nl->simd_padding_width = arch_and_padding[i].simd_padding_width;
 278             }
 279         }
 280
 281         /* For now, the accelerated kernels cannot handle the combination of switch functions for both
 282          * electrostatics and VdW that use different switch radius or switch cutoff distances
 283          * (both of them enter in the switch function calculation). This would require
 284          * us to evaluate two completely separate switch functions for every interaction.
 285          * Instead, we disable such kernels by setting the pointer to NULL.
 286          * This will cause the generic kernel (which can handle it) to be called instead.
 287          *
 288          * Note that we typically already enable tabulated coulomb interactions for this case,
 289          * so this is mostly a safe-guard to make sure we call the generic kernel if the
 290          * tables are disabled.
 291          */
 292         if ((nl->ielec != GMX_NBKERNEL_ELEC_NONE) && (nl->ielecmod == eintmodPOTSWITCH) &&
 293             (nl->ivdw  != GMX_NBKERNEL_VDW_NONE)  && (nl->ivdwmod  == eintmodPOTSWITCH) &&
 294             bElecAndVdwSwitchDiffers)
 295         {
 296             nl->kernelptr_vf = NULL;
 297             nl->kernelptr_f  = NULL;
 298         }
 299
 300         /* Give up, pick a generic one instead.
 301          * We only do this for particle-particle kernels; by leaving the water-optimized kernel
 302          * pointers to NULL, the water optimization will automatically be disabled for this interaction.
 303          */
 304         if (nl->kernelptr_vf == NULL && !gmx_strcasecmp_min(geom, "Particle-Particle"))
 305         {
 306             nl->kernelptr_vf       = (void *) gmx_nb_generic_kernel;
 307             nl->kernelptr_f        = (void *) gmx_nb_generic_kernel;
 308             nl->simd_padding_width = 1;
 309             if (debug)
 310             {
 311                 fprintf(debug,
 312                         "WARNING - Slow generic NB kernel used for neighborlist with\n"
 313                         "    Elec: '%s', Modifier: '%s'\n"
 314                         "    Vdw:  '%s', Modifier: '%s'\n",
 315                         elec, elec_mod, vdw, vdw_mod);
 316             }
 317         }
 318     }
 319     return;
 320 }
 321
 322 void do_nonbonded(t_forcerec *fr,
 323                   rvec x[], rvec f_shortrange[], rvec f_longrange[], t_mdatoms *mdatoms, t_blocka *excl,
 324                   gmx_grppairener_t *grppener,
 325                   t_nrnb *nrnb, real *lambda, real *dvdl,
 326                   int nls, int eNL, int flags)
 327 {
 328     t_nblist *        nlist;
 329     int               n, n0, n1, i, i0, i1, range;
 330     t_nblists *       nblists;
 331     nb_kernel_data_t  kernel_data;
 332     nb_kernel_t *     kernelptr = NULL;
 333     rvec *            f;
 334
 335     kernel_data.flags                   = flags;
 336     kernel_data.exclusions              = excl;
 337     kernel_data.lambda                  = lambda;
 338     kernel_data.dvdl                    = dvdl;
 339
 340     if (fr->bAllvsAll)
 341     {
 342         gmx_incons("All-vs-all kernels have not been implemented in version 4.6");
 343         return;
 344     }
 345
 346     if (eNL >= 0)
 347     {
 348         i0 = eNL;
 349         i1 = i0+1;
 350     }
 351     else
 352     {
 353         i0 = 0;
 354         i1 = eNL_NR;
 355     }
 356
 357     if (nls >= 0)
 358     {
 359         n0 = nls;
 360         n1 = nls+1;
 361     }
 362     else
 363     {
 364         n0 = 0;
 365         n1 = fr->nnblists;
 366     }
 367
 368     for (n = n0; (n < n1); n++)
 369     {
 370         nblists = &fr->nblists[n];
 371
 372         /* Tabulated kernels hard-code a lot of assumptions about the
 373          * structure of these tables, but that's not worth fixing with
 374          * the group scheme due for removal soon. As a token
 375          * improvement, this assertion will stop code segfaulting if
 376          * someone assumes that extending the group-scheme table-type
 377          * enumeration is something that GROMACS supports. */
 378         /* cppcheck-suppress duplicateExpression */
 379         assert(etiNR == 3);
 380
 381         kernel_data.table_elec              = &nblists->table_elec;
 382         kernel_data.table_vdw               = &nblists->table_vdw;
 383         kernel_data.table_elec_vdw          = &nblists->table_elec_vdw;
 384
 385         for (range = 0; range < 2; range++)
 386         {
 387             /* Are we doing short/long-range? */
 388             if (range == 0)
 389             {
 390                 /* Short-range */
 391                 if (!(flags & GMX_NONBONDED_DO_SR))
 392                 {
 393                     continue;
 394                 }
 395                 kernel_data.energygrp_elec          = grppener->ener[egCOULSR];
 396                 kernel_data.energygrp_vdw           = grppener->ener[fr->bBHAM ? egBHAMSR : egLJSR];
 397                 kernel_data.energygrp_polarization  = grppener->ener[egGB];
 398                 nlist = nblists->nlist_sr;
 399                 f                                   = f_shortrange;
 400             }
 401             else
 402             {
 403                 /* Long-range */
 404                 if (!(flags & GMX_NONBONDED_DO_LR))
 405                 {
 406                     continue;
 407                 }
 408                 kernel_data.energygrp_elec          = grppener->ener[egCOULLR];
 409                 kernel_data.energygrp_vdw           = grppener->ener[fr->bBHAM ? egBHAMLR : egLJLR];
 410                 kernel_data.energygrp_polarization  = grppener->ener[egGB];
 411                 nlist = nblists->nlist_lr;
 412                 f                                   = f_longrange;
 413             }
 414
 415             for (i = i0; (i < i1); i++)
 416             {
 417                 if (nlist[i].nri > 0)
 418                 {
 419                     if (flags & GMX_NONBONDED_DO_POTENTIAL)
 420                     {
 421                         /* Potential and force */
 422                         kernelptr = (nb_kernel_t *)nlist[i].kernelptr_vf;
 423                     }
 424                     else
 425                     {
 426                         /* Force only, no potential */
 427                         kernelptr = (nb_kernel_t *)nlist[i].kernelptr_f;
 428                     }
 429
 430                     if (nlist[i].type != GMX_NBLIST_INTERACTION_FREE_ENERGY && (flags & GMX_NONBONDED_DO_FOREIGNLAMBDA))
 431                     {
 432                         /* We don't need the non-perturbed interactions */
 433                         continue;
 434                     }
 435                     /* Neighborlists whose kernelptr==NULL will always be empty */
 436                     if (kernelptr != NULL)
 437                     {
 438                         (*kernelptr)(&(nlist[i]), x, f, fr, mdatoms, &kernel_data, nrnb);
 439                     }
 440                     else
 441                     {
 442                         gmx_fatal(FARGS, "Non-empty neighborlist does not have any kernel pointer assigned.");
 443                     }
 444                 }
 445             }
 446         }
 447     }
 448 }