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 "gromacs/legacyheaders/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/gmxlib/nonbonded/nb_free_energy.h"
  50 #include "gromacs/gmxlib/nonbonded/nb_generic.h"
  51 #include "gromacs/gmxlib/nonbonded/nb_generic_adress.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/txtdump.h"
  57 #include "gromacs/legacyheaders/types/forcerec.h"
  58 #include "gromacs/legacyheaders/types/mdatom.h"
  59 #include "gromacs/legacyheaders/types/nblist.h"
  60 #include "gromacs/legacyheaders/types/nrnb.h"
  61 #include "gromacs/listed-forces/bonded.h"
  62 #include "gromacs/math/utilities.h"
  63 #include "gromacs/math/vec.h"
  64 #include "gromacs/pbcutil/ishift.h"
  65 #include "gromacs/pbcutil/mshift.h"
  66 #include "gromacs/pbcutil/pbc.h"
  67 #include "gromacs/simd/simd.h"
  68 #include "gromacs/tables/forcetable.h"
  69 #include "gromacs/utility/arraysize.h"
  70 #include "gromacs/utility/basedefinitions.h"
  71 #include "gromacs/utility/cstringutil.h"
  72 #include "gromacs/utility/fatalerror.h"
  73 #include "gromacs/utility/smalloc.h"
  74
  75 /* Different default (c) and SIMD instructions interaction-specific kernels */
  76 #include "gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_c.h"
  77
  78 #if GMX_SIMD_X86_SSE2 && !(defined GMX_DOUBLE)
  79 #    include "gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_sse2_single.h"
  80 #endif
  81 #if GMX_SIMD_X86_SSE4_1 && !(defined GMX_DOUBLE)
  82 #    include "gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_sse4_1_single.h"
  83 #endif
  84 #if GMX_SIMD_X86_AVX_128_FMA && !(defined GMX_DOUBLE)
  85 #    include "gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_avx_128_fma_single.h"
  86 #endif
  87 #if GMX_SIMD_X86_AVX_256_OR_HIGHER && !(defined GMX_DOUBLE)
  88 #    include "gromacs/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_avx_256_single.h"
  89 #endif
  90 #if GMX_SIMD_X86_SSE2 && defined GMX_DOUBLE
  91 #    include "gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_sse2_double.h"
  92 #endif
  93 #if GMX_SIMD_X86_SSE4_1 && defined GMX_DOUBLE
  94 #    include "gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_sse4_1_double.h"
  95 #endif
  96 #if GMX_SIMD_X86_AVX_128_FMA && defined GMX_DOUBLE
  97 #    include "gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_avx_128_fma_double.h"
  98 #endif
  99 #if GMX_SIMD_X86_AVX_256_OR_HIGHER && defined GMX_DOUBLE
 100 #    include "gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_avx_256_double.h"
 101 #endif
 102 #if GMX_SIMD_SPARC64_HPC_ACE && defined GMX_DOUBLE
 103 #    include "gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_sparc64_hpc_ace_double.h"
 104 #endif
 105
 106
 107 static tMPI_Thread_mutex_t nonbonded_setup_mutex = TMPI_THREAD_MUTEX_INITIALIZER;
 108 static gmx_bool            nonbonded_setup_done  = FALSE;
 109
 110
 111 void
 112 gmx_nonbonded_setup(t_forcerec *   fr,
 113                     gmx_bool       bGenericKernelOnly)
 114 {
 115     tMPI_Thread_mutex_lock(&nonbonded_setup_mutex);
 116     /* Here we are guaranteed only one thread made it. */
 117     if (nonbonded_setup_done == FALSE)
 118     {
 119         if (bGenericKernelOnly == FALSE)
 120         {
 121             /* Add the generic kernels to the structure stored statically in nb_kernel.c */
 122             nb_kernel_list_add_kernels(kernellist_c, kernellist_c_size);
 123
 124             if (!(fr != NULL && fr->use_simd_kernels == FALSE))
 125             {
 126                 /* Add interaction-specific kernels for different architectures */
 127                 /* Single precision */
 128 #if GMX_SIMD_X86_SSE2 && !(defined GMX_DOUBLE)
 129                 nb_kernel_list_add_kernels(kernellist_sse2_single, kernellist_sse2_single_size);
 130 #endif
 131 #if GMX_SIMD_X86_SSE4_1 && !(defined GMX_DOUBLE)
 132                 nb_kernel_list_add_kernels(kernellist_sse4_1_single, kernellist_sse4_1_single_size);
 133 #endif
 134 #if GMX_SIMD_X86_AVX_128_FMA && !(defined GMX_DOUBLE)
 135                 nb_kernel_list_add_kernels(kernellist_avx_128_fma_single, kernellist_avx_128_fma_single_size);
 136 #endif
 137 #if GMX_SIMD_X86_AVX_256_OR_HIGHER && !(defined GMX_DOUBLE)
 138                 nb_kernel_list_add_kernels(kernellist_avx_256_single, kernellist_avx_256_single_size);
 139 #endif
 140                 /* Double precision */
 141 #if GMX_SIMD_X86_SSE2 && defined GMX_DOUBLE
 142                 nb_kernel_list_add_kernels(kernellist_sse2_double, kernellist_sse2_double_size);
 143 #endif
 144 #if GMX_SIMD_X86_SSE4_1 && defined GMX_DOUBLE
 145                 nb_kernel_list_add_kernels(kernellist_sse4_1_double, kernellist_sse4_1_double_size);
 146 #endif
 147 #if GMX_SIMD_X86_AVX_128_FMA && defined GMX_DOUBLE
 148                 nb_kernel_list_add_kernels(kernellist_avx_128_fma_double, kernellist_avx_128_fma_double_size);
 149 #endif
 150 #if GMX_SIMD_X86_AVX_256_OR_HIGHER && defined GMX_DOUBLE
 151                 nb_kernel_list_add_kernels(kernellist_avx_256_double, kernellist_avx_256_double_size);
 152 #endif
 153 #if GMX_SIMD_SPARC64_HPC_ACE && defined GMX_DOUBLE
 154                 nb_kernel_list_add_kernels(kernellist_sparc64_hpc_ace_double, kernellist_sparc64_hpc_ace_double_size);
 155 #endif
 156                 ; /* empty statement to avoid a completely empty block */
 157             }
 158         }
 159         /* Create a hash for faster lookups */
 160         nb_kernel_list_hash_init();
 161
 162         nonbonded_setup_done = TRUE;
 163     }
 164     tMPI_Thread_mutex_unlock(&nonbonded_setup_mutex);
 165 }
 166
 167
 168
 169 void
 170 gmx_nonbonded_set_kernel_pointers(FILE *log, t_nblist *nl, gmx_bool bElecAndVdwSwitchDiffers)
 171 {
 172     const char *     elec;
 173     const char *     elec_mod;
 174     const char *     vdw;
 175     const char *     vdw_mod;
 176     const char *     geom;
 177     const char *     other;
 178
 179     struct
 180     {
 181         const char *  arch;
 182         int           simd_padding_width;
 183     }
 184     arch_and_padding[] =
 185     {
 186         /* Single precision */
 187 #if GMX_SIMD_X86_AVX_256_OR_HIGHER && !(defined GMX_DOUBLE)
 188         { "avx_256_single", 8 },
 189 #endif
 190 #if GMX_SIMD_X86_AVX_128_FMA && !(defined GMX_DOUBLE)
 191         { "avx_128_fma_single", 4 },
 192 #endif
 193 #if GMX_SIMD_X86_SSE4_1 && !(defined GMX_DOUBLE)
 194         { "sse4_1_single", 4 },
 195 #endif
 196 #if GMX_SIMD_X86_SSE2 && !(defined GMX_DOUBLE)
 197         { "sse2_single", 4 },
 198 #endif
 199         /* Double precision */
 200 #if GMX_SIMD_X86_AVX_256_OR_HIGHER && defined GMX_DOUBLE
 201         { "avx_256_double", 4 },
 202 #endif
 203 #if GMX_SIMD_X86_AVX_128_FMA && defined GMX_DOUBLE
 204         /* Sic. Double precision 2-way SIMD does not require neighbor list padding,
 205          * since the kernels execute a loop unrolled a factor 2, followed by
 206          * a possible single odd-element epilogue.
 207          */
 208         { "avx_128_fma_double", 1 },
 209 #endif
 210 #if GMX_SIMD_X86_SSE2 && defined GMX_DOUBLE
 211         /* No padding - see comment above */
 212         { "sse2_double", 1 },
 213 #endif
 214 #if GMX_SIMD_X86_SSE4_1 && defined GMX_DOUBLE
 215         /* No padding - see comment above */
 216         { "sse4_1_double", 1 },
 217 #endif
 218 #if GMX_SIMD_SPARC64_HPC_ACE && defined GMX_DOUBLE
 219         /* No padding - see comment above */
 220         { "sparc64_hpc_ace_double", 1 },
 221 #endif
 222         { "c", 1 },
 223     };
 224     int              narch = asize(arch_and_padding);
 225     int              i;
 226
 227     if (nonbonded_setup_done == FALSE)
 228     {
 229         /* We typically call this setup routine before starting timers,
 230          * but if that has not been done for whatever reason we do it now.
 231          */
 232         gmx_nonbonded_setup(NULL, FALSE);
 233     }
 234
 235     /* Not used yet */
 236     other = "";
 237
 238     nl->kernelptr_vf = NULL;
 239     nl->kernelptr_v  = NULL;
 240     nl->kernelptr_f  = NULL;
 241
 242     elec     = gmx_nbkernel_elec_names[nl->ielec];
 243     elec_mod = eintmod_names[nl->ielecmod];
 244     vdw      = gmx_nbkernel_vdw_names[nl->ivdw];
 245     vdw_mod  = eintmod_names[nl->ivdwmod];
 246     geom     = gmx_nblist_geometry_names[nl->igeometry];
 247
 248     if (nl->type == GMX_NBLIST_INTERACTION_ADRESS)
 249     {
 250         nl->kernelptr_vf       = (void *) gmx_nb_generic_adress_kernel;
 251         nl->kernelptr_f        = (void *) gmx_nb_generic_adress_kernel;
 252         nl->simd_padding_width = 1;
 253         return;
 254     }
 255
 256     if (nl->type == GMX_NBLIST_INTERACTION_FREE_ENERGY)
 257     {
 258         nl->kernelptr_vf       = (void *) gmx_nb_free_energy_kernel;
 259         nl->kernelptr_f        = (void *) gmx_nb_free_energy_kernel;
 260         nl->simd_padding_width = 1;
 261     }
 262     else if (!gmx_strcasecmp_min(geom, "CG-CG"))
 263     {
 264         nl->kernelptr_vf       = (void *) gmx_nb_generic_cg_kernel;
 265         nl->kernelptr_f        = (void *) gmx_nb_generic_cg_kernel;
 266         nl->simd_padding_width = 1;
 267     }
 268     else
 269     {
 270         /* Try to find a specific kernel first */
 271
 272         for (i = 0; i < narch && nl->kernelptr_vf == NULL; i++)
 273         {
 274             nl->kernelptr_vf       = (void *) nb_kernel_list_findkernel(log, arch_and_padding[i].arch, elec, elec_mod, vdw, vdw_mod, geom, other, "PotentialAndForce");
 275             nl->simd_padding_width = arch_and_padding[i].simd_padding_width;
 276         }
 277         for (i = 0; i < narch && nl->kernelptr_f == NULL; i++)
 278         {
 279             nl->kernelptr_f        = (void *) nb_kernel_list_findkernel(log, arch_and_padding[i].arch, elec, elec_mod, vdw, vdw_mod, geom, other, "Force");
 280             nl->simd_padding_width = arch_and_padding[i].simd_padding_width;
 281
 282             /* If there is not force-only optimized kernel, is there a potential & force one? */
 283             if (nl->kernelptr_f == NULL)
 284             {
 285                 nl->kernelptr_f        = (void *) nb_kernel_list_findkernel(NULL, arch_and_padding[i].arch, elec, elec_mod, vdw, vdw_mod, geom, other, "PotentialAndForce");
 286                 nl->simd_padding_width = arch_and_padding[i].simd_padding_width;
 287             }
 288         }
 289
 290         /* For now, the accelerated kernels cannot handle the combination of switch functions for both
 291          * electrostatics and VdW that use different switch radius or switch cutoff distances
 292          * (both of them enter in the switch function calculation). This would require
 293          * us to evaluate two completely separate switch functions for every interaction.
 294          * Instead, we disable such kernels by setting the pointer to NULL.
 295          * This will cause the generic kernel (which can handle it) to be called instead.
 296          *
 297          * Note that we typically already enable tabulated coulomb interactions for this case,
 298          * so this is mostly a safe-guard to make sure we call the generic kernel if the
 299          * tables are disabled.
 300          */
 301         if ((nl->ielec != GMX_NBKERNEL_ELEC_NONE) && (nl->ielecmod == eintmodPOTSWITCH) &&
 302             (nl->ivdw  != GMX_NBKERNEL_VDW_NONE)  && (nl->ivdwmod  == eintmodPOTSWITCH) &&
 303             bElecAndVdwSwitchDiffers)
 304         {
 305             nl->kernelptr_vf = NULL;
 306             nl->kernelptr_f  = NULL;
 307         }
 308
 309         /* Give up, pick a generic one instead.
 310          * We only do this for particle-particle kernels; by leaving the water-optimized kernel
 311          * pointers to NULL, the water optimization will automatically be disabled for this interaction.
 312          */
 313         if (nl->kernelptr_vf == NULL && !gmx_strcasecmp_min(geom, "Particle-Particle"))
 314         {
 315             nl->kernelptr_vf       = (void *) gmx_nb_generic_kernel;
 316             nl->kernelptr_f        = (void *) gmx_nb_generic_kernel;
 317             nl->simd_padding_width = 1;
 318             if (debug)
 319             {
 320                 fprintf(debug,
 321                         "WARNING - Slow generic NB kernel used for neighborlist with\n"
 322                         "    Elec: '%s', Modifier: '%s'\n"
 323                         "    Vdw:  '%s', Modifier: '%s'\n",
 324                         elec, elec_mod, vdw, vdw_mod);
 325             }
 326         }
 327     }
 328     return;
 329 }
 330
 331 void do_nonbonded(t_forcerec *fr,
 332                   rvec x[], rvec f_shortrange[], rvec f_longrange[], t_mdatoms *mdatoms, t_blocka *excl,
 333                   gmx_grppairener_t *grppener,
 334                   t_nrnb *nrnb, real *lambda, real *dvdl,
 335                   int nls, int eNL, int flags)
 336 {
 337     t_nblist *        nlist;
 338     int               n, n0, n1, i, i0, i1, range;
 339     t_nblists *       nblists;
 340     nb_kernel_data_t  kernel_data;
 341     nb_kernel_t *     kernelptr = NULL;
 342     rvec *            f;
 343
 344     kernel_data.flags                   = flags;
 345     kernel_data.exclusions              = excl;
 346     kernel_data.lambda                  = lambda;
 347     kernel_data.dvdl                    = dvdl;
 348
 349     if (fr->bAllvsAll)
 350     {
 351         gmx_incons("All-vs-all kernels have not been implemented in version 4.6");
 352         return;
 353     }
 354
 355     if (eNL >= 0)
 356     {
 357         i0 = eNL;
 358         i1 = i0+1;
 359     }
 360     else
 361     {
 362         i0 = 0;
 363         i1 = eNL_NR;
 364     }
 365
 366     if (nls >= 0)
 367     {
 368         n0 = nls;
 369         n1 = nls+1;
 370     }
 371     else
 372     {
 373         n0 = 0;
 374         n1 = fr->nnblists;
 375     }
 376
 377     for (n = n0; (n < n1); n++)
 378     {
 379         nblists = &fr->nblists[n];
 380
 381         /* Tabulated kernels hard-code a lot of assumptions about the
 382          * structure of these tables, but that's not worth fixing with
 383          * the group scheme due for removal soon. As a token
 384          * improvement, this assertion will stop code segfaulting if
 385          * someone assumes that extending the group-scheme table-type
 386          * enumeration is something that GROMACS supports. */
 387         /* cppcheck-suppress duplicateExpression */
 388         assert(etiNR == 3);
 389
 390         kernel_data.table_elec              = &nblists->table_elec;
 391         kernel_data.table_vdw               = &nblists->table_vdw;
 392         kernel_data.table_elec_vdw          = &nblists->table_elec_vdw;
 393
 394         for (range = 0; range < 2; range++)
 395         {
 396             /* Are we doing short/long-range? */
 397             if (range == 0)
 398             {
 399                 /* Short-range */
 400                 if (!(flags & GMX_NONBONDED_DO_SR))
 401                 {
 402                     continue;
 403                 }
 404                 kernel_data.energygrp_elec          = grppener->ener[egCOULSR];
 405                 kernel_data.energygrp_vdw           = grppener->ener[fr->bBHAM ? egBHAMSR : egLJSR];
 406                 kernel_data.energygrp_polarization  = grppener->ener[egGB];
 407                 nlist = nblists->nlist_sr;
 408                 f                                   = f_shortrange;
 409             }
 410             else
 411             {
 412                 /* Long-range */
 413                 if (!(flags & GMX_NONBONDED_DO_LR))
 414                 {
 415                     continue;
 416                 }
 417                 kernel_data.energygrp_elec          = grppener->ener[egCOULLR];
 418                 kernel_data.energygrp_vdw           = grppener->ener[fr->bBHAM ? egBHAMLR : egLJLR];
 419                 kernel_data.energygrp_polarization  = grppener->ener[egGB];
 420                 nlist = nblists->nlist_lr;
 421                 f                                   = f_longrange;
 422             }
 423
 424             for (i = i0; (i < i1); i++)
 425             {
 426                 if (nlist[i].nri > 0)
 427                 {
 428                     if (flags & GMX_NONBONDED_DO_POTENTIAL)
 429                     {
 430                         /* Potential and force */
 431                         kernelptr = (nb_kernel_t *)nlist[i].kernelptr_vf;
 432                     }
 433                     else
 434                     {
 435                         /* Force only, no potential */
 436                         kernelptr = (nb_kernel_t *)nlist[i].kernelptr_f;
 437                     }
 438
 439                     if (nlist[i].type != GMX_NBLIST_INTERACTION_FREE_ENERGY && (flags & GMX_NONBONDED_DO_FOREIGNLAMBDA))
 440                     {
 441                         /* We don't need the non-perturbed interactions */
 442                         continue;
 443                     }
 444                     /* Neighborlists whose kernelptr==NULL will always be empty */
 445                     if (kernelptr != NULL)
 446                     {
 447                         (*kernelptr)(&(nlist[i]), x, f, fr, mdatoms, &kernel_data, nrnb);
 448                     }
 449                     else
 450                     {
 451                         gmx_fatal(FARGS, "Non-empty neighborlist does not have any kernel pointer assigned.");
 452                     }
 453                 }
 454             }
 455         }
 456     }
 457 }