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, 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.
42 #include "types/commrec.h"
44 #include "gmx_fatal.h"
57 static const t_nrnb_data nbdata[eNRNB] = {
58 /* These are re-used for different NB kernels, since there are so many.
59 * The actual number of flops is set dynamically.
61 { "NB VdW [V&F]", 1 },
63 { "NB Elec. [V&F]", 1 },
64 { "NB Elec. [F]", 1 },
65 { "NB Elec. [W3,V&F]", 1 },
66 { "NB Elec. [W3,F]", 1 },
67 { "NB Elec. [W3-W3,V&F]", 1 },
68 { "NB Elec. [W3-W3,F]", 1 },
69 { "NB Elec. [W4,V&F]", 1 },
70 { "NB Elec. [W4,F]", 1 },
71 { "NB Elec. [W4-W4,V&F]", 1 },
72 { "NB Elec. [W4-W4,F]", 1 },
73 { "NB VdW & Elec. [V&F]", 1 },
74 { "NB VdW & Elec. [F]", 1 },
75 { "NB VdW & Elec. [W3,V&F]", 1 },
76 { "NB VdW & Elec. [W3,F]", 1 },
77 { "NB VdW & Elec. [W3-W3,V&F]", 1 },
78 { "NB VdW & Elec. [W3-W3,F]", 1 },
79 { "NB VdW & Elec. [W4,V&F]", 1 },
80 { "NB VdW & Elec. [W4,F]", 1 },
81 { "NB VdW & Elec. [W4-W4,V&F]", 1 },
82 { "NB VdW & Elec. [W4-W4,F]", 1 },
84 { "NB Generic kernel", 1 },
85 { "NB Generic charge grp kernel", 1 },
86 { "NB Generic AdResS kernel", 1 },
87 { "NB Free energy kernel", 1 },
88 { "NB All-vs-all", 1 },
89 { "NB All-vs-all, GB", 1 },
91 { "Pair Search distance check", 9 }, /* nbnxn pair dist. check */
92 /* nbnxn kernel flops are based on inner-loops without exclusion checks.
93 * Plain Coulomb runs through the RF kernels, except with CUDA.
94 * invsqrt is counted as 6 flops: 1 for _mm_rsqt_ps + 5 for iteration.
95 * The flops are equal for plain-C, x86 SIMD and CUDA, except for:
96 * - plain-C kernel uses one flop more for Coulomb-only (F) than listed
97 * - x86 SIMD LJ geom-comb.rule kernels (fastest) use 2 more flops
98 * - x86 SIMD LJ LB-comb.rule kernels (fast) use 3 (8 for F+E) more flops
99 * - GPU always does exclusions, which requires 2-4 flops, but as invsqrt
100 * is always counted as 6 flops, this roughly compensates.
102 { "NxN RF Elec. + LJ [F]", 38 }, /* nbnxn kernel LJ+RF, no ener */
103 { "NxN RF Elec. + LJ [V&F]", 54 },
104 { "NxN QSTab Elec. + LJ [F]", 41 }, /* nbnxn kernel LJ+tab, no en */
105 { "NxN QSTab Elec. + LJ [V&F]", 59 },
106 { "NxN Ewald Elec. + LJ [F]", 66 }, /* nbnxn kernel LJ+Ewald, no en */
107 { "NxN Ewald Elec. + LJ [V&F]", 107 },
108 { "NxN LJ [F]", 33 }, /* nbnxn kernel LJ, no ener */
109 { "NxN LJ [V&F]", 43 },
110 { "NxN RF Electrostatics [F]", 31 }, /* nbnxn kernel RF, no ener */
111 { "NxN RF Electrostatics [V&F]", 36 },
112 { "NxN QSTab Elec. [F]", 34 }, /* nbnxn kernel tab, no ener */
113 { "NxN QSTab Elec. [V&F]", 41 },
114 { "NxN Ewald Elec. [F]", 61 }, /* nbnxn kernel Ewald, no ener */
115 { "NxN Ewald Elec. [V&F]", 84 },
116 /* The switch function flops should be added to the LJ kernels above */
117 { "NxN LJ add F-switch [F]", 12 }, /* extra cost for LJ F-switch */
118 { "NxN LJ add F-switch [V&F]", 22 },
119 { "NxN LJ add P-switch [F]", 27 }, /* extra cost for LJ P-switch */
120 { "NxN LJ add P-switch [V&F]", 20 },
121 { "1,4 nonbonded interactions", 90 },
122 { "Born radii (Still)", 47 },
123 { "Born radii (HCT/OBC)", 183 },
124 { "Born force chain rule", 15 },
125 { "All-vs-All Still radii", 1 },
126 { "All-vs-All HCT/OBC radii", 1 },
127 { "All-vs-All Born chain rule", 1 },
128 { "Calc Weights", 36 },
130 { "Spread Q Bspline", 2 },
132 { "Gather F Bspline", 6 },
134 { "Convolution", 4 },
137 { "Reset In Box", 3 },
143 { "FENE Bonds", 58 },
144 { "Tab. Bonds", 62 },
145 { "Restraint Potential", 86 },
146 { "Linear Angles", 57 },
148 { "G96Angles", 150 },
149 { "Quartic Angles", 160 },
150 { "Tab. Angles", 169 },
152 { "Impropers", 208 },
153 { "RB-Dihedrals", 247 },
154 { "Four. Dihedrals", 247 },
155 { "Tab. Dihedrals", 227 },
156 { "Dist. Restr.", 200 },
157 { "Orient. Restr.", 200 },
158 { "Dihedral Restr.", 200 },
159 { "Pos. Restr.", 50 },
160 { "Flat-bottom posres", 50 },
161 { "Angle Restr.", 191 },
162 { "Angle Restr. Z", 164 },
163 { "Morse Potent.", 83 },
164 { "Cubic Bonds", 54 },
166 { "Polarization", 59 },
167 { "Anharmonic Polarization", 72 },
168 { "Water Pol.", 62 },
169 { "Thole Pol.", 296 },
172 { "Ext.ens. Update", 54 },
179 { "Constraint-V", 8 },
180 { "Shake-Init", 10 },
181 { "Constraint-Vir", 24 },
183 { "Virtual Site 2", 23 },
184 { "Virtual Site 3", 37 },
185 { "Virtual Site 3fd", 95 },
186 { "Virtual Site 3fad", 176 },
187 { "Virtual Site 3out", 87 },
188 { "Virtual Site 4fd", 110 },
189 { "Virtual Site 4fdn", 254 },
190 { "Virtual Site N", 15 },
191 { "Mixed Generalized Born stuff", 10 }
194 static void pr_two(FILE *out, int c, int i)
198 fprintf(out, "%c0%1d", c, i);
202 fprintf(out, "%c%2d", c, i);
206 static void pr_difftime(FILE *out, double dt)
208 int ndays, nhours, nmins, nsecs;
209 gmx_bool bPrint, bPrinted;
211 ndays = dt/(24*3600);
212 dt = dt-24*3600*ndays;
218 bPrint = (ndays > 0);
222 fprintf(out, "%d", ndays);
224 bPrint = bPrint || (nhours > 0);
229 pr_two(out, 'd', nhours);
233 fprintf(out, "%d", nhours);
236 bPrinted = bPrinted || bPrint;
237 bPrint = bPrint || (nmins > 0);
242 pr_two(out, 'h', nmins);
246 fprintf(out, "%d", nmins);
249 bPrinted = bPrinted || bPrint;
252 pr_two(out, ':', nsecs);
256 fprintf(out, "%ds", nsecs);
261 void init_nrnb(t_nrnb *nrnb)
265 for (i = 0; (i < eNRNB); i++)
271 void cp_nrnb(t_nrnb *dest, t_nrnb *src)
275 for (i = 0; (i < eNRNB); i++)
277 dest->n[i] = src->n[i];
281 void add_nrnb(t_nrnb *dest, t_nrnb *s1, t_nrnb *s2)
285 for (i = 0; (i < eNRNB); i++)
287 dest->n[i] = s1->n[i]+s2->n[i];
291 void print_nrnb(FILE *out, t_nrnb *nrnb)
295 for (i = 0; (i < eNRNB); i++)
299 fprintf(out, " %-26s %10.0f.\n", nbdata[i].name, nrnb->n[i]);
304 void _inc_nrnb(t_nrnb *nrnb, int enr, int inc, char gmx_unused *file, int gmx_unused line)
308 printf("nrnb %15s(%2d) incremented with %8d from file %s line %d\n",
309 nbdata[enr].name, enr, inc, file, line);
313 /* Returns in enr is the index of a full nbnxn VdW kernel */
314 static gmx_bool nrnb_is_nbnxn_vdw_kernel(int enr)
316 return (enr >= eNR_NBNXN_LJ_RF && enr <= eNR_NBNXN_LJ_E);
319 /* Returns in enr is the index of an nbnxn kernel addition (switch function) */
320 static gmx_bool nrnb_is_nbnxn_kernel_addition(int enr)
322 return (enr >= eNR_NBNXN_LJ_FSW && enr <= eNR_NBNXN_LJ_PSW_E);
325 void print_flop(FILE *out, t_nrnb *nrnb, double *nbfs, double *mflop)
328 double mni, frac, tfrac, tflop;
329 const char *myline = "-----------------------------------------------------------------------------";
332 for (i = 0; (i < eNR_NBKERNEL_ALLVSALLGB); i++)
334 if (strstr(nbdata[i].name, "W3-W3") != NULL)
336 *nbfs += 9e-6*nrnb->n[i];
338 else if (strstr(nbdata[i].name, "W3") != NULL)
340 *nbfs += 3e-6*nrnb->n[i];
342 else if (strstr(nbdata[i].name, "W4-W4") != NULL)
344 *nbfs += 10e-6*nrnb->n[i];
346 else if (strstr(nbdata[i].name, "W4") != NULL)
348 *nbfs += 4e-6*nrnb->n[i];
352 *nbfs += 1e-6*nrnb->n[i];
356 for (i = 0; (i < eNRNB); i++)
358 tflop += 1e-6*nrnb->n[i]*nbdata[i].flop;
363 fprintf(out, "No MEGA Flopsen this time\n");
368 fprintf(out, "\n\tM E G A - F L O P S A C C O U N T I N G\n\n");
373 fprintf(out, " NB=Group-cutoff nonbonded kernels NxN=N-by-N cluster Verlet kernels\n");
374 fprintf(out, " RF=Reaction-Field VdW=Van der Waals QSTab=quadratic-spline table\n");
375 fprintf(out, " W3=SPC/TIP3p W4=TIP4p (single or pairs)\n");
376 fprintf(out, " V&F=Potential and force V=Potential only F=Force only\n\n");
378 fprintf(out, " %-32s %16s %15s %7s\n",
379 "Computing:", "M-Number", "M-Flops", "% Flops");
380 fprintf(out, "%s\n", myline);
384 for (i = 0; (i < eNRNB); i++)
386 mni = 1e-6*nrnb->n[i];
387 /* Skip empty entries and nbnxn additional flops,
388 * which have been added to the kernel entry.
390 if (mni > 0 && !nrnb_is_nbnxn_kernel_addition(i))
394 flop = nbdata[i].flop;
395 if (nrnb_is_nbnxn_vdw_kernel(i))
397 /* Possibly add the cost of a switch function */
398 for (j = eNR_NBNXN_LJ_FSW; j <= eNR_NBNXN_LJ_PSW; j += 2)
402 /* Select the force or energy flop count */
403 e_kernel_add = j + ((i - eNR_NBNXN_LJ_RF) % 2);
405 if (nrnb->n[e_kernel_add] > 0)
407 flop += nbdata[e_kernel_add].flop;
412 frac = 100.0*mni*flop/tflop;
416 fprintf(out, " %-32s %16.6f %15.3f %6.1f\n",
417 nbdata[i].name, mni, mni*flop, frac);
423 fprintf(out, "%s\n", myline);
424 fprintf(out, " %-32s %16s %15.3f %6.1f\n",
425 "Total", "", *mflop, tfrac);
426 fprintf(out, "%s\n\n", myline);
430 void print_perf(FILE *out, double time_per_thread, double time_per_node,
431 gmx_int64_t nsteps, real delta_t,
432 double nbfs, double mflop)
438 if (time_per_node > 0)
440 fprintf(out, "%12s %12s %12s %10s\n", "", "Core t (s)", "Wall t (s)", "(%)");
441 fprintf(out, "%12s %12.3f %12.3f %10.1f\n", "Time:",
442 time_per_thread, time_per_node, 100.0*time_per_thread/time_per_node);
443 /* only print day-hour-sec format if time_per_node is more than 30 min */
444 if (time_per_node > 30*60)
446 fprintf(out, "%12s %12s", "", "");
447 pr_difftime(out, time_per_node);
451 mflop = mflop/time_per_node;
452 wallclocktime = nsteps*delta_t;
454 if (getenv("GMX_DETAILED_PERF_STATS") == NULL)
456 fprintf(out, "%12s %12s %12s\n",
457 "", "(ns/day)", "(hour/ns)");
458 fprintf(out, "%12s %12.3f %12.3f\n", "Performance:",
459 wallclocktime*24*3.6/time_per_node, 1000*time_per_node/(3600*wallclocktime));
463 fprintf(out, "%12s %12s %12s %12s %12s\n",
464 "", "(Mnbf/s)", (mflop > 1000) ? "(GFlops)" : "(MFlops)",
465 "(ns/day)", "(hour/ns)");
466 fprintf(out, "%12s %12.3f %12.3f %12.3f %12.3f\n", "Performance:",
467 nbfs/time_per_node, (mflop > 1000) ? (mflop/1000) : mflop,
468 wallclocktime*24*3.6/time_per_node, 1000*time_per_node/(3600*wallclocktime));
473 if (getenv("GMX_DETAILED_PERF_STATS") == NULL)
475 fprintf(out, "%12s %14s\n",
477 fprintf(out, "%12s %14.1f\n", "Performance:",
478 nsteps*3600.0/time_per_node);
482 fprintf(out, "%12s %12s %12s %14s\n",
483 "", "(Mnbf/s)", (mflop > 1000) ? "(GFlops)" : "(MFlops)",
485 fprintf(out, "%12s %12.3f %12.3f %14.1f\n", "Performance:",
486 nbfs/time_per_node, (mflop > 1000) ? (mflop/1000) : mflop,
487 nsteps*3600.0/time_per_node);
493 int cost_nrnb(int enr)
495 return nbdata[enr].flop;
498 const char *nrnb_str(int enr)
500 return nbdata[enr].name;
503 static const int force_index[] = {
504 eNR_BONDS, eNR_ANGLES, eNR_PROPER, eNR_IMPROPER,
505 eNR_RB, eNR_DISRES, eNR_ORIRES, eNR_POSRES,
506 eNR_FBPOSRES, eNR_NS,
508 #define NFORCE_INDEX asize(force_index)
510 static const int constr_index[] = {
511 eNR_SHAKE, eNR_SHAKE_RIJ, eNR_SETTLE, eNR_UPDATE, eNR_PCOUPL,
512 eNR_CONSTR_VIR, eNR_CONSTR_V
514 #define NCONSTR_INDEX asize(constr_index)
516 static double pr_av(FILE *log, t_commrec *cr,
517 double fav, double ftot[], const char *title)
525 fav /= cr->nnodes - cr->npmenodes;
526 fprintf(log, "\n %-26s", title);
527 for (i = 0; (i < cr->nnodes); i++)
529 dperc = (100.0*ftot[i])/fav;
530 unb = max(unb, dperc);
532 fprintf(log, "%3d ", perc);
537 fprintf(log, "%6d%%\n\n", perc);
541 fprintf(log, "\n\n");
547 void pr_load(FILE *log, t_commrec *cr, t_nrnb nrnb[])
550 double dperc, unb, uf, us;
556 snew(ftot, cr->nnodes);
557 snew(stot, cr->nnodes);
559 for (i = 0; (i < cr->nnodes); i++)
561 add_nrnb(av, av, &(nrnb[i]));
562 /* Cost due to forces */
563 for (j = 0; (j < eNR_NBKERNEL_ALLVSALLGB); j++)
565 ftot[i] += nrnb[i].n[j]*cost_nrnb(j);
567 for (j = 0; (j < NFORCE_INDEX); j++)
569 ftot[i] += nrnb[i].n[force_index[j]]*cost_nrnb(force_index[j]);
572 for (j = 0; (j < NCONSTR_INDEX); j++)
574 stot[i] += nrnb[i].n[constr_index[j]]*cost_nrnb(constr_index[j]);
577 for (j = 0; (j < eNRNB); j++)
579 av->n[j] = av->n[j]/(double)(cr->nnodes - cr->npmenodes);
582 fprintf(log, "\nDetailed load balancing info in percentage of average\n");
584 fprintf(log, " Type NODE:");
585 for (i = 0; (i < cr->nnodes); i++)
587 fprintf(log, "%3d ", i);
589 fprintf(log, "Scaling\n");
590 fprintf(log, "---------------------------");
591 for (i = 0; (i < cr->nnodes); i++)
593 fprintf(log, "----");
595 fprintf(log, "-------\n");
597 for (j = 0; (j < eNRNB); j++)
602 fprintf(log, " %-26s", nrnb_str(j));
603 for (i = 0; (i < cr->nnodes); i++)
605 dperc = (100.0*nrnb[i].n[j])/av->n[j];
606 unb = max(unb, dperc);
608 fprintf(log, "%3d ", perc);
613 fprintf(log, "%6d%%\n", perc);
622 for (i = 0; (i < cr->nnodes); i++)
627 uf = pr_av(log, cr, fav, ftot, "Total Force");
628 us = pr_av(log, cr, sav, stot, "Total Constr.");
630 unb = (uf*fav+us*sav)/(fav+sav);
634 fprintf(log, "\nTotal Scaling: %.0f%% of max performance\n\n", unb);