3 * This source code is part of
7 * GROningen MAchine for Chemical Simulations
10 * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
11 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
12 * Copyright (c) 2001-2004, The GROMACS development team,
13 * check out http://www.gromacs.org for more information.
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version 2
18 * of the License, or (at your option) any later version.
20 * If you want to redistribute modifications, please consider that
21 * scientific software is very special. Version control is crucial -
22 * bugs must be traceable. We will be happy to consider code for
23 * inclusion in the official distribution, but derived work must not
24 * be called official GROMACS. Details are found in the README & COPYING
25 * files - if they are missing, get the official version at www.gromacs.org.
27 * To help us fund GROMACS development, we humbly ask that you cite
28 * the papers on the package - you can find them in the top README file.
30 * For more info, check our website at http://www.gromacs.org
33 * GROningen Mixture of Alchemy and Childrens' Stories
40 #include "types/commrec.h"
42 #include "gmx_fatal.h"
55 static const t_nrnb_data nbdata[eNRNB] = {
56 /* These are re-used for different NB kernels, since there are so many.
57 * The actual number of flops is set dynamically.
59 { "NB VdW [V&F]", 1 },
61 { "NB Elec. [V&F]", 1 },
62 { "NB Elec. [F]", 1 },
63 { "NB Elec. [W3,V&F]", 1 },
64 { "NB Elec. [W3,F]", 1 },
65 { "NB Elec. [W3-W3,V&F]", 1 },
66 { "NB Elec. [W3-W3,F]", 1 },
67 { "NB Elec. [W4,V&F]", 1 },
68 { "NB Elec. [W4,F]", 1 },
69 { "NB Elec. [W4-W4,V&F]", 1 },
70 { "NB Elec. [W4-W4,F]", 1 },
71 { "NB VdW & Elec. [V&F]", 1 },
72 { "NB VdW & Elec. [F]", 1 },
73 { "NB VdW & Elec. [W3,V&F]", 1 },
74 { "NB VdW & Elec. [W3,F]", 1 },
75 { "NB VdW & Elec. [W3-W3,V&F]", 1 },
76 { "NB VdW & Elec. [W3-W3,F]", 1 },
77 { "NB VdW & Elec. [W4,V&F]", 1 },
78 { "NB VdW & Elec. [W4,F]", 1 },
79 { "NB VdW & Elec. [W4-W4,V&F]", 1 },
80 { "NB VdW & Elec. [W4-W4,F]", 1 },
82 { "NB Generic kernel", 1 },
83 { "NB Generic charge grp kernel", 1 },
84 { "NB Generic AdResS kernel", 1 },
85 { "NB Free energy kernel", 1 },
86 { "NB All-vs-all", 1 },
87 { "NB All-vs-all, GB", 1 },
89 { "Pair Search distance check", 9 }, /* nbnxn pair dist. check */
90 /* nbnxn kernel flops are based on inner-loops without exclusion checks.
91 * Plain Coulomb runs through the RF kernels, except with CUDA.
92 * invsqrt is counted as 6 flops: 1 for _mm_rsqt_ps + 5 for iteration.
93 * The flops are equal for plain-C, x86 SIMD and CUDA, except for:
94 * - plain-C kernel uses one flop more for Coulomb-only (F) than listed
95 * - x86 SIMD LJ geom-comb.rule kernels (fastest) use 2 more flops
96 * - x86 SIMD LJ LB-comb.rule kernels (fast) use 3 (8 for F+E) more flops
97 * - GPU always does exclusions, which requires 2-4 flops, but as invsqrt
98 * is always counted as 6 flops, this roughly compensates.
100 { "NxN RF Elec. + VdW [F]", 38 }, /* nbnxn kernel LJ+RF, no ener */
101 { "NxN RF Elec. + VdW [V&F]", 54 },
102 { "NxN QSTab Elec. + VdW [F]", 41 }, /* nbnxn kernel LJ+tab, no en */
103 { "NxN QSTab Elec. + VdW [V&F]", 59 },
104 { "NxN Ewald Elec. + VdW [F]", 66 }, /* nbnxn kernel LJ+Ewald, no en */
105 { "NxN Ewald Elec. + VdW [V&F]", 107 },
106 { "NxN VdW [F]", 33 }, /* nbnxn kernel LJ, no ener */
107 { "NxN VdW [V&F]", 43 },
108 { "NxN RF Electrostatics [F]", 31 }, /* nbnxn kernel RF, no ener */
109 { "NxN RF Electrostatics [V&F]", 36 },
110 { "NxN QSTab Elec. [F]", 34 }, /* nbnxn kernel tab, no ener */
111 { "NxN QSTab Elec. [V&F]", 41 },
112 { "NxN Ewald Elec. [F]", 61 }, /* nbnxn kernel Ewald, no ener */
113 { "NxN Ewald Elec. [V&F]", 84 },
114 { "1,4 nonbonded interactions", 90 },
115 { "Born radii (Still)", 47 },
116 { "Born radii (HCT/OBC)", 183 },
117 { "Born force chain rule", 15 },
118 { "All-vs-All Still radii", 1 },
119 { "All-vs-All HCT/OBC radii", 1 },
120 { "All-vs-All Born chain rule", 1 },
121 { "Calc Weights", 36 },
123 { "Spread Q Bspline", 2 },
125 { "Gather F Bspline", 6 },
127 { "Convolution", 4 },
130 { "Reset In Box", 3 },
136 { "FENE Bonds", 58 },
137 { "Tab. Bonds", 62 },
138 { "Restraint Potential", 86 },
139 { "Linear Angles", 57 },
141 { "G96Angles", 150 },
142 { "Quartic Angles", 160 },
143 { "Tab. Angles", 169 },
145 { "Impropers", 208 },
146 { "RB-Dihedrals", 247 },
147 { "Four. Dihedrals", 247 },
148 { "Tab. Dihedrals", 227 },
149 { "Dist. Restr.", 200 },
150 { "Orient. Restr.", 200 },
151 { "Dihedral Restr.", 200 },
152 { "Pos. Restr.", 50 },
153 { "Flat-bottom posres", 50 },
154 { "Angle Restr.", 191 },
155 { "Angle Restr. Z", 164 },
156 { "Morse Potent.", 83 },
157 { "Cubic Bonds", 54 },
159 { "Polarization", 59 },
160 { "Anharmonic Polarization", 72 },
161 { "Water Pol.", 62 },
162 { "Thole Pol.", 296 },
165 { "Ext.ens. Update", 54 },
172 { "Constraint-V", 8 },
173 { "Shake-Init", 10 },
174 { "Constraint-Vir", 24 },
176 { "Virtual Site 2", 23 },
177 { "Virtual Site 3", 37 },
178 { "Virtual Site 3fd", 95 },
179 { "Virtual Site 3fad", 176 },
180 { "Virtual Site 3out", 87 },
181 { "Virtual Site 4fd", 110 },
182 { "Virtual Site 4fdn", 254 },
183 { "Virtual Site N", 15 },
184 { "Mixed Generalized Born stuff", 10 }
187 static void pr_two(FILE *out, int c, int i)
191 fprintf(out, "%c0%1d", c, i);
195 fprintf(out, "%c%2d", c, i);
199 static void pr_difftime(FILE *out, double dt)
201 int ndays, nhours, nmins, nsecs;
202 gmx_bool bPrint, bPrinted;
204 ndays = dt/(24*3600);
205 dt = dt-24*3600*ndays;
211 bPrint = (ndays > 0);
215 fprintf(out, "%d", ndays);
217 bPrint = bPrint || (nhours > 0);
222 pr_two(out, 'd', nhours);
226 fprintf(out, "%d", nhours);
229 bPrinted = bPrinted || bPrint;
230 bPrint = bPrint || (nmins > 0);
235 pr_two(out, 'h', nmins);
239 fprintf(out, "%d", nmins);
242 bPrinted = bPrinted || bPrint;
245 pr_two(out, ':', nsecs);
249 fprintf(out, "%ds", nsecs);
254 void init_nrnb(t_nrnb *nrnb)
258 for (i = 0; (i < eNRNB); i++)
264 void cp_nrnb(t_nrnb *dest, t_nrnb *src)
268 for (i = 0; (i < eNRNB); i++)
270 dest->n[i] = src->n[i];
274 void add_nrnb(t_nrnb *dest, t_nrnb *s1, t_nrnb *s2)
278 for (i = 0; (i < eNRNB); i++)
280 dest->n[i] = s1->n[i]+s2->n[i];
284 void print_nrnb(FILE *out, t_nrnb *nrnb)
288 for (i = 0; (i < eNRNB); i++)
292 fprintf(out, " %-26s %10.0f.\n", nbdata[i].name, nrnb->n[i]);
297 void _inc_nrnb(t_nrnb *nrnb, int enr, int inc, char gmx_unused *file, int gmx_unused line)
301 printf("nrnb %15s(%2d) incremented with %8d from file %s line %d\n",
302 nbdata[enr].name, enr, inc, file, line);
306 void print_flop(FILE *out, t_nrnb *nrnb, double *nbfs, double *mflop)
309 double mni, frac, tfrac, tflop;
310 const char *myline = "-----------------------------------------------------------------------------";
313 for (i = 0; (i < eNR_NBKERNEL_ALLVSALLGB); i++)
315 if (strstr(nbdata[i].name, "W3-W3") != NULL)
317 *nbfs += 9e-6*nrnb->n[i];
319 else if (strstr(nbdata[i].name, "W3") != NULL)
321 *nbfs += 3e-6*nrnb->n[i];
323 else if (strstr(nbdata[i].name, "W4-W4") != NULL)
325 *nbfs += 10e-6*nrnb->n[i];
327 else if (strstr(nbdata[i].name, "W4") != NULL)
329 *nbfs += 4e-6*nrnb->n[i];
333 *nbfs += 1e-6*nrnb->n[i];
337 for (i = 0; (i < eNRNB); i++)
339 tflop += 1e-6*nrnb->n[i]*nbdata[i].flop;
344 fprintf(out, "No MEGA Flopsen this time\n");
349 fprintf(out, "\n\tM E G A - F L O P S A C C O U N T I N G\n\n");
354 fprintf(out, " NB=Group-cutoff nonbonded kernels NxN=N-by-N cluster Verlet kernels\n");
355 fprintf(out, " RF=Reaction-Field VdW=Van der Waals QSTab=quadratic-spline table\n");
356 fprintf(out, " W3=SPC/TIP3p W4=TIP4p (single or pairs)\n");
357 fprintf(out, " V&F=Potential and force V=Potential only F=Force only\n\n");
359 fprintf(out, " %-32s %16s %15s %7s\n",
360 "Computing:", "M-Number", "M-Flops", "% Flops");
361 fprintf(out, "%s\n", myline);
365 for (i = 0; (i < eNRNB); i++)
367 mni = 1e-6*nrnb->n[i];
368 *mflop += mni*nbdata[i].flop;
369 frac = 100.0*mni*nbdata[i].flop/tflop;
373 fprintf(out, " %-32s %16.6f %15.3f %6.1f\n",
374 nbdata[i].name, mni, mni*nbdata[i].flop, frac);
379 fprintf(out, "%s\n", myline);
380 fprintf(out, " %-32s %16s %15.3f %6.1f\n",
381 "Total", "", *mflop, tfrac);
382 fprintf(out, "%s\n\n", myline);
386 void print_perf(FILE *out, double time_per_thread, double time_per_node,
387 gmx_large_int_t nsteps, real delta_t,
388 double nbfs, double mflop)
394 if (time_per_node > 0)
396 fprintf(out, "%12s %12s %12s %10s\n", "", "Core t (s)", "Wall t (s)", "(%)");
397 fprintf(out, "%12s %12.3f %12.3f %10.1f\n", "Time:",
398 time_per_thread, time_per_node, 100.0*time_per_thread/time_per_node);
399 /* only print day-hour-sec format if time_per_node is more than 30 min */
400 if (time_per_node > 30*60)
402 fprintf(out, "%12s %12s", "", "");
403 pr_difftime(out, time_per_node);
407 mflop = mflop/time_per_node;
408 wallclocktime = nsteps*delta_t;
410 if (getenv("GMX_DETAILED_PERF_STATS") == NULL)
412 fprintf(out, "%12s %12s %12s\n",
413 "", "(ns/day)", "(hour/ns)");
414 fprintf(out, "%12s %12.3f %12.3f\n", "Performance:",
415 wallclocktime*24*3.6/time_per_node, 1000*time_per_node/(3600*wallclocktime));
419 fprintf(out, "%12s %12s %12s %12s %12s\n",
420 "", "(Mnbf/s)", (mflop > 1000) ? "(GFlops)" : "(MFlops)",
421 "(ns/day)", "(hour/ns)");
422 fprintf(out, "%12s %12.3f %12.3f %12.3f %12.3f\n", "Performance:",
423 nbfs/time_per_node, (mflop > 1000) ? (mflop/1000) : mflop,
424 wallclocktime*24*3.6/time_per_node, 1000*time_per_node/(3600*wallclocktime));
429 if (getenv("GMX_DETAILED_PERF_STATS") == NULL)
431 fprintf(out, "%12s %14s\n",
433 fprintf(out, "%12s %14.1f\n", "Performance:",
434 nsteps*3600.0/time_per_node);
438 fprintf(out, "%12s %12s %12s %14s\n",
439 "", "(Mnbf/s)", (mflop > 1000) ? "(GFlops)" : "(MFlops)",
441 fprintf(out, "%12s %12.3f %12.3f %14.1f\n", "Performance:",
442 nbfs/time_per_node, (mflop > 1000) ? (mflop/1000) : mflop,
443 nsteps*3600.0/time_per_node);
449 int cost_nrnb(int enr)
451 return nbdata[enr].flop;
454 const char *nrnb_str(int enr)
456 return nbdata[enr].name;
459 static const int force_index[] = {
460 eNR_BONDS, eNR_ANGLES, eNR_PROPER, eNR_IMPROPER,
461 eNR_RB, eNR_DISRES, eNR_ORIRES, eNR_POSRES,
462 eNR_FBPOSRES, eNR_NS,
464 #define NFORCE_INDEX asize(force_index)
466 static const int constr_index[] = {
467 eNR_SHAKE, eNR_SHAKE_RIJ, eNR_SETTLE, eNR_UPDATE, eNR_PCOUPL,
468 eNR_CONSTR_VIR, eNR_CONSTR_V
470 #define NCONSTR_INDEX asize(constr_index)
472 static double pr_av(FILE *log, t_commrec *cr,
473 double fav, double ftot[], const char *title)
481 fav /= cr->nnodes - cr->npmenodes;
482 fprintf(log, "\n %-26s", title);
483 for (i = 0; (i < cr->nnodes); i++)
485 dperc = (100.0*ftot[i])/fav;
486 unb = max(unb, dperc);
488 fprintf(log, "%3d ", perc);
493 fprintf(log, "%6d%%\n\n", perc);
497 fprintf(log, "\n\n");
503 void pr_load(FILE *log, t_commrec *cr, t_nrnb nrnb[])
506 double dperc, unb, uf, us;
512 snew(ftot, cr->nnodes);
513 snew(stot, cr->nnodes);
515 for (i = 0; (i < cr->nnodes); i++)
517 add_nrnb(av, av, &(nrnb[i]));
518 /* Cost due to forces */
519 for (j = 0; (j < eNR_NBKERNEL_ALLVSALLGB); j++)
521 ftot[i] += nrnb[i].n[j]*cost_nrnb(j);
523 for (j = 0; (j < NFORCE_INDEX); j++)
525 ftot[i] += nrnb[i].n[force_index[j]]*cost_nrnb(force_index[j]);
528 for (j = 0; (j < NCONSTR_INDEX); j++)
530 stot[i] += nrnb[i].n[constr_index[j]]*cost_nrnb(constr_index[j]);
533 for (j = 0; (j < eNRNB); j++)
535 av->n[j] = av->n[j]/(double)(cr->nnodes - cr->npmenodes);
538 fprintf(log, "\nDetailed load balancing info in percentage of average\n");
540 fprintf(log, " Type NODE:");
541 for (i = 0; (i < cr->nnodes); i++)
543 fprintf(log, "%3d ", i);
545 fprintf(log, "Scaling\n");
546 fprintf(log, "---------------------------");
547 for (i = 0; (i < cr->nnodes); i++)
549 fprintf(log, "----");
551 fprintf(log, "-------\n");
553 for (j = 0; (j < eNRNB); j++)
558 fprintf(log, " %-26s", nrnb_str(j));
559 for (i = 0; (i < cr->nnodes); i++)
561 dperc = (100.0*nrnb[i].n[j])/av->n[j];
562 unb = max(unb, dperc);
564 fprintf(log, "%3d ", perc);
569 fprintf(log, "%6d%%\n", perc);
578 for (i = 0; (i < cr->nnodes); i++)
583 uf = pr_av(log, cr, fav, ftot, "Total Force");
584 us = pr_av(log, cr, sav, stot, "Total Constr.");
586 unb = (uf*fav+us*sav)/(fav+sav);
590 fprintf(log, "\nTotal Scaling: %.0f%% of max performance\n\n", unb);