2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2004, The GROMACS development team.
6 * Copyright (c) 2013,2014,2015,2017,2018, 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.
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.
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.
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.
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.
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.
39 #include "nonbonded.h"
47 #include "thread_mpi/threads.h"
49 #include "gromacs/gmxlib/nrnb.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/listed-forces/bonded.h"
55 #include "gromacs/math/utilities.h"
56 #include "gromacs/math/vec.h"
57 #include "gromacs/mdtypes/enerdata.h"
58 #include "gromacs/mdtypes/forcerec.h"
59 #include "gromacs/mdtypes/md_enums.h"
60 #include "gromacs/mdtypes/mdatom.h"
61 #include "gromacs/mdtypes/nblist.h"
62 #include "gromacs/pbcutil/ishift.h"
63 #include "gromacs/pbcutil/mshift.h"
64 #include "gromacs/pbcutil/pbc.h"
65 #include "gromacs/tables/forcetable.h"
66 #include "gromacs/utility/arraysize.h"
67 #include "gromacs/utility/basedefinitions.h"
68 #include "gromacs/utility/cstringutil.h"
69 #include "gromacs/utility/fatalerror.h"
70 #include "gromacs/utility/smalloc.h"
72 /* Different default (c) and SIMD instructions interaction-specific kernels */
73 #include "gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_c.h"
75 #if GMX_SIMD_X86_SSE2 && !GMX_DOUBLE
76 # include "gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_sse2_single.h"
78 #if GMX_SIMD_X86_SSE4_1 && !GMX_DOUBLE
79 # include "gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_sse4_1_single.h"
81 #if GMX_SIMD_X86_AVX_128_FMA && !GMX_DOUBLE
82 # include "gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_avx_128_fma_single.h"
84 #if (GMX_SIMD_X86_AVX_256 || GMX_SIMD_X86_AVX2_256) && !GMX_DOUBLE
85 # include "gromacs/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_avx_256_single.h"
87 #if GMX_SIMD_X86_SSE2 && GMX_DOUBLE
88 # include "gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_sse2_double.h"
90 #if GMX_SIMD_X86_SSE4_1 && GMX_DOUBLE
91 # include "gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_sse4_1_double.h"
93 #if GMX_SIMD_X86_AVX_128_FMA && GMX_DOUBLE
94 # include "gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_avx_128_fma_double.h"
96 #if (GMX_SIMD_X86_AVX_256 || GMX_SIMD_X86_AVX2_256) && GMX_DOUBLE
97 # include "gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_avx_256_double.h"
99 #if GMX_SIMD_SPARC64_HPC_ACE && GMX_DOUBLE
100 # include "gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_sparc64_hpc_ace_double.h"
104 static tMPI_Thread_mutex_t nonbonded_setup_mutex = TMPI_THREAD_MUTEX_INITIALIZER;
105 static gmx_bool nonbonded_setup_done = FALSE;
109 gmx_nonbonded_setup(t_forcerec * fr,
110 gmx_bool bGenericKernelOnly)
112 tMPI_Thread_mutex_lock(&nonbonded_setup_mutex);
113 /* Here we are guaranteed only one thread made it. */
114 if (nonbonded_setup_done == FALSE)
116 if (bGenericKernelOnly == FALSE)
118 /* Add the generic kernels to the structure stored statically in nb_kernel.c */
119 nb_kernel_list_add_kernels(kernellist_c, kernellist_c_size);
121 if (!(fr != nullptr && fr->use_simd_kernels == FALSE))
123 /* Add interaction-specific kernels for different architectures */
124 /* Single precision */
125 #if GMX_SIMD_X86_SSE2 && !GMX_DOUBLE
126 nb_kernel_list_add_kernels(kernellist_sse2_single, kernellist_sse2_single_size);
128 #if GMX_SIMD_X86_SSE4_1 && !GMX_DOUBLE
129 nb_kernel_list_add_kernels(kernellist_sse4_1_single, kernellist_sse4_1_single_size);
131 #if GMX_SIMD_X86_AVX_128_FMA && !GMX_DOUBLE
132 nb_kernel_list_add_kernels(kernellist_avx_128_fma_single, kernellist_avx_128_fma_single_size);
134 #if (GMX_SIMD_X86_AVX_256 || GMX_SIMD_X86_AVX2_256) && !GMX_DOUBLE
135 nb_kernel_list_add_kernels(kernellist_avx_256_single, kernellist_avx_256_single_size);
137 /* Double precision */
138 #if GMX_SIMD_X86_SSE2 && GMX_DOUBLE
139 nb_kernel_list_add_kernels(kernellist_sse2_double, kernellist_sse2_double_size);
141 #if GMX_SIMD_X86_SSE4_1 && GMX_DOUBLE
142 nb_kernel_list_add_kernels(kernellist_sse4_1_double, kernellist_sse4_1_double_size);
144 #if GMX_SIMD_X86_AVX_128_FMA && GMX_DOUBLE
145 nb_kernel_list_add_kernels(kernellist_avx_128_fma_double, kernellist_avx_128_fma_double_size);
147 #if (GMX_SIMD_X86_AVX_256 || GMX_SIMD_X86_AVX2_256) && GMX_DOUBLE
148 nb_kernel_list_add_kernels(kernellist_avx_256_double, kernellist_avx_256_double_size);
150 #if GMX_SIMD_SPARC64_HPC_ACE && GMX_DOUBLE
151 nb_kernel_list_add_kernels(kernellist_sparc64_hpc_ace_double, kernellist_sparc64_hpc_ace_double_size);
153 ; /* empty statement to avoid a completely empty block */
156 /* Create a hash for faster lookups */
157 nb_kernel_list_hash_init();
159 nonbonded_setup_done = TRUE;
161 tMPI_Thread_mutex_unlock(&nonbonded_setup_mutex);
167 gmx_nonbonded_set_kernel_pointers(FILE *log, t_nblist *nl, gmx_bool bElecAndVdwSwitchDiffers)
170 const char * elec_mod;
172 const char * vdw_mod;
179 int simd_padding_width;
183 /* Single precision */
184 #if (GMX_SIMD_X86_AVX_256 || GMX_SIMD_X86_AVX2_256) && !GMX_DOUBLE
185 { "avx_256_single", 8 },
187 #if GMX_SIMD_X86_AVX_128_FMA && !GMX_DOUBLE
188 { "avx_128_fma_single", 4 },
190 #if GMX_SIMD_X86_SSE4_1 && !GMX_DOUBLE
191 { "sse4_1_single", 4 },
193 #if GMX_SIMD_X86_SSE2 && !GMX_DOUBLE
194 { "sse2_single", 4 },
196 /* Double precision */
197 #if (GMX_SIMD_X86_AVX_256 || GMX_SIMD_X86_AVX2_256) && GMX_DOUBLE
198 { "avx_256_double", 4 },
200 #if GMX_SIMD_X86_AVX_128_FMA && GMX_DOUBLE
201 /* Sic. Double precision 2-way SIMD does not require neighbor list padding,
202 * since the kernels execute a loop unrolled a factor 2, followed by
203 * a possible single odd-element epilogue.
205 { "avx_128_fma_double", 1 },
207 #if GMX_SIMD_X86_SSE2 && GMX_DOUBLE
208 /* No padding - see comment above */
209 { "sse2_double", 1 },
211 #if GMX_SIMD_X86_SSE4_1 && GMX_DOUBLE
212 /* No padding - see comment above */
213 { "sse4_1_double", 1 },
215 #if GMX_SIMD_SPARC64_HPC_ACE && GMX_DOUBLE
216 /* No padding - see comment above */
217 { "sparc64_hpc_ace_double", 1 },
221 int narch = asize(arch_and_padding);
224 if (nonbonded_setup_done == FALSE)
226 /* We typically call this setup routine before starting timers,
227 * but if that has not been done for whatever reason we do it now.
229 gmx_nonbonded_setup(nullptr, FALSE);
235 nl->kernelptr_vf = nullptr;
236 nl->kernelptr_v = nullptr;
237 nl->kernelptr_f = nullptr;
239 elec = gmx_nbkernel_elec_names[nl->ielec];
240 elec_mod = eintmod_names[nl->ielecmod];
241 vdw = gmx_nbkernel_vdw_names[nl->ivdw];
242 vdw_mod = eintmod_names[nl->ivdwmod];
243 geom = gmx_nblist_geometry_names[nl->igeometry];
245 if (nl->type == GMX_NBLIST_INTERACTION_FREE_ENERGY)
247 nl->kernelptr_vf = (void *) gmx_nb_free_energy_kernel;
248 nl->kernelptr_f = (void *) gmx_nb_free_energy_kernel;
249 nl->simd_padding_width = 1;
251 else if (!gmx_strcasecmp_min(geom, "CG-CG"))
253 nl->kernelptr_vf = (void *) gmx_nb_generic_cg_kernel;
254 nl->kernelptr_f = (void *) gmx_nb_generic_cg_kernel;
255 nl->simd_padding_width = 1;
259 /* Try to find a specific kernel first */
261 for (i = 0; i < narch && nl->kernelptr_vf == nullptr; i++)
263 nl->kernelptr_vf = (void *) nb_kernel_list_findkernel(log, arch_and_padding[i].arch, elec, elec_mod, vdw, vdw_mod, geom, other, "PotentialAndForce");
264 nl->simd_padding_width = arch_and_padding[i].simd_padding_width;
266 for (i = 0; i < narch && nl->kernelptr_f == nullptr; i++)
268 nl->kernelptr_f = (void *) nb_kernel_list_findkernel(log, arch_and_padding[i].arch, elec, elec_mod, vdw, vdw_mod, geom, other, "Force");
269 nl->simd_padding_width = arch_and_padding[i].simd_padding_width;
271 /* If there is not force-only optimized kernel, is there a potential & force one? */
272 if (nl->kernelptr_f == nullptr)
274 nl->kernelptr_f = (void *) nb_kernel_list_findkernel(nullptr, 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;
279 /* For now, the accelerated kernels cannot handle the combination of switch functions for both
280 * electrostatics and VdW that use different switch radius or switch cutoff distances
281 * (both of them enter in the switch function calculation). This would require
282 * us to evaluate two completely separate switch functions for every interaction.
283 * Instead, we disable such kernels by setting the pointer to NULL.
284 * This will cause the generic kernel (which can handle it) to be called instead.
286 * Note that we typically already enable tabulated coulomb interactions for this case,
287 * so this is mostly a safe-guard to make sure we call the generic kernel if the
288 * tables are disabled.
290 if ((nl->ielec != GMX_NBKERNEL_ELEC_NONE) && (nl->ielecmod == eintmodPOTSWITCH) &&
291 (nl->ivdw != GMX_NBKERNEL_VDW_NONE) && (nl->ivdwmod == eintmodPOTSWITCH) &&
292 bElecAndVdwSwitchDiffers)
294 nl->kernelptr_vf = nullptr;
295 nl->kernelptr_f = nullptr;
298 /* Give up, pick a generic one instead.
299 * We only do this for particle-particle kernels; by leaving the water-optimized kernel
300 * pointers to NULL, the water optimization will automatically be disabled for this interaction.
302 if (nl->kernelptr_vf == nullptr && !gmx_strcasecmp_min(geom, "Particle-Particle"))
304 nl->kernelptr_vf = (void *) gmx_nb_generic_kernel;
305 nl->kernelptr_f = (void *) gmx_nb_generic_kernel;
306 nl->simd_padding_width = 1;
310 "WARNING - Slow generic NB kernel used for neighborlist with\n"
311 " Elec: '%s', Modifier: '%s'\n"
312 " Vdw: '%s', Modifier: '%s'\n",
313 elec, elec_mod, vdw, vdw_mod);
319 void do_nonbonded(const t_forcerec *fr,
322 const t_mdatoms *mdatoms,
323 const t_blocka *excl,
324 gmx_grppairener_t *grppener,
333 int n, n0, n1, i, i0, i1;
335 nb_kernel_data_t kernel_data;
336 nb_kernel_t * kernelptr = nullptr;
339 kernel_data.flags = flags;
340 kernel_data.exclusions = excl;
341 kernel_data.lambda = lambda;
342 kernel_data.dvdl = dvdl;
346 gmx_incons("All-vs-all kernels have not been implemented in version 4.6");
371 for (n = n0; (n < n1); n++)
373 nblists = &fr->nblists[n];
375 /* Tabulated kernels hard-code a lot of assumptions about the
376 * structure of these tables, but that's not worth fixing with
377 * the group scheme due for removal soon. As a token
378 * improvement, this assertion will stop code segfaulting if
379 * someone assumes that extending the group-scheme table-type
380 * enumeration is something that GROMACS supports. */
381 /* cppcheck-suppress duplicateExpression */
382 static_assert(etiNR == 3, "");
384 kernel_data.table_elec = nblists->table_elec;
385 kernel_data.table_vdw = nblists->table_vdw;
386 kernel_data.table_elec_vdw = nblists->table_elec_vdw;
391 if (!(flags & GMX_NONBONDED_DO_SR))
395 kernel_data.energygrp_elec = grppener->ener[egCOULSR];
396 kernel_data.energygrp_vdw = grppener->ener[fr->bBHAM ? egBHAMSR : egLJSR];
397 nlist = nblists->nlist_sr;
401 for (i = i0; (i < i1); i++)
403 if (nlist[i].nri > 0)
405 if (flags & GMX_NONBONDED_DO_POTENTIAL)
407 /* Potential and force */
408 kernelptr = (nb_kernel_t *)nlist[i].kernelptr_vf;
412 /* Force only, no potential */
413 kernelptr = (nb_kernel_t *)nlist[i].kernelptr_f;
416 if (nlist[i].type != GMX_NBLIST_INTERACTION_FREE_ENERGY && (flags & GMX_NONBONDED_DO_FOREIGNLAMBDA))
418 /* We don't need the non-perturbed interactions */
421 /* Neighborlists whose kernelptr==NULL will always be empty */
422 if (kernelptr != nullptr)
424 (*kernelptr)(&(nlist[i]), x, f, fr, mdatoms, &kernel_data, nrnb);
428 gmx_fatal(FARGS, "Non-empty neighborlist does not have any kernel pointer assigned.");