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, 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.
37 /* This file is completely threadsafe - keep it that way! */
48 #include "gromacs/domdec/domdec.h"
49 #include "gromacs/fileio/gmxfio.h"
50 #include "gromacs/legacyheaders/constr.h"
51 #include "gromacs/legacyheaders/copyrite.h"
52 #include "gromacs/legacyheaders/gmx_omp_nthreads.h"
53 #include "gromacs/legacyheaders/mdrun.h"
54 #include "gromacs/legacyheaders/nrnb.h"
55 #include "gromacs/legacyheaders/types/commrec.h"
56 #include "gromacs/math/units.h"
57 #include "gromacs/math/vec.h"
58 #include "gromacs/pbcutil/pbc-simd.h"
59 #include "gromacs/pbcutil/pbc.h"
60 #include "gromacs/simd/simd.h"
61 #include "gromacs/simd/simd_math.h"
62 #include "gromacs/simd/vector_operations.h"
63 #include "gromacs/topology/block.h"
64 #include "gromacs/topology/mtop_util.h"
65 #include "gromacs/utility/bitmask.h"
66 #include "gromacs/utility/fatalerror.h"
67 #include "gromacs/utility/gmxomp.h"
68 #include "gromacs/utility/smalloc.h"
70 /* MSVC 2010 produces buggy SIMD PBC code, disable SIMD for MSVC <= 2010 */
71 #if defined GMX_SIMD_HAVE_REAL && !(defined _MSC_VER && _MSC_VER < 1700) && !defined(__ICL)
76 #if defined(GMX_SIMD_X86_AVX_256) || defined(GMX_SIMD_X86_AVX2_256)
78 // This was originally work-in-progress for augmenting the SIMD module with
79 // masked load/store operations. Instead, that turned into and extended SIMD
80 // interface that supports gather/scatter in all platforms, which will be
81 // part of a future Gromacs version. However, since the code for bonded
82 // interactions and LINCS was already written it would be a pity not to get
83 // the performance gains in Gromacs-5.1. For this reason we have added it as
84 // a bit of a hack in the two files that use it. It will be replaced with the
85 // new generic functionality after version 5.1
88 static gmx_inline void gmx_simdcall
89 gmx_hack_simd_transpose4_r(gmx_simd_double_t *row0,
90 gmx_simd_double_t *row1,
91 gmx_simd_double_t *row2,
92 gmx_simd_double_t *row3)
94 __m256d tmp0, tmp1, tmp2, tmp3;
96 tmp0 = _mm256_unpacklo_pd(*row0, *row1);
97 tmp2 = _mm256_unpacklo_pd(*row2, *row3);
98 tmp1 = _mm256_unpackhi_pd(*row0, *row1);
99 tmp3 = _mm256_unpackhi_pd(*row2, *row3);
100 *row0 = _mm256_permute2f128_pd(tmp0, tmp2, 0b00100000);
101 *row1 = _mm256_permute2f128_pd(tmp1, tmp3, 0b00100000);
102 *row2 = _mm256_permute2f128_pd(tmp0, tmp2, 0b00110001);
103 *row3 = _mm256_permute2f128_pd(tmp1, tmp3, 0b00110001);
106 static gmx_inline void gmx_simdcall
107 gmx_hack_simd4_transpose_to_simd_r(const gmx_simd4_double_t *a,
108 gmx_simd_double_t *row0,
109 gmx_simd_double_t *row1,
110 gmx_simd_double_t *row2,
111 gmx_simd_double_t *row3)
118 gmx_hack_simd_transpose4_r(row0, row1, row2, row3);
121 static gmx_inline void gmx_simdcall
122 gmx_hack_simd_transpose_to_simd4_r(gmx_simd_double_t row0,
123 gmx_simd_double_t row1,
124 gmx_simd_double_t row2,
125 gmx_simd_double_t row3,
126 gmx_simd4_double_t *a)
133 gmx_hack_simd_transpose4_r(&a[0], &a[1], &a[2], &a[3]);
137 # ifdef GMX_SIMD_X86_AVX_GCC_MASKLOAD_BUG
138 # define gmx_hack_simd4_load3_r(mem) _mm256_maskload_pd((mem), _mm_castsi128_ps(_mm256_set_epi32(0, 0, -1, -1, -1, -1, -1, -1)))
139 # define gmx_hack_simd4_store3_r(mem, x) _mm256_maskstore_pd((mem), _mm_castsi128_ps(_mm256_set_epi32(0, 0, -1, -1, -1, -1, -1, -1)), (x))
141 # define gmx_hack_simd4_load3_r(mem) _mm256_maskload_pd((mem), _mm256_set_epi32(0, 0, -1, -1, -1, -1, -1, -1))
142 # define gmx_hack_simd4_store3_r(mem, x) _mm256_maskstore_pd((mem), _mm256_set_epi32(0, 0, -1, -1, -1, -1, -1, -1), (x))
145 # else /* single instead of double */
146 static gmx_inline void gmx_simdcall
147 gmx_hack_simd_transpose4_r(gmx_simd_float_t *row0,
148 gmx_simd_float_t *row1,
149 gmx_simd_float_t *row2,
150 gmx_simd_float_t *row3)
152 __m256 tmp0, tmp1, tmp2, tmp3;
154 tmp0 = _mm256_unpacklo_ps(*row0, *row1);
155 tmp2 = _mm256_unpacklo_ps(*row2, *row3);
156 tmp1 = _mm256_unpackhi_ps(*row0, *row1);
157 tmp3 = _mm256_unpackhi_ps(*row2, *row3);
158 *row0 = _mm256_shuffle_ps(tmp0, tmp2, 0b0100010001000100);
159 *row1 = _mm256_shuffle_ps(tmp0, tmp2, 0b1110111011101110);
160 *row2 = _mm256_shuffle_ps(tmp1, tmp3, 0b0100010001000100);
161 *row3 = _mm256_shuffle_ps(tmp1, tmp3, 0b1110111011101110);
164 static gmx_inline void gmx_simdcall
165 gmx_hack_simd4_transpose_to_simd_r(const gmx_simd4_float_t *a,
166 gmx_simd_float_t *row0,
167 gmx_simd_float_t *row1,
168 gmx_simd_float_t *row2,
169 gmx_simd_float_t *row3)
171 *row0 = _mm256_insertf128_ps(_mm256_castps128_ps256(a[0]), a[4], 1);
172 *row1 = _mm256_insertf128_ps(_mm256_castps128_ps256(a[1]), a[5], 1);
173 *row2 = _mm256_insertf128_ps(_mm256_castps128_ps256(a[2]), a[6], 1);
174 *row3 = _mm256_insertf128_ps(_mm256_castps128_ps256(a[3]), a[7], 1);
176 gmx_hack_simd_transpose4_r(row0, row1, row2, row3);
179 static gmx_inline void gmx_simdcall
180 gmx_hack_simd_transpose_to_simd4_r(gmx_simd_float_t row0,
181 gmx_simd_float_t row1,
182 gmx_simd_float_t row2,
183 gmx_simd_float_t row3,
184 gmx_simd4_float_t *a)
186 gmx_hack_simd_transpose4_r(&row0, &row1, &row2, &row3);
188 a[0] = _mm256_extractf128_ps(row0, 0);
189 a[1] = _mm256_extractf128_ps(row1, 0);
190 a[2] = _mm256_extractf128_ps(row2, 0);
191 a[3] = _mm256_extractf128_ps(row3, 0);
192 a[4] = _mm256_extractf128_ps(row0, 1);
193 a[5] = _mm256_extractf128_ps(row1, 1);
194 a[6] = _mm256_extractf128_ps(row2, 1);
195 a[7] = _mm256_extractf128_ps(row3, 1);
197 #ifdef GMX_SIMD_X86_AVX_GCC_MASKLOAD_BUG
198 # define gmx_hack_simd4_load3_r(mem) _mm_maskload_ps((mem), _mm_castsi256_pd(_mm_set_epi32(0, -1, -1, -1)))
199 # define gmx_hack_simd4_store3_r(mem, x) _mm_maskstore_ps((mem), _mm_castsi256_pd(_mm_set_epi32(0, -1, -1, -1)), (x))
201 # define gmx_hack_simd4_load3_r(mem) _mm_maskload_ps((mem), _mm_set_epi32(0, -1, -1, -1))
202 # define gmx_hack_simd4_store3_r(mem, x) _mm_maskstore_ps((mem), _mm_set_epi32(0, -1, -1, -1), (x))
209 #ifdef GMX_SIMD_HAVE_REAL
210 /*! \brief Store differences between indexed rvecs in SIMD registers.
212 * Returns SIMD register with the difference vectors:
213 * v[pair_index[i*2]] - v[pair_index[i*2 + 1]]
215 * \param[in] v Array of rvecs
216 * \param[in] pair_index Index pairs for GMX_SIMD_REAL_WIDTH vector pairs
217 * \param[in,out] buf Aligned tmp buffer of size 3*GMX_SIMD_REAL_WIDTH
218 * \param[out] dx SIMD register with x difference
219 * \param[out] dy SIMD register with y difference
220 * \param[out] dz SIMD register with z difference
222 static gmx_inline void gmx_simdcall
223 gmx_hack_simd_gather_rvec_dist_pair_index(const rvec *v,
224 const int *pair_index,
225 real gmx_unused *buf,
230 #if defined(GMX_SIMD_X86_AVX_256) || defined(GMX_SIMD_X86_AVX2_256)
232 gmx_simd4_real_t d[GMX_SIMD_REAL_WIDTH];
235 for (i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
237 d[i] = gmx_simd4_sub_r(gmx_hack_simd4_load3_r(&(v[pair_index[i*2 + 0]][0])),
238 gmx_hack_simd4_load3_r(&(v[pair_index[i*2 + 1]][0])));
241 gmx_hack_simd4_transpose_to_simd_r(d, dx, dy, dz, &tmp);
244 GMX_ALIGNED(real, GMX_SIMD_REAL_WIDTH) buf_aligned[3*GMX_SIMD_REAL_WIDTH];
246 real* buf_aligned = buf;
251 for (i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
253 /* Store the distances packed and aligned */
254 for (m = 0; m < DIM; m++)
256 buf_aligned[m*GMX_SIMD_REAL_WIDTH + i] =
257 v[pair_index[i*2]][m] - v[pair_index[i*2 + 1]][m];
260 *dx = gmx_simd_load_r(buf_aligned + 0*GMX_SIMD_REAL_WIDTH);
261 *dy = gmx_simd_load_r(buf_aligned + 1*GMX_SIMD_REAL_WIDTH);
262 *dz = gmx_simd_load_r(buf_aligned + 2*GMX_SIMD_REAL_WIDTH);
267 /*! \brief Stores SIMD vector into multiple rvecs.
269 * \param[in] x SIMD register with x-components of the vectors
270 * \param[in] y SIMD register with y-components of the vectors
271 * \param[in] z SIMD register with z-components of the vectors
272 * \param[in,out] buf Aligned tmp buffer of size 3*GMX_SIMD_REAL_WIDTH
273 * \param[out] v Array of GMX_SIMD_REAL_WIDTH rvecs
275 static gmx_inline void gmx_simdcall
276 gmx_simd_store_vec_to_rvec(gmx_simd_real_t x,
279 real gmx_unused *buf,
282 #if defined(GMX_SIMD_X86_AVX_256) || defined(GMX_SIMD_X86_AVX2_256)
284 gmx_simd4_real_t s4[GMX_SIMD_REAL_WIDTH];
285 gmx_simd_real_t zero = gmx_simd_setzero_r();
287 gmx_hack_simd_transpose_to_simd4_r(x, y, z, zero, s4);
289 for (i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
291 gmx_hack_simd4_store3_r(v[i], s4[i]);
295 GMX_ALIGNED(real, GMX_SIMD_REAL_WIDTH) buf_aligned[3*GMX_SIMD_REAL_WIDTH];
297 real* buf_aligned = buf;
302 gmx_simd_store_r(buf_aligned + 0*GMX_SIMD_REAL_WIDTH, x);
303 gmx_simd_store_r(buf_aligned + 1*GMX_SIMD_REAL_WIDTH, y);
304 gmx_simd_store_r(buf_aligned + 2*GMX_SIMD_REAL_WIDTH, z);
306 for (i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
308 for (m = 0; m < DIM; m++)
310 v[i][m] = buf_aligned[m*GMX_SIMD_REAL_WIDTH + i];
315 #endif /* GMX_SIMD_HAVE_REAL */
319 int b0; /* first constraint for this task */
320 int b1; /* b1-1 is the last constraint for this task */
321 int ntriangle; /* the number of constraints in triangles */
322 int *triangle; /* the list of triangle constraints */
323 int *tri_bits; /* the bits tell if the matrix element should be used */
324 int tri_alloc; /* allocation size of triangle and tri_bits */
325 int nind; /* number of indices */
326 int *ind; /* constraint index for updating atom data */
327 int nind_r; /* number of indices */
328 int *ind_r; /* constraint index for updating atom data */
329 int ind_nalloc; /* allocation size of ind and ind_r */
330 tensor vir_r_m_dr; /* temporary variable for virial calculation */
331 real dhdlambda; /* temporary variable for lambda derivative */
332 real *simd_buf; /* Aligned work array for SIMD */
335 typedef struct gmx_lincsdata {
336 int ncg; /* the global number of constraints */
337 int ncg_flex; /* the global number of flexible constraints */
338 int ncg_triangle; /* the global number of constraints in triangles */
339 int nIter; /* the number of iterations */
340 int nOrder; /* the order of the matrix expansion */
341 int max_connect; /* the maximum number of constrains connected to a single atom */
343 int nc_real; /* the number of real constraints */
344 int nc; /* the number of constraints including padding for SIMD */
345 int nc_alloc; /* the number we allocated memory for */
346 int ncc; /* the number of constraint connections */
347 int ncc_alloc; /* the number we allocated memory for */
348 real matlam; /* the FE lambda value used for filling blc and blmf */
349 int *con_index; /* mapping from topology to LINCS constraints */
350 real *bllen0; /* the reference distance in topology A */
351 real *ddist; /* the reference distance in top B - the r.d. in top A */
352 int *bla; /* the atom pairs involved in the constraints */
353 real *blc; /* 1/sqrt(invmass1 + invmass2) */
354 real *blc1; /* as blc, but with all masses 1 */
355 int *blnr; /* index into blbnb and blmf */
356 int *blbnb; /* list of constraint connections */
357 int ntriangle; /* the local number of constraints in triangles */
358 int ncc_triangle; /* the number of constraint connections in triangles */
359 gmx_bool bCommIter; /* communicate before each LINCS interation */
360 real *blmf; /* matrix of mass factors for constraint connections */
361 real *blmf1; /* as blmf, but with all masses 1 */
362 real *bllen; /* the reference bond length */
363 int *nlocat; /* the local atom count per constraint, can be NULL */
365 int ntask; /* The number of tasks = #threads for LINCS */
366 lincs_task_t *task; /* LINCS thread division */
367 gmx_bitmask_t *atf; /* atom flags for thread parallelization */
368 int atf_nalloc; /* allocation size of atf */
369 gmx_bool bTaskDep; /* are the LINCS tasks interdependent? */
370 /* arrays for temporary storage in the LINCS algorithm */
377 real *mlambda; /* the Lagrange multipliers * -1 */
378 /* storage for the constraint RMS relative deviation output */
382 /* Define simd_width for memory allocation used for SIMD code */
384 static const int simd_width = GMX_SIMD_REAL_WIDTH;
386 static const int simd_width = 1;
388 /* We can't use small memory alignments on many systems, so use min 64 bytes*/
389 static const int align_bytes = std::max<int>(64, simd_width*sizeof(real));
391 real *lincs_rmsd_data(struct gmx_lincsdata *lincsd)
393 return lincsd->rmsd_data;
396 real lincs_rmsd(struct gmx_lincsdata *lincsd, gmx_bool bSD2)
398 if (lincsd->rmsd_data[0] > 0)
400 return sqrt(lincsd->rmsd_data[bSD2 ? 2 : 1]/lincsd->rmsd_data[0]);
408 /* Do a set of nrec LINCS matrix multiplications.
409 * This function will return with up to date thread-local
410 * constraint data, without an OpenMP barrier.
412 static void lincs_matrix_expand(const struct gmx_lincsdata *lincsd,
413 const lincs_task_t *li_task,
415 real *rhs1, real *rhs2, real *sol)
417 int b0, b1, nrec, rec;
418 const int *blnr = lincsd->blnr;
419 const int *blbnb = lincsd->blbnb;
423 nrec = lincsd->nOrder;
425 for (rec = 0; rec < nrec; rec++)
429 if (lincsd->bTaskDep)
433 for (b = b0; b < b1; b++)
439 for (n = blnr[b]; n < blnr[b+1]; n++)
441 mvb = mvb + blcc[n]*rhs1[blbnb[n]];
444 sol[b] = sol[b] + mvb;
452 } /* nrec*(ncons+2*nrtot) flops */
454 if (lincsd->ntriangle > 0)
456 /* Perform an extra nrec recursions for only the constraints
457 * involved in rigid triangles.
458 * In this way their accuracy should come close to those of the other
459 * constraints, since traingles of constraints can produce eigenvalues
460 * around 0.7, while the effective eigenvalue for bond constraints
461 * is around 0.4 (and 0.7*0.7=0.5).
464 if (lincsd->bTaskDep)
466 /* We need a barrier here, since other threads might still be
467 * reading the contents of rhs1 and/o rhs2.
468 * We could avoid this barrier by introducing two extra rhs
469 * arrays for the triangle constraints only.
474 /* Constraints involved in a triangle are ensured to be in the same
475 * LINCS task. This means no barriers are required during the extra
476 * iterations for the triangle constraints.
478 const int *triangle = li_task->triangle;
479 const int *tri_bits = li_task->tri_bits;
481 for (rec = 0; rec < nrec; rec++)
485 for (tb = 0; tb < li_task->ntriangle; tb++)
487 int b, bits, nr0, nr1, n;
495 for (n = nr0; n < nr1; n++)
497 if (bits & (1 << (n - nr0)))
499 mvb = mvb + blcc[n]*rhs1[blbnb[n]];
503 sol[b] = sol[b] + mvb;
511 } /* nrec*(ntriangle + ncc_triangle*2) flops */
515 static void lincs_update_atoms_noind(int ncons, const int *bla,
517 const real *fac, rvec *r,
522 real mvb, im1, im2, tmp0, tmp1, tmp2;
526 for (b = 0; b < ncons; b++)
542 } /* 16 ncons flops */
546 for (b = 0; b < ncons; b++)
564 static void lincs_update_atoms_ind(int ncons, const int *ind, const int *bla,
566 const real *fac, rvec *r,
571 real mvb, im1, im2, tmp0, tmp1, tmp2;
575 for (bi = 0; bi < ncons; bi++)
592 } /* 16 ncons flops */
596 for (bi = 0; bi < ncons; bi++)
611 } /* 16 ncons flops */
615 static void lincs_update_atoms(struct gmx_lincsdata *li, int th,
617 const real *fac, rvec *r,
623 /* Single thread, we simply update for all constraints */
624 lincs_update_atoms_noind(li->nc_real,
625 li->bla, prefac, fac, r, invmass, x);
629 /* Update the atom vector components for our thread local
630 * constraints that only access our local atom range.
631 * This can be done without a barrier.
633 lincs_update_atoms_ind(li->task[th].nind, li->task[th].ind,
634 li->bla, prefac, fac, r, invmass, x);
636 if (li->task[li->ntask].nind > 0)
638 /* Update the constraints that operate on atoms
639 * in multiple thread atom blocks on the master thread.
644 lincs_update_atoms_ind(li->task[li->ntask].nind,
645 li->task[li->ntask].ind,
646 li->bla, prefac, fac, r, invmass, x);
653 /* Calculate the constraint distance vectors r to project on from x.
654 * Determine the right-hand side of the matrix equation using quantity f.
655 * This function only differs from calc_dr_x_xp_simd below in that
656 * no constraint length is subtracted and no PBC is used for f.
658 static void gmx_simdcall
659 calc_dr_x_f_simd(int b0,
662 const rvec * gmx_restrict x,
663 const rvec * gmx_restrict f,
664 const real * gmx_restrict blc,
665 const pbc_simd_t * pbc_simd,
666 real * gmx_restrict vbuf1,
667 real * gmx_restrict vbuf2,
668 rvec * gmx_restrict r,
669 real * gmx_restrict rhs,
670 real * gmx_restrict sol)
674 assert(b0 % GMX_SIMD_REAL_WIDTH == 0);
676 for (bs = b0; bs < b1; bs += GMX_SIMD_REAL_WIDTH)
678 gmx_simd_real_t rx_S, ry_S, rz_S, n2_S, il_S;
679 gmx_simd_real_t fx_S, fy_S, fz_S, ip_S, rhs_S;
681 gmx_hack_simd_gather_rvec_dist_pair_index(x, bla + bs*2, vbuf1,
682 &rx_S, &ry_S, &rz_S);
684 pbc_correct_dx_simd(&rx_S, &ry_S, &rz_S, pbc_simd);
686 n2_S = gmx_simd_norm2_r(rx_S, ry_S, rz_S);
687 il_S = gmx_simd_invsqrt_r(n2_S);
689 rx_S = gmx_simd_mul_r(rx_S, il_S);
690 ry_S = gmx_simd_mul_r(ry_S, il_S);
691 rz_S = gmx_simd_mul_r(rz_S, il_S);
693 gmx_simd_store_vec_to_rvec(rx_S, ry_S, rz_S, vbuf1, r + bs);
695 gmx_hack_simd_gather_rvec_dist_pair_index(f, bla + bs*2, vbuf2,
696 &fx_S, &fy_S, &fz_S);
698 ip_S = gmx_simd_iprod_r(rx_S, ry_S, rz_S,
701 rhs_S = gmx_simd_mul_r(gmx_simd_load_r(blc + bs), ip_S);
703 gmx_simd_store_r(rhs + bs, rhs_S);
704 gmx_simd_store_r(sol + bs, rhs_S);
707 #endif /* LINCS_SIMD */
709 /* LINCS projection, works on derivatives of the coordinates */
710 static void do_lincsp(rvec *x, rvec *f, rvec *fp, t_pbc *pbc,
711 struct gmx_lincsdata *lincsd, int th,
713 int econq, gmx_bool bCalcDHDL,
714 gmx_bool bCalcVir, tensor rmdf)
717 int *bla, *blnr, *blbnb;
719 real *blc, *blmf, *blcc, *rhs1, *rhs2, *sol;
721 b0 = lincsd->task[th].b0;
722 b1 = lincsd->task[th].b1;
727 blbnb = lincsd->blbnb;
728 if (econq != econqForce)
730 /* Use mass-weighted parameters */
736 /* Use non mass-weighted parameters */
738 blmf = lincsd->blmf1;
740 blcc = lincsd->tmpncc;
747 /* This SIMD code does the same as the plain-C code after the #else.
748 * The only difference is that we always call pbc code, as with SIMD
749 * the overhead of pbc computation (when not needed) is small.
753 /* Convert the pbc struct for SIMD */
754 set_pbc_simd(pbc, &pbc_simd);
756 /* Compute normalized x i-j vectors, store in r.
757 * Compute the inner product of r and xp i-j and store in rhs1.
759 calc_dr_x_f_simd(b0, b1, bla, x, f, blc,
761 lincsd->task[th].simd_buf,
762 lincsd->task[th].simd_buf + GMX_SIMD_REAL_WIDTH*DIM,
765 #else /* LINCS_SIMD */
767 /* Compute normalized i-j vectors */
770 for (b = b0; b < b1; b++)
774 pbc_dx_aiuc(pbc, x[bla[2*b]], x[bla[2*b+1]], dx);
780 for (b = b0; b < b1; b++)
784 rvec_sub(x[bla[2*b]], x[bla[2*b+1]], dx);
786 } /* 16 ncons flops */
789 for (b = b0; b < b1; b++)
796 mvb = blc[b]*(r[b][0]*(f[i][0] - f[j][0]) +
797 r[b][1]*(f[i][1] - f[j][1]) +
798 r[b][2]*(f[i][2] - f[j][2]));
804 #endif /* LINCS_SIMD */
806 if (lincsd->bTaskDep)
808 /* We need a barrier, since the matrix construction below
809 * can access entries in r of other threads.
814 /* Construct the (sparse) LINCS matrix */
815 for (b = b0; b < b1; b++)
819 for (n = blnr[b]; n < blnr[b+1]; n++)
821 blcc[n] = blmf[n]*iprod(r[b], r[blbnb[n]]);
824 /* Together: 23*ncons + 6*nrtot flops */
826 lincs_matrix_expand(lincsd, &lincsd->task[th], blcc, rhs1, rhs2, sol);
827 /* nrec*(ncons+2*nrtot) flops */
829 if (econq == econqDeriv_FlexCon)
831 /* We only want to constraint the flexible constraints,
832 * so we mask out the normal ones by setting sol to 0.
834 for (b = b0; b < b1; b++)
836 if (!(lincsd->bllen0[b] == 0 && lincsd->ddist[b] == 0))
843 /* We multiply sol by blc, so we can use lincs_update_atoms for OpenMP */
844 for (b = b0; b < b1; b++)
849 /* When constraining forces, we should not use mass weighting,
850 * so we pass invmass=NULL, which results in the use of 1 for all atoms.
852 lincs_update_atoms(lincsd, th, 1.0, sol, r,
853 (econq != econqForce) ? invmass : NULL, fp);
860 for (b = b0; b < b1; b++)
862 dhdlambda -= sol[b]*lincsd->ddist[b];
865 lincsd->task[th].dhdlambda = dhdlambda;
870 /* Constraint virial,
871 * determines sum r_bond x delta f,
872 * where delta f is the constraint correction
873 * of the quantity that is being constrained.
875 for (b = b0; b < b1; b++)
880 mvb = lincsd->bllen[b]*sol[b];
881 for (i = 0; i < DIM; i++)
884 for (j = 0; j < DIM; j++)
886 rmdf[i][j] += tmp1*r[b][j];
889 } /* 23 ncons flops */
894 /* Calculate the constraint distance vectors r to project on from x.
895 * Determine the right-hand side of the matrix equation using coordinates xp.
897 static void gmx_simdcall
898 calc_dr_x_xp_simd(int b0,
901 const rvec * gmx_restrict x,
902 const rvec * gmx_restrict xp,
903 const real * gmx_restrict bllen,
904 const real * gmx_restrict blc,
905 const pbc_simd_t * pbc_simd,
906 real * gmx_restrict vbuf1,
907 real * gmx_restrict vbuf2,
908 rvec * gmx_restrict r,
909 real * gmx_restrict rhs,
910 real * gmx_restrict sol)
914 assert(b0 % GMX_SIMD_REAL_WIDTH == 0);
916 for (bs = b0; bs < b1; bs += GMX_SIMD_REAL_WIDTH)
918 gmx_simd_real_t rx_S, ry_S, rz_S, n2_S, il_S;
919 gmx_simd_real_t rxp_S, ryp_S, rzp_S, ip_S, rhs_S;
921 gmx_hack_simd_gather_rvec_dist_pair_index(x, bla + bs*2, vbuf1,
922 &rx_S, &ry_S, &rz_S);
924 pbc_correct_dx_simd(&rx_S, &ry_S, &rz_S, pbc_simd);
926 n2_S = gmx_simd_norm2_r(rx_S, ry_S, rz_S);
927 il_S = gmx_simd_invsqrt_r(n2_S);
929 rx_S = gmx_simd_mul_r(rx_S, il_S);
930 ry_S = gmx_simd_mul_r(ry_S, il_S);
931 rz_S = gmx_simd_mul_r(rz_S, il_S);
933 gmx_simd_store_vec_to_rvec(rx_S, ry_S, rz_S, vbuf1, r + bs);
935 gmx_hack_simd_gather_rvec_dist_pair_index(xp, bla + bs*2, vbuf2,
936 &rxp_S, &ryp_S, &rzp_S);
938 pbc_correct_dx_simd(&rxp_S, &ryp_S, &rzp_S, pbc_simd);
940 ip_S = gmx_simd_iprod_r(rx_S, ry_S, rz_S,
941 rxp_S, ryp_S, rzp_S);
943 rhs_S = gmx_simd_mul_r(gmx_simd_load_r(blc + bs),
944 gmx_simd_sub_r(ip_S, gmx_simd_load_r(bllen + bs)));
946 gmx_simd_store_r(rhs + bs, rhs_S);
947 gmx_simd_store_r(sol + bs, rhs_S);
950 #endif /* LINCS_SIMD */
952 /* Determine the distances and right-hand side for the next iteration */
953 static void calc_dist_iter(int b0,
956 const rvec * gmx_restrict xp,
957 const real * gmx_restrict bllen,
958 const real * gmx_restrict blc,
961 real * gmx_restrict rhs,
962 real * gmx_restrict sol,
967 for (b = b0; b < b1; b++)
969 real len, len2, dlen2, mvb;
975 pbc_dx_aiuc(pbc, xp[bla[2*b]], xp[bla[2*b+1]], dx);
979 rvec_sub(xp[bla[2*b]], xp[bla[2*b+1]], dx);
982 dlen2 = 2*len2 - norm2(dx);
983 if (dlen2 < wfac*len2)
985 /* not race free - see detailed comment in caller */
990 mvb = blc[b]*(len - dlen2*gmx_invsqrt(dlen2));
998 } /* 20*ncons flops */
1002 /* As the function above, but using SIMD intrinsics */
1003 static void gmx_simdcall
1004 calc_dist_iter_simd(int b0,
1007 const rvec * gmx_restrict x,
1008 const real * gmx_restrict bllen,
1009 const real * gmx_restrict blc,
1010 const pbc_simd_t * pbc_simd,
1012 real * gmx_restrict vbuf,
1013 real * gmx_restrict rhs,
1014 real * gmx_restrict sol,
1017 gmx_simd_real_t min_S = gmx_simd_set1_r(GMX_REAL_MIN);
1018 gmx_simd_real_t two_S = gmx_simd_set1_r(2.0);
1019 gmx_simd_real_t wfac_S = gmx_simd_set1_r(wfac);
1020 gmx_simd_bool_t warn_S;
1024 assert(b0 % GMX_SIMD_REAL_WIDTH == 0);
1026 /* Initialize all to FALSE */
1027 warn_S = gmx_simd_cmplt_r(two_S, gmx_simd_setzero_r());
1029 for (bs = b0; bs < b1; bs += GMX_SIMD_REAL_WIDTH)
1031 gmx_simd_real_t rx_S, ry_S, rz_S, n2_S;
1032 gmx_simd_real_t len_S, len2_S, dlen2_S, lc_S, blc_S;
1034 gmx_hack_simd_gather_rvec_dist_pair_index(x, bla + bs*2, vbuf,
1035 &rx_S, &ry_S, &rz_S);
1037 pbc_correct_dx_simd(&rx_S, &ry_S, &rz_S, pbc_simd);
1039 n2_S = gmx_simd_norm2_r(rx_S, ry_S, rz_S);
1041 len_S = gmx_simd_load_r(bllen + bs);
1042 len2_S = gmx_simd_mul_r(len_S, len_S);
1044 dlen2_S = gmx_simd_fmsub_r(two_S, len2_S, n2_S);
1046 warn_S = gmx_simd_or_b(warn_S,
1047 gmx_simd_cmplt_r(dlen2_S,
1048 gmx_simd_mul_r(wfac_S, len2_S)));
1050 /* Avoid 1/0 by taking the max with REAL_MIN.
1051 * Note: when dlen2 is close to zero (90 degree constraint rotation),
1052 * the accuracy of the algorithm is no longer relevant.
1054 dlen2_S = gmx_simd_max_r(dlen2_S, min_S);
1056 lc_S = gmx_simd_fnmadd_r(dlen2_S, gmx_simd_invsqrt_r(dlen2_S), len_S);
1058 blc_S = gmx_simd_load_r(blc + bs);
1060 lc_S = gmx_simd_mul_r(blc_S, lc_S);
1062 gmx_simd_store_r(rhs + bs, lc_S);
1063 gmx_simd_store_r(sol + bs, lc_S);
1066 if (gmx_simd_anytrue_b(warn_S))
1071 #endif /* LINCS_SIMD */
1073 static void do_lincs(rvec *x, rvec *xp, matrix box, t_pbc *pbc,
1074 struct gmx_lincsdata *lincsd, int th,
1075 const real *invmass,
1078 real wangle, gmx_bool *bWarn,
1079 real invdt, rvec * gmx_restrict v,
1080 gmx_bool bCalcVir, tensor vir_r_m_dr)
1082 int b0, b1, b, i, j, n, iter;
1083 int *bla, *blnr, *blbnb;
1085 real *blc, *blmf, *bllen, *blcc, *rhs1, *rhs2, *sol, *blc_sol, *mlambda;
1088 b0 = lincsd->task[th].b0;
1089 b1 = lincsd->task[th].b1;
1093 blnr = lincsd->blnr;
1094 blbnb = lincsd->blbnb;
1096 blmf = lincsd->blmf;
1097 bllen = lincsd->bllen;
1098 blcc = lincsd->tmpncc;
1099 rhs1 = lincsd->tmp1;
1100 rhs2 = lincsd->tmp2;
1102 blc_sol = lincsd->tmp4;
1103 mlambda = lincsd->mlambda;
1104 nlocat = lincsd->nlocat;
1108 /* This SIMD code does the same as the plain-C code after the #else.
1109 * The only difference is that we always call pbc code, as with SIMD
1110 * the overhead of pbc computation (when not needed) is small.
1112 pbc_simd_t pbc_simd;
1114 /* Convert the pbc struct for SIMD */
1115 set_pbc_simd(pbc, &pbc_simd);
1117 /* Compute normalized x i-j vectors, store in r.
1118 * Compute the inner product of r and xp i-j and store in rhs1.
1120 calc_dr_x_xp_simd(b0, b1, bla, x, xp, bllen, blc,
1122 lincsd->task[th].simd_buf,
1123 lincsd->task[th].simd_buf + GMX_SIMD_REAL_WIDTH*DIM,
1126 #else /* LINCS_SIMD */
1130 /* Compute normalized i-j vectors */
1131 for (b = b0; b < b1; b++)
1136 pbc_dx_aiuc(pbc, x[bla[2*b]], x[bla[2*b+1]], dx);
1139 pbc_dx_aiuc(pbc, xp[bla[2*b]], xp[bla[2*b+1]], dx);
1140 mvb = blc[b]*(iprod(r[b], dx) - bllen[b]);
1147 /* Compute normalized i-j vectors */
1148 for (b = b0; b < b1; b++)
1150 real tmp0, tmp1, tmp2, rlen, mvb;
1154 tmp0 = x[i][0] - x[j][0];
1155 tmp1 = x[i][1] - x[j][1];
1156 tmp2 = x[i][2] - x[j][2];
1157 rlen = gmx_invsqrt(tmp0*tmp0 + tmp1*tmp1 + tmp2*tmp2);
1158 r[b][0] = rlen*tmp0;
1159 r[b][1] = rlen*tmp1;
1160 r[b][2] = rlen*tmp2;
1161 /* 16 ncons flops */
1165 mvb = blc[b]*(r[b][0]*(xp[i][0] - xp[j][0]) +
1166 r[b][1]*(xp[i][1] - xp[j][1]) +
1167 r[b][2]*(xp[i][2] - xp[j][2]) - bllen[b]);
1172 /* Together: 26*ncons + 6*nrtot flops */
1175 #endif /* LINCS_SIMD */
1177 if (lincsd->bTaskDep)
1179 /* We need a barrier, since the matrix construction below
1180 * can access entries in r of other threads.
1185 /* Construct the (sparse) LINCS matrix */
1186 for (b = b0; b < b1; b++)
1188 for (n = blnr[b]; n < blnr[b+1]; n++)
1190 blcc[n] = blmf[n]*iprod(r[b], r[blbnb[n]]);
1193 /* Together: 26*ncons + 6*nrtot flops */
1195 lincs_matrix_expand(lincsd, &lincsd->task[th], blcc, rhs1, rhs2, sol);
1196 /* nrec*(ncons+2*nrtot) flops */
1199 for (b = b0; b < b1; b++)
1201 mlambda[b] = blc[b]*sol[b];
1204 for (b = b0; b < b1; b += GMX_SIMD_REAL_WIDTH)
1206 gmx_simd_store_r(mlambda + b,
1207 gmx_simd_mul_r(gmx_simd_load_r(blc + b),
1208 gmx_simd_load_r(sol + b)));
1212 /* Update the coordinates */
1213 lincs_update_atoms(lincsd, th, 1.0, mlambda, r, invmass, xp);
1216 ******** Correction for centripetal effects ********
1221 wfac = cos(DEG2RAD*wangle);
1224 for (iter = 0; iter < lincsd->nIter; iter++)
1226 if ((lincsd->bCommIter && DOMAINDECOMP(cr) && cr->dd->constraints))
1231 /* Communicate the corrected non-local coordinates */
1232 if (DOMAINDECOMP(cr))
1234 dd_move_x_constraints(cr->dd, box, xp, NULL, FALSE);
1239 else if (lincsd->bTaskDep)
1245 calc_dist_iter_simd(b0, b1, bla, xp, bllen, blc, &pbc_simd, wfac,
1246 lincsd->task[th].simd_buf, rhs1, sol, bWarn);
1248 calc_dist_iter(b0, b1, bla, xp, bllen, blc, pbc, wfac,
1250 /* 20*ncons flops */
1253 lincs_matrix_expand(lincsd, &lincsd->task[th], blcc, rhs1, rhs2, sol);
1254 /* nrec*(ncons+2*nrtot) flops */
1257 for (b = b0; b < b1; b++)
1261 mvb = blc[b]*sol[b];
1266 for (b = b0; b < b1; b += GMX_SIMD_REAL_WIDTH)
1268 gmx_simd_real_t mvb;
1270 mvb = gmx_simd_mul_r(gmx_simd_load_r(blc + b),
1271 gmx_simd_load_r(sol + b));
1272 gmx_simd_store_r(blc_sol + b, mvb);
1273 gmx_simd_store_r(mlambda + b,
1274 gmx_simd_add_r(gmx_simd_load_r(mlambda + b), mvb));
1278 /* Update the coordinates */
1279 lincs_update_atoms(lincsd, th, 1.0, blc_sol, r, invmass, xp);
1281 /* nit*ncons*(37+9*nrec) flops */
1285 /* Update the velocities */
1286 lincs_update_atoms(lincsd, th, invdt, mlambda, r, invmass, v);
1287 /* 16 ncons flops */
1290 if (nlocat != NULL && (bCalcDHDL || bCalcVir))
1292 if (lincsd->bTaskDep)
1294 /* In lincs_update_atoms threads might cross-read mlambda */
1298 /* Only account for local atoms */
1299 for (b = b0; b < b1; b++)
1301 mlambda[b] *= 0.5*nlocat[b];
1310 for (b = b0; b < b1; b++)
1312 /* Note that this this is dhdl*dt^2, the dt^2 factor is corrected
1313 * later after the contributions are reduced over the threads.
1315 dhdl -= lincsd->mlambda[b]*lincsd->ddist[b];
1317 lincsd->task[th].dhdlambda = dhdl;
1322 /* Constraint virial */
1323 for (b = b0; b < b1; b++)
1327 tmp0 = -bllen[b]*mlambda[b];
1328 for (i = 0; i < DIM; i++)
1330 tmp1 = tmp0*r[b][i];
1331 for (j = 0; j < DIM; j++)
1333 vir_r_m_dr[i][j] -= tmp1*r[b][j];
1336 } /* 22 ncons flops */
1340 * 26*ncons + 6*nrtot + nrec*(ncons+2*nrtot)
1341 * + nit * (20*ncons + nrec*(ncons+2*nrtot) + 17 ncons)
1343 * (26+nrec)*ncons + (6+2*nrec)*nrtot
1344 * + nit * ((37+nrec)*ncons + 2*nrec*nrtot)
1346 * (63+nrec)*ncons + (6+4*nrec)*nrtot
1350 /* Sets the elements in the LINCS matrix for task li_task */
1351 static void set_lincs_matrix_task(struct gmx_lincsdata *li,
1352 lincs_task_t *li_task,
1353 const real *invmass,
1358 /* Construct the coupling coefficient matrix blmf */
1359 li_task->ntriangle = 0;
1361 for (i = li_task->b0; i < li_task->b1; i++)
1366 a2 = li->bla[2*i+1];
1367 for (n = li->blnr[i]; (n < li->blnr[i+1]); n++)
1369 int k, sign, center, end;
1373 /* If we are using multiple, independent tasks for LINCS,
1374 * the calls to check_assign_connected should have
1375 * put all connected constraints in our task.
1377 assert(li->bTaskDep || (k >= li_task->b0 && k < li_task->b1));
1379 if (a1 == li->bla[2*k] || a2 == li->bla[2*k+1])
1387 if (a1 == li->bla[2*k] || a1 == li->bla[2*k+1])
1397 li->blmf[n] = sign*invmass[center]*li->blc[i]*li->blc[k];
1398 li->blmf1[n] = sign*0.5;
1399 if (li->ncg_triangle > 0)
1403 /* Look for constraint triangles */
1404 for (nk = li->blnr[k]; (nk < li->blnr[k+1]); nk++)
1407 if (kk != i && kk != k &&
1408 (li->bla[2*kk] == end || li->bla[2*kk+1] == end))
1410 /* If we are using multiple tasks for LINCS,
1411 * the calls to check_assign_triangle should have
1412 * put all constraints in the triangle in our task.
1414 assert(k >= li_task->b0 && k < li_task->b1);
1415 assert(kk >= li_task->b0 && kk < li_task->b1);
1417 if (li_task->ntriangle == 0 ||
1418 li_task->triangle[li_task->ntriangle - 1] < i)
1420 /* Add this constraint to the triangle list */
1421 li_task->triangle[li_task->ntriangle] = i;
1422 li_task->tri_bits[li_task->ntriangle] = 0;
1423 li_task->ntriangle++;
1424 if (li->blnr[i+1] - li->blnr[i] > static_cast<int>(sizeof(li_task->tri_bits[0])*8 - 1))
1426 gmx_fatal(FARGS, "A constraint is connected to %d constraints, this is more than the %d allowed for constraints participating in triangles",
1427 li->blnr[i+1] - li->blnr[i],
1428 sizeof(li_task->tri_bits[0])*8-1);
1431 li_task->tri_bits[li_task->ntriangle-1] |= (1 << (n - li->blnr[i]));
1440 /* Sets the elements in the LINCS matrix */
1441 void set_lincs_matrix(struct gmx_lincsdata *li, real *invmass, real lambda)
1444 const real invsqrt2 = 0.7071067811865475244;
1446 for (i = 0; (i < li->nc); i++)
1451 a2 = li->bla[2*i+1];
1452 li->blc[i] = gmx_invsqrt(invmass[a1] + invmass[a2]);
1453 li->blc1[i] = invsqrt2;
1456 /* Construct the coupling coefficient matrix blmf */
1457 int th, ntriangle = 0, ncc_triangle = 0;
1458 #pragma omp parallel for reduction(+: ntriangle, ncc_triangle) num_threads(li->ntask) schedule(static)
1459 for (th = 0; th < li->ntask; th++)
1461 set_lincs_matrix_task(li, &li->task[th], invmass, &ncc_triangle);
1462 ntriangle = li->task[th].ntriangle;
1464 li->ntriangle = ntriangle;
1465 li->ncc_triangle = ncc_triangle;
1469 fprintf(debug, "Of the %d constraints %d participate in triangles\n",
1470 li->nc, li->ntriangle);
1471 fprintf(debug, "There are %d couplings of which %d in triangles\n",
1472 li->ncc, li->ncc_triangle);
1476 * so we know with which lambda value the masses have been set.
1478 li->matlam = lambda;
1481 static int count_triangle_constraints(t_ilist *ilist, t_blocka *at2con)
1483 int ncon1, ncon_tot;
1484 int c0, a00, a01, n1, c1, a10, a11, ac1, n2, c2, a20, a21;
1487 t_iatom *ia1, *ia2, *iap;
1489 ncon1 = ilist[F_CONSTR].nr/3;
1490 ncon_tot = ncon1 + ilist[F_CONSTRNC].nr/3;
1492 ia1 = ilist[F_CONSTR].iatoms;
1493 ia2 = ilist[F_CONSTRNC].iatoms;
1496 for (c0 = 0; c0 < ncon_tot; c0++)
1499 iap = constr_iatomptr(ncon1, ia1, ia2, c0);
1502 for (n1 = at2con->index[a01]; n1 < at2con->index[a01+1]; n1++)
1507 iap = constr_iatomptr(ncon1, ia1, ia2, c1);
1518 for (n2 = at2con->index[ac1]; n2 < at2con->index[ac1+1]; n2++)
1521 if (c2 != c0 && c2 != c1)
1523 iap = constr_iatomptr(ncon1, ia1, ia2, c2);
1526 if (a20 == a00 || a21 == a00)
1540 return ncon_triangle;
1543 static gmx_bool more_than_two_sequential_constraints(const t_ilist *ilist,
1544 const t_blocka *at2con)
1546 t_iatom *ia1, *ia2, *iap;
1547 int ncon1, ncon_tot, c;
1549 gmx_bool bMoreThanTwoSequentialConstraints;
1551 ncon1 = ilist[F_CONSTR].nr/3;
1552 ncon_tot = ncon1 + ilist[F_CONSTRNC].nr/3;
1554 ia1 = ilist[F_CONSTR].iatoms;
1555 ia2 = ilist[F_CONSTRNC].iatoms;
1557 bMoreThanTwoSequentialConstraints = FALSE;
1558 for (c = 0; c < ncon_tot && !bMoreThanTwoSequentialConstraints; c++)
1560 iap = constr_iatomptr(ncon1, ia1, ia2, c);
1563 /* Check if this constraint has constraints connected at both atoms */
1564 if (at2con->index[a1+1] - at2con->index[a1] > 1 &&
1565 at2con->index[a2+1] - at2con->index[a2] > 1)
1567 bMoreThanTwoSequentialConstraints = TRUE;
1571 return bMoreThanTwoSequentialConstraints;
1574 static int int_comp(const void *a, const void *b)
1576 return (*(int *)a) - (*(int *)b);
1579 gmx_lincsdata_t init_lincs(FILE *fplog, gmx_mtop_t *mtop,
1580 int nflexcon_global, t_blocka *at2con,
1581 gmx_bool bPLINCS, int nIter, int nProjOrder)
1583 struct gmx_lincsdata *li;
1585 gmx_moltype_t *molt;
1586 gmx_bool bMoreThanTwoSeq;
1590 fprintf(fplog, "\nInitializing%s LINear Constraint Solver\n",
1591 bPLINCS ? " Parallel" : "");
1597 gmx_mtop_ftype_count(mtop, F_CONSTR) +
1598 gmx_mtop_ftype_count(mtop, F_CONSTRNC);
1599 li->ncg_flex = nflexcon_global;
1602 li->nOrder = nProjOrder;
1604 li->max_connect = 0;
1605 for (mt = 0; mt < mtop->nmoltype; mt++)
1609 molt = &mtop->moltype[mt];
1610 for (a = 0; a < molt->atoms.nr; a++)
1612 li->max_connect = std::max(li->max_connect,
1613 at2con[mt].index[a + 1] - at2con[mt].index[a]);
1617 li->ncg_triangle = 0;
1618 bMoreThanTwoSeq = FALSE;
1619 for (mb = 0; mb < mtop->nmolblock; mb++)
1621 molt = &mtop->moltype[mtop->molblock[mb].type];
1624 mtop->molblock[mb].nmol*
1625 count_triangle_constraints(molt->ilist,
1626 &at2con[mtop->molblock[mb].type]);
1628 if (!bMoreThanTwoSeq &&
1629 more_than_two_sequential_constraints(molt->ilist, &at2con[mtop->molblock[mb].type]))
1631 bMoreThanTwoSeq = TRUE;
1635 /* Check if we need to communicate not only before LINCS,
1636 * but also before each iteration.
1637 * The check for only two sequential constraints is only
1638 * useful for the common case of H-bond only constraints.
1639 * With more effort we could also make it useful for small
1640 * molecules with nr. sequential constraints <= nOrder-1.
1642 li->bCommIter = (bPLINCS && (li->nOrder < 1 || bMoreThanTwoSeq));
1644 if (debug && bPLINCS)
1646 fprintf(debug, "PLINCS communication before each iteration: %d\n",
1650 /* LINCS can run on any number of threads.
1651 * Currently the number is fixed for the whole simulation,
1652 * but it could be set in set_lincs().
1653 * The current constraint to task assignment code can create independent
1654 * tasks only when not more than two constraints are connected sequentially.
1656 li->ntask = gmx_omp_nthreads_get(emntLINCS);
1657 li->bTaskDep = (li->ntask > 1 && bMoreThanTwoSeq);
1660 fprintf(debug, "LINCS: using %d threads, tasks are %sdependent\n",
1661 li->ntask, li->bTaskDep ? "" : "in");
1669 /* Allocate an extra elements for "task-overlap" constraints */
1670 snew(li->task, li->ntask + 1);
1673 #pragma omp parallel for num_threads(li->ntask)
1674 for (th = 0; th < li->ntask; th++)
1676 /* Per thread SIMD load buffer for loading 2 simd_width rvecs */
1677 snew_aligned(li->task[th].simd_buf, 2*simd_width*DIM,
1681 if (bPLINCS || li->ncg_triangle > 0)
1683 please_cite(fplog, "Hess2008a");
1687 please_cite(fplog, "Hess97a");
1692 fprintf(fplog, "The number of constraints is %d\n", li->ncg);
1695 fprintf(fplog, "There are inter charge-group constraints,\n"
1696 "will communicate selected coordinates each lincs iteration\n");
1698 if (li->ncg_triangle > 0)
1701 "%d constraints are involved in constraint triangles,\n"
1702 "will apply an additional matrix expansion of order %d for couplings\n"
1703 "between constraints inside triangles\n",
1704 li->ncg_triangle, li->nOrder);
1711 /* Sets up the work division over the threads */
1712 static void lincs_thread_setup(struct gmx_lincsdata *li, int natoms)
1719 if (natoms > li->atf_nalloc)
1721 li->atf_nalloc = over_alloc_large(natoms);
1722 srenew(li->atf, li->atf_nalloc);
1726 /* Clear the atom flags */
1727 for (a = 0; a < natoms; a++)
1729 bitmask_clear(&atf[a]);
1732 if (li->ntask > BITMASK_SIZE)
1734 gmx_fatal(FARGS, "More than %d threads is not supported for LINCS.", BITMASK_SIZE);
1737 for (th = 0; th < li->ntask; th++)
1739 lincs_task_t *li_task;
1742 li_task = &li->task[th];
1744 /* For each atom set a flag for constraints from each */
1745 for (b = li_task->b0; b < li_task->b1; b++)
1747 bitmask_set_bit(&atf[li->bla[b*2 ]], th);
1748 bitmask_set_bit(&atf[li->bla[b*2 + 1]], th);
1752 #pragma omp parallel for num_threads(li->ntask) schedule(static)
1753 for (th = 0; th < li->ntask; th++)
1755 lincs_task_t *li_task;
1759 li_task = &li->task[th];
1761 if (li_task->b1 - li_task->b0 > li_task->ind_nalloc)
1763 li_task->ind_nalloc = over_alloc_large(li_task->b1-li_task->b0);
1764 srenew(li_task->ind, li_task->ind_nalloc);
1765 srenew(li_task->ind_r, li_task->ind_nalloc);
1768 bitmask_init_low_bits(&mask, th);
1771 li_task->nind_r = 0;
1772 for (b = li_task->b0; b < li_task->b1; b++)
1774 /* We let the constraint with the lowest thread index
1775 * operate on atoms with constraints from multiple threads.
1777 if (bitmask_is_disjoint(atf[li->bla[b*2]], mask) &&
1778 bitmask_is_disjoint(atf[li->bla[b*2+1]], mask))
1780 /* Add the constraint to the local atom update index */
1781 li_task->ind[li_task->nind++] = b;
1785 /* Add the constraint to the rest block */
1786 li_task->ind_r[li_task->nind_r++] = b;
1791 /* We need to copy all constraints which have not be assigned
1792 * to a thread to a separate list which will be handled by one thread.
1794 li_m = &li->task[li->ntask];
1797 for (th = 0; th < li->ntask; th++)
1799 lincs_task_t *li_task;
1802 li_task = &li->task[th];
1804 if (li_m->nind + li_task->nind_r > li_m->ind_nalloc)
1806 li_m->ind_nalloc = over_alloc_large(li_m->nind+li_task->nind_r);
1807 srenew(li_m->ind, li_m->ind_nalloc);
1810 for (b = 0; b < li_task->nind_r; b++)
1812 li_m->ind[li_m->nind++] = li_task->ind_r[b];
1817 fprintf(debug, "LINCS thread %d: %d constraints\n",
1824 fprintf(debug, "LINCS thread r: %d constraints\n",
1829 /* There is no realloc with alignment, so here we make one for reals.
1830 * Note that this function does not preserve the contents of the memory.
1832 static void resize_real_aligned(real **ptr, int nelem)
1834 sfree_aligned(*ptr);
1835 snew_aligned(*ptr, nelem, align_bytes);
1838 static void assign_constraint(struct gmx_lincsdata *li,
1839 int constraint_index,
1841 real lenA, real lenB,
1842 const t_blocka *at2con)
1848 /* Make an mapping of local topology constraint index to LINCS index */
1849 li->con_index[constraint_index] = con;
1851 li->bllen0[con] = lenA;
1852 li->ddist[con] = lenB - lenA;
1853 /* Set the length to the topology A length */
1854 li->bllen[con] = lenA;
1855 li->bla[2*con] = a1;
1856 li->bla[2*con+1] = a2;
1858 /* Make space in the constraint connection matrix for constraints
1859 * connected to both end of the current constraint.
1862 at2con->index[a1 + 1] - at2con->index[a1] - 1 +
1863 at2con->index[a2 + 1] - at2con->index[a2] - 1;
1865 li->blnr[con + 1] = li->ncc;
1867 /* Increase the constraint count */
1871 /* Check if constraint with topology index constraint_index is connected
1872 * to other constraints, and if so add those connected constraints to our task.
1874 static void check_assign_connected(struct gmx_lincsdata *li,
1875 const t_iatom *iatom,
1879 const t_blocka *at2con)
1881 /* Currently this function only supports constraint groups
1882 * in which all constraints share at least one atom
1883 * (e.g. H-bond constraints).
1884 * Check both ends of the current constraint for
1885 * connected constraints. We need to assign those
1890 for (end = 0; end < 2; end++)
1894 a = (end == 0 ? a1 : a2);
1896 for (k = at2con->index[a]; k < at2con->index[a + 1]; k++)
1901 /* Check if constraint cc has not yet been assigned */
1902 if (li->con_index[cc] == -1)
1908 lenA = idef->iparams[type].constr.dA;
1909 lenB = idef->iparams[type].constr.dB;
1911 if (bDynamics || lenA != 0 || lenB != 0)
1913 assign_constraint(li, cc, iatom[3*cc + 1], iatom[3*cc + 2], lenA, lenB, at2con);
1920 /* Check if constraint with topology index constraint_index is involved
1921 * in a constraint triangle, and if so add the other two constraints
1922 * in the triangle to our task.
1924 static void check_assign_triangle(struct gmx_lincsdata *li,
1925 const t_iatom *iatom,
1928 int constraint_index,
1930 const t_blocka *at2con)
1932 int nca, cc[32], ca[32], k;
1933 int c_triangle[2] = { -1, -1 };
1936 for (k = at2con->index[a1]; k < at2con->index[a1 + 1]; k++)
1941 if (c != constraint_index)
1945 aa1 = iatom[c*3 + 1];
1946 aa2 = iatom[c*3 + 2];
1962 for (k = at2con->index[a2]; k < at2con->index[a2 + 1]; k++)
1967 if (c != constraint_index)
1971 aa1 = iatom[c*3 + 1];
1972 aa2 = iatom[c*3 + 2];
1975 for (i = 0; i < nca; i++)
1979 c_triangle[0] = cc[i];
1986 for (i = 0; i < nca; i++)
1990 c_triangle[0] = cc[i];
1998 if (c_triangle[0] >= 0)
2002 for (end = 0; end < 2; end++)
2004 /* Check if constraint c_triangle[end] has not yet been assigned */
2005 if (li->con_index[c_triangle[end]] == -1)
2010 i = c_triangle[end]*3;
2012 lenA = idef->iparams[type].constr.dA;
2013 lenB = idef->iparams[type].constr.dB;
2015 if (bDynamics || lenA != 0 || lenB != 0)
2017 assign_constraint(li, c_triangle[end], iatom[i + 1], iatom[i + 2], lenA, lenB, at2con);
2024 static void set_matrix_indices(struct gmx_lincsdata *li,
2025 const lincs_task_t *li_task,
2026 const t_blocka *at2con,
2027 gmx_bool bSortMatrix)
2031 for (b = li_task->b0; b < li_task->b1; b++)
2036 a2 = li->bla[b*2 + 1];
2039 for (k = at2con->index[a1]; k < at2con->index[a1 + 1]; k++)
2043 concon = li->con_index[at2con->a[k]];
2046 li->blbnb[i++] = concon;
2049 for (k = at2con->index[a2]; k < at2con->index[a2 + 1]; k++)
2053 concon = li->con_index[at2con->a[k]];
2056 li->blbnb[i++] = concon;
2062 /* Order the blbnb matrix to optimize memory access */
2063 qsort(&(li->blbnb[li->blnr[b]]), li->blnr[b + 1] - li->blnr[b],
2064 sizeof(li->blbnb[0]), int_comp);
2069 void set_lincs(t_idef *idef, t_mdatoms *md,
2070 gmx_bool bDynamics, t_commrec *cr,
2071 struct gmx_lincsdata *li)
2073 int natoms, nflexcon;
2076 int i, ncc_alloc_old, ncon_tot;
2081 /* Zero the thread index ranges.
2082 * Otherwise without local constraints we could return with old ranges.
2084 for (i = 0; i < li->ntask; i++)
2088 li->task[i].nind = 0;
2092 li->task[li->ntask].nind = 0;
2095 /* This is the local topology, so there are only F_CONSTR constraints */
2096 if (idef->il[F_CONSTR].nr == 0)
2098 /* There are no constraints,
2099 * we do not need to fill any data structures.
2106 fprintf(debug, "Building the LINCS connectivity\n");
2109 if (DOMAINDECOMP(cr))
2111 if (cr->dd->constraints)
2115 dd_get_constraint_range(cr->dd, &start, &natoms);
2119 natoms = cr->dd->nat_home;
2124 natoms = md->homenr;
2126 at2con = make_at2con(0, natoms, idef->il, idef->iparams, bDynamics,
2129 ncon_tot = idef->il[F_CONSTR].nr/3;
2131 /* Ensure we have enough padding for aligned loads for each thread */
2132 if (ncon_tot + li->ntask*simd_width > li->nc_alloc || li->nc_alloc == 0)
2134 li->nc_alloc = over_alloc_dd(ncon_tot + li->ntask*simd_width);
2135 srenew(li->con_index, li->nc_alloc);
2136 resize_real_aligned(&li->bllen0, li->nc_alloc);
2137 resize_real_aligned(&li->ddist, li->nc_alloc);
2138 srenew(li->bla, 2*li->nc_alloc);
2139 resize_real_aligned(&li->blc, li->nc_alloc);
2140 resize_real_aligned(&li->blc1, li->nc_alloc);
2141 srenew(li->blnr, li->nc_alloc + 1);
2142 resize_real_aligned(&li->bllen, li->nc_alloc);
2143 srenew(li->tmpv, li->nc_alloc);
2144 if (DOMAINDECOMP(cr))
2146 srenew(li->nlocat, li->nc_alloc);
2148 resize_real_aligned(&li->tmp1, li->nc_alloc);
2149 resize_real_aligned(&li->tmp2, li->nc_alloc);
2150 resize_real_aligned(&li->tmp3, li->nc_alloc);
2151 resize_real_aligned(&li->tmp4, li->nc_alloc);
2152 resize_real_aligned(&li->mlambda, li->nc_alloc);
2155 iatom = idef->il[F_CONSTR].iatoms;
2157 ncc_alloc_old = li->ncc_alloc;
2158 li->blnr[0] = li->ncc;
2160 /* Assign the constraints for li->ntask LINCS tasks.
2161 * We target a uniform distribution of constraints over the tasks.
2162 * Note that when flexible constraints are present, but are removed here
2163 * (e.g. because we are doing EM) we get imbalance, but since that doesn't
2164 * happen during normal MD, that's ok.
2166 int ncon_assign, ncon_target, con, th;
2168 /* Determine the number of constraints we need to assign here */
2169 ncon_assign = ncon_tot;
2172 /* With energy minimization, flexible constraints are ignored
2173 * (and thus minimized, as they should be).
2175 ncon_assign -= nflexcon;
2178 /* Set the target constraint count per task to exactly uniform,
2179 * this might be overridden below.
2181 ncon_target = (ncon_assign + li->ntask - 1)/li->ntask;
2183 /* Mark all constraints as unassigned by setting their index to -1 */
2184 for (con = 0; con < ncon_tot; con++)
2186 li->con_index[con] = -1;
2190 for (th = 0; th < li->ntask; th++)
2192 lincs_task_t *li_task;
2194 li_task = &li->task[th];
2197 /* With indepedent tasks we likely have H-bond constraints or constraint
2198 * pairs. The connected constraints will be pulled into the task, so the
2199 * constraints per task will often exceed ncon_target.
2200 * Triangle constraints can also increase the count, but there are
2201 * relatively few of those, so we usually expect to get ncon_target.
2205 /* We round ncon_target to a multiple of GMX_SIMD_WIDTH,
2206 * since otherwise a lot of operations can be wasted.
2207 * There are several ways to round here, we choose the one
2208 * that alternates block sizes, which helps with Intel HT.
2210 ncon_target = ((ncon_assign*(th + 1))/li->ntask - li->nc_real + GMX_SIMD_REAL_WIDTH - 1) & ~(GMX_SIMD_REAL_WIDTH - 1);
2214 /* Continue filling the arrays where we left off with the previous task,
2215 * including padding for SIMD.
2217 li_task->b0 = li->nc;
2219 while (con < ncon_tot && li->nc - li_task->b0 < ncon_target)
2221 if (li->con_index[con] == -1)
2226 type = iatom[3*con];
2227 a1 = iatom[3*con + 1];
2228 a2 = iatom[3*con + 2];
2229 lenA = idef->iparams[type].constr.dA;
2230 lenB = idef->iparams[type].constr.dB;
2231 /* Skip the flexible constraints when not doing dynamics */
2232 if (bDynamics || lenA != 0 || lenB != 0)
2234 assign_constraint(li, con, a1, a2, lenA, lenB, &at2con);
2236 if (li->ntask > 1 && !li->bTaskDep)
2238 /* We can generate independent tasks. Check if we
2239 * need to assign connected constraints to our task.
2241 check_assign_connected(li, iatom, idef, bDynamics,
2244 if (li->ntask > 1 && li->ncg_triangle > 0)
2246 /* Ensure constraints in one triangle are assigned
2249 check_assign_triangle(li, iatom, idef, bDynamics,
2250 con, a1, a2, &at2con);
2258 li_task->b1 = li->nc;
2262 /* Copy the last atom pair indices and lengths for constraints
2263 * up to a multiple of simd_width, such that we can do all
2264 * SIMD operations without having to worry about end effects.
2268 li->nc = ((li_task->b1 + simd_width - 1)/simd_width)*simd_width;
2269 last = li_task->b1 - 1;
2270 for (i = li_task->b1; i < li->nc; i++)
2272 li->bla[i*2 ] = li->bla[last*2 ];
2273 li->bla[i*2 + 1] = li->bla[last*2 + 1];
2274 li->bllen0[i] = li->bllen0[last];
2275 li->ddist[i] = li->ddist[last];
2276 li->bllen[i] = li->bllen[last];
2277 li->blnr[i + 1] = li->blnr[last + 1];
2281 /* Keep track of how many constraints we assigned */
2282 li->nc_real += li_task->b1 - li_task->b0;
2286 fprintf(debug, "LINCS task %d constraints %d - %d\n",
2287 th, li_task->b0, li_task->b1);
2291 assert(li->nc_real == ncon_assign);
2293 gmx_bool bSortMatrix;
2295 /* Without DD we order the blbnb matrix to optimize memory access.
2296 * With DD the overhead of sorting is more than the gain during access.
2298 bSortMatrix = !DOMAINDECOMP(cr);
2300 if (li->ncc > li->ncc_alloc)
2302 li->ncc_alloc = over_alloc_small(li->ncc);
2303 srenew(li->blbnb, li->ncc_alloc);
2306 #pragma omp parallel for num_threads(li->ntask) schedule(static)
2307 for (th = 0; th < li->ntask; th++)
2309 lincs_task_t *li_task;
2311 li_task = &li->task[th];
2313 if (li->ncg_triangle > 0 &&
2314 li_task->b1 - li_task->b0 > li_task->tri_alloc)
2316 /* This is allocating too much, but it is difficult to improve */
2317 li_task->tri_alloc = over_alloc_dd(li_task->b1 - li_task->b0);
2318 srenew(li_task->triangle, li_task->tri_alloc);
2319 srenew(li_task->tri_bits, li_task->tri_alloc);
2322 set_matrix_indices(li, li_task, &at2con, bSortMatrix);
2325 done_blocka(&at2con);
2329 /* Since the matrix is static, we should free some memory */
2330 li->ncc_alloc = li->ncc;
2331 srenew(li->blbnb, li->ncc_alloc);
2334 if (li->ncc_alloc > ncc_alloc_old)
2336 srenew(li->blmf, li->ncc_alloc);
2337 srenew(li->blmf1, li->ncc_alloc);
2338 srenew(li->tmpncc, li->ncc_alloc);
2341 if (DOMAINDECOMP(cr) && dd_constraints_nlocalatoms(cr->dd) != NULL)
2345 nlocat_dd = dd_constraints_nlocalatoms(cr->dd);
2347 /* Convert nlocat from local topology to LINCS constraint indexing */
2348 for (con = 0; con < ncon_tot; con++)
2350 li->nlocat[li->con_index[con]] = nlocat_dd[con];
2360 fprintf(debug, "Number of constraints is %d, padded %d, couplings %d\n",
2361 li->nc_real, li->nc, li->ncc);
2366 lincs_thread_setup(li, md->nr);
2369 set_lincs_matrix(li, md->invmass, md->lambda);
2372 static void lincs_warning(FILE *fplog,
2373 gmx_domdec_t *dd, rvec *x, rvec *xprime, t_pbc *pbc,
2374 int ncons, int *bla, real *bllen, real wangle,
2375 int maxwarn, int *warncount)
2379 real wfac, d0, d1, cosine;
2382 wfac = cos(DEG2RAD*wangle);
2384 sprintf(buf, "bonds that rotated more than %g degrees:\n"
2385 " atom 1 atom 2 angle previous, current, constraint length\n",
2387 fprintf(stderr, "%s", buf);
2390 fprintf(fplog, "%s", buf);
2393 for (b = 0; b < ncons; b++)
2399 pbc_dx_aiuc(pbc, x[i], x[j], v0);
2400 pbc_dx_aiuc(pbc, xprime[i], xprime[j], v1);
2404 rvec_sub(x[i], x[j], v0);
2405 rvec_sub(xprime[i], xprime[j], v1);
2409 cosine = iprod(v0, v1)/(d0*d1);
2412 sprintf(buf, " %6d %6d %5.1f %8.4f %8.4f %8.4f\n",
2413 ddglatnr(dd, i), ddglatnr(dd, j),
2414 RAD2DEG*acos(cosine), d0, d1, bllen[b]);
2415 fprintf(stderr, "%s", buf);
2418 fprintf(fplog, "%s", buf);
2420 if (!gmx_isfinite(d1))
2422 gmx_fatal(FARGS, "Bond length not finite.");
2428 if (*warncount > maxwarn)
2430 too_many_constraint_warnings(econtLINCS, *warncount);
2434 static void cconerr(const struct gmx_lincsdata *lincsd,
2435 rvec *x, t_pbc *pbc,
2436 real *ncons_loc, real *ssd, real *max, int *imax)
2438 const int *bla, *nlocat;
2441 int count, im, task;
2444 bllen = lincsd->bllen;
2445 nlocat = lincsd->nlocat;
2451 for (task = 0; task < lincsd->ntask; task++)
2455 for (b = lincsd->task[task].b0; b < lincsd->task[task].b1; b++)
2462 pbc_dx_aiuc(pbc, x[bla[2*b]], x[bla[2*b+1]], dx);
2466 rvec_sub(x[bla[2*b]], x[bla[2*b+1]], dx);
2469 len = r2*gmx_invsqrt(r2);
2470 d = fabs(len/bllen[b]-1);
2471 if (d > ma && (nlocat == NULL || nlocat[b]))
2483 ssd2 += nlocat[b]*d*d;
2489 *ncons_loc = (nlocat ? 0.5 : 1)*count;
2490 *ssd = (nlocat ? 0.5 : 1)*ssd2;
2495 gmx_bool constrain_lincs(FILE *fplog, gmx_bool bLog, gmx_bool bEner,
2498 struct gmx_lincsdata *lincsd, t_mdatoms *md,
2500 rvec *x, rvec *xprime, rvec *min_proj,
2501 matrix box, t_pbc *pbc,
2502 real lambda, real *dvdlambda,
2503 real invdt, rvec *v,
2504 gmx_bool bCalcVir, tensor vir_r_m_dr,
2507 int maxwarn, int *warncount)
2510 char buf[STRLEN], buf2[22], buf3[STRLEN];
2512 real ncons_loc, p_ssd, p_max = 0;
2514 gmx_bool bOK, bWarn;
2518 /* This boolean should be set by a flag passed to this routine.
2519 * We can also easily check if any constraint length is changed,
2520 * if not dH/dlambda=0 and we can also set the boolean to FALSE.
2522 bCalcDHDL = (ir->efep != efepNO && dvdlambda != NULL);
2524 if (lincsd->nc == 0 && cr->dd == NULL)
2528 lincsd->rmsd_data[0] = 0;
2529 if (ir->eI == eiSD2 && v == NULL)
2537 lincsd->rmsd_data[i] = 0;
2543 if (econq == econqCoord)
2545 /* We can't use bCalcDHDL here, since NULL can be passed for dvdlambda
2546 * also with efep!=fepNO.
2548 if (ir->efep != efepNO)
2550 if (md->nMassPerturbed && lincsd->matlam != md->lambda)
2552 set_lincs_matrix(lincsd, md->invmass, md->lambda);
2555 for (i = 0; i < lincsd->nc; i++)
2557 lincsd->bllen[i] = lincsd->bllen0[i] + lambda*lincsd->ddist[i];
2561 if (lincsd->ncg_flex)
2563 /* Set the flexible constraint lengths to the old lengths */
2566 for (i = 0; i < lincsd->nc; i++)
2568 if (lincsd->bllen[i] == 0)
2570 pbc_dx_aiuc(pbc, x[lincsd->bla[2*i]], x[lincsd->bla[2*i+1]], dx);
2571 lincsd->bllen[i] = norm(dx);
2577 for (i = 0; i < lincsd->nc; i++)
2579 if (lincsd->bllen[i] == 0)
2582 sqrt(distance2(x[lincsd->bla[2*i]],
2583 x[lincsd->bla[2*i+1]]));
2591 cconerr(lincsd, xprime, pbc,
2592 &ncons_loc, &p_ssd, &p_max, &p_imax);
2595 /* This bWarn var can be updated by multiple threads
2596 * at the same time. But as we only need to detect
2597 * if a warning occured or not, this is not an issue.
2601 /* The OpenMP parallel region of constrain_lincs for coords */
2602 #pragma omp parallel num_threads(lincsd->ntask)
2604 int th = gmx_omp_get_thread_num();
2606 clear_mat(lincsd->task[th].vir_r_m_dr);
2608 do_lincs(x, xprime, box, pbc, lincsd, th,
2611 ir->LincsWarnAngle, &bWarn,
2613 th == 0 ? vir_r_m_dr : lincsd->task[th].vir_r_m_dr);
2616 if (bLog && fplog && lincsd->nc > 0)
2618 fprintf(fplog, " Rel. Constraint Deviation: RMS MAX between atoms\n");
2619 fprintf(fplog, " Before LINCS %.6f %.6f %6d %6d\n",
2620 sqrt(p_ssd/ncons_loc), p_max,
2621 ddglatnr(cr->dd, lincsd->bla[2*p_imax]),
2622 ddglatnr(cr->dd, lincsd->bla[2*p_imax+1]));
2626 cconerr(lincsd, xprime, pbc,
2627 &ncons_loc, &p_ssd, &p_max, &p_imax);
2628 /* Check if we are doing the second part of SD */
2629 if (ir->eI == eiSD2 && v == NULL)
2637 lincsd->rmsd_data[0] = ncons_loc;
2638 lincsd->rmsd_data[i] = p_ssd;
2642 lincsd->rmsd_data[0] = 0;
2643 lincsd->rmsd_data[1] = 0;
2644 lincsd->rmsd_data[2] = 0;
2646 if (bLog && fplog && lincsd->nc > 0)
2649 " After LINCS %.6f %.6f %6d %6d\n\n",
2650 sqrt(p_ssd/ncons_loc), p_max,
2651 ddglatnr(cr->dd, lincsd->bla[2*p_imax]),
2652 ddglatnr(cr->dd, lincsd->bla[2*p_imax+1]));
2659 cconerr(lincsd, xprime, pbc,
2660 &ncons_loc, &p_ssd, &p_max, &p_imax);
2663 sprintf(buf3, " in simulation %d", cr->ms->sim);
2669 sprintf(buf, "\nStep %s, time %g (ps) LINCS WARNING%s\n"
2670 "relative constraint deviation after LINCS:\n"
2671 "rms %.6f, max %.6f (between atoms %d and %d)\n",
2672 gmx_step_str(step, buf2), ir->init_t+step*ir->delta_t,
2674 sqrt(p_ssd/ncons_loc), p_max,
2675 ddglatnr(cr->dd, lincsd->bla[2*p_imax]),
2676 ddglatnr(cr->dd, lincsd->bla[2*p_imax+1]));
2679 fprintf(fplog, "%s", buf);
2681 fprintf(stderr, "%s", buf);
2682 lincs_warning(fplog, cr->dd, x, xprime, pbc,
2683 lincsd->nc, lincsd->bla, lincsd->bllen,
2684 ir->LincsWarnAngle, maxwarn, warncount);
2686 bOK = (p_max < 0.5);
2689 if (lincsd->ncg_flex)
2691 for (i = 0; (i < lincsd->nc); i++)
2693 if (lincsd->bllen0[i] == 0 && lincsd->ddist[i] == 0)
2695 lincsd->bllen[i] = 0;
2702 /* The OpenMP parallel region of constrain_lincs for derivatives */
2703 #pragma omp parallel num_threads(lincsd->ntask)
2705 int th = gmx_omp_get_thread_num();
2707 do_lincsp(x, xprime, min_proj, pbc, lincsd, th,
2708 md->invmass, econq, bCalcDHDL,
2709 bCalcVir, th == 0 ? vir_r_m_dr : lincsd->task[th].vir_r_m_dr);
2715 /* Reduce the dH/dlambda contributions over the threads */
2720 for (th = 0; th < lincsd->ntask; th++)
2722 dhdlambda += lincsd->task[th].dhdlambda;
2726 /* dhdlambda contains dH/dlambda*dt^2, correct for this */
2727 /* TODO This should probably use invdt, so that sd integrator scaling works properly */
2728 dhdlambda /= ir->delta_t*ir->delta_t;
2730 *dvdlambda += dhdlambda;
2733 if (bCalcVir && lincsd->ntask > 1)
2735 for (i = 1; i < lincsd->ntask; i++)
2737 m_add(vir_r_m_dr, lincsd->task[i].vir_r_m_dr, vir_r_m_dr);
2741 /* count assuming nit=1 */
2742 inc_nrnb(nrnb, eNR_LINCS, lincsd->nc_real);
2743 inc_nrnb(nrnb, eNR_LINCSMAT, (2+lincsd->nOrder)*lincsd->ncc);
2744 if (lincsd->ntriangle > 0)
2746 inc_nrnb(nrnb, eNR_LINCSMAT, lincsd->nOrder*lincsd->ncc_triangle);
2750 inc_nrnb(nrnb, eNR_CONSTR_V, lincsd->nc_real*2);
2754 inc_nrnb(nrnb, eNR_CONSTR_VIR, lincsd->nc_real);