2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2008, 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.
39 #include "wallcycle.h"
45 #include "gromacs/legacyheaders/md_logging.h"
46 #include "gromacs/legacyheaders/types/commrec.h"
47 #include "gromacs/timing/cyclecounter.h"
48 #include "gromacs/timing/gpu_timing.h"
49 #include "gromacs/utility/cstringutil.h"
50 #include "gromacs/utility/gmxmpi.h"
51 #include "gromacs/utility/smalloc.h"
52 #include "gromacs/utility/snprintf.h"
54 /* DEBUG_WCYCLE adds consistency checking for the counters.
55 * It checks if you stop a counter different from the last
56 * one that was opened and if you do nest too deep.
58 /* #define DEBUG_WCYCLE */
61 #include "gromacs/utility/fatalerror.h"
72 typedef struct gmx_wallcycle
75 /* variables for testing/debugging */
81 int counterlist[DEPTH_MAX];
85 gmx_cycles_t cycle_prev;
86 gmx_int64_t reset_counters;
88 MPI_Comm mpi_comm_mygroup;
92 #ifdef GMX_CYCLE_SUBCOUNTERS
98 /* Each name should not exceed 19 printing characters
99 (ie. terminating null can be twentieth) */
100 static const char *wcn[ewcNR] =
102 "Run", "Step", "PP during PME", "Domain decomp.", "DD comm. load",
103 "DD comm. bounds", "Vsite constr.", "Send X to PME", "Neighbor search", "Launch GPU ops.",
104 "Comm. coord.", "Born radii", "Force", "Wait + Comm. F", "PME mesh",
105 "PME redist. X/F", "PME spread/gather", "PME 3D-FFT", "PME 3D-FFT Comm.", "PME solve LJ", "PME solve Elec",
106 "PME wait for PP", "Wait + Recv. PME F", "Wait GPU nonlocal", "Wait GPU local", "Wait GPU loc. est.", "NB X/F buffer ops.",
107 "Vsite spread", "COM pull force",
108 "Write traj.", "Update", "Constraints", "Comm. energies",
109 "Enforced rotation", "Add rot. forces", "Coordinate swapping", "IMD", "Test"
112 static const char *wcsn[ewcsNR] =
114 "DD redist.", "DD NS grid + sort", "DD setup comm.",
115 "DD make top.", "DD make constr.", "DD top. other",
116 "NS grid local", "NS grid non-loc.", "NS search local", "NS search non-loc.",
120 "Listed buffer ops.",
122 "Ewald F correction",
127 gmx_bool wallcycle_have_counter(void)
129 return gmx_cycles_have_counter();
132 gmx_wallcycle_t wallcycle_init(FILE *fplog, int resetstep, t_commrec gmx_unused *cr,
133 int nthreads_pp, int nthreads_pme)
138 if (!wallcycle_have_counter())
145 wc->wc_barrier = FALSE;
149 wc->reset_counters = resetstep;
150 wc->nthreads_pp = nthreads_pp;
151 wc->nthreads_pme = nthreads_pme;
152 wc->cycles_sum = NULL;
155 if (PAR(cr) && getenv("GMX_CYCLE_BARRIER") != NULL)
159 fprintf(fplog, "\nWill call MPI_Barrier before each cycle start/stop call\n\n");
161 wc->wc_barrier = TRUE;
162 wc->mpi_comm_mygroup = cr->mpi_comm_mygroup;
166 snew(wc->wcc, ewcNR);
167 if (getenv("GMX_CYCLE_ALL") != NULL)
171 fprintf(fplog, "\nWill time all the code during the run\n\n");
173 snew(wc->wcc_all, ewcNR*ewcNR);
176 #ifdef GMX_CYCLE_SUBCOUNTERS
177 snew(wc->wcsc, ewcsNR);
187 void wallcycle_destroy(gmx_wallcycle_t wc)
198 if (wc->wcc_all != NULL)
202 #ifdef GMX_CYCLE_SUBCOUNTERS
203 if (wc->wcsc != NULL)
211 static void wallcycle_all_start(gmx_wallcycle_t wc, int ewc, gmx_cycles_t cycle)
214 wc->cycle_prev = cycle;
217 static void wallcycle_all_stop(gmx_wallcycle_t wc, int ewc, gmx_cycles_t cycle)
219 wc->wcc_all[wc->ewc_prev*ewcNR+ewc].n += 1;
220 wc->wcc_all[wc->ewc_prev*ewcNR+ewc].c += cycle - wc->cycle_prev;
225 static void debug_start_check(gmx_wallcycle_t wc, int ewc)
227 /* fprintf(stderr,"wcycle_start depth %d, %s\n",wc->count_depth,wcn[ewc]); */
229 if (wc->count_depth < 0 || wc->count_depth >= DEPTH_MAX)
231 gmx_fatal(FARGS, "wallcycle counter depth out of range: %d",
234 wc->counterlist[wc->count_depth] = ewc;
238 static void debug_stop_check(gmx_wallcycle_t wc, int ewc)
242 /* fprintf(stderr,"wcycle_stop depth %d, %s\n",wc->count_depth,wcn[ewc]); */
244 if (wc->count_depth < 0)
246 gmx_fatal(FARGS, "wallcycle counter depth out of range when stopping %s: %d", wcn[ewc], wc->count_depth);
248 if (wc->counterlist[wc->count_depth] != ewc)
250 gmx_fatal(FARGS, "wallcycle mismatch at stop, start %s, stop %s",
251 wcn[wc->counterlist[wc->count_depth]], wcn[ewc]);
256 void wallcycle_start(gmx_wallcycle_t wc, int ewc)
268 MPI_Barrier(wc->mpi_comm_mygroup);
273 debug_start_check(wc, ewc);
276 cycle = gmx_cycles_read();
277 wc->wcc[ewc].start = cycle;
278 if (wc->wcc_all != NULL)
283 wallcycle_all_start(wc, ewc, cycle);
285 else if (wc->wc_depth == 3)
287 wallcycle_all_stop(wc, ewc, cycle);
292 void wallcycle_start_nocount(gmx_wallcycle_t wc, int ewc)
299 wallcycle_start(wc, ewc);
303 double wallcycle_stop(gmx_wallcycle_t wc, int ewc)
305 gmx_cycles_t cycle, last;
315 MPI_Barrier(wc->mpi_comm_mygroup);
320 debug_stop_check(wc, ewc);
323 cycle = gmx_cycles_read();
324 last = cycle - wc->wcc[ewc].start;
325 wc->wcc[ewc].c += last;
332 wallcycle_all_stop(wc, ewc, cycle);
334 else if (wc->wc_depth == 2)
336 wallcycle_all_start(wc, ewc, cycle);
343 void wallcycle_get(gmx_wallcycle_t wc, int ewc, int *n, double *c)
346 *c = (double)wc->wcc[ewc].c;
349 void wallcycle_reset_all(gmx_wallcycle_t wc)
358 for (i = 0; i < ewcNR; i++)
365 for (i = 0; i < ewcNR*ewcNR; i++)
367 wc->wcc_all[i].n = 0;
368 wc->wcc_all[i].c = 0;
371 #ifdef GMX_CYCLE_SUBCOUNTERS
372 for (i = 0; i < ewcsNR; i++)
380 static gmx_bool is_pme_counter(int ewc)
382 return (ewc >= ewcPMEMESH && ewc <= ewcPMEWAITCOMM);
385 static gmx_bool is_pme_subcounter(int ewc)
387 return (ewc >= ewcPME_REDISTXF && ewc < ewcPMEWAITCOMM);
390 void wallcycle_sum(t_commrec *cr, gmx_wallcycle_t wc)
393 double cycles[ewcNR+ewcsNR];
394 double cycles_n[ewcNR+ewcsNR], buf[ewcNR+ewcsNR], *cyc_all, *buf_all;
403 snew(wc->cycles_sum, ewcNR+ewcsNR);
407 /* The GPU wait estimate counter is used for load balancing only
408 * and will mess up the total due to double counting: clear it.
410 wcc[ewcWAIT_GPU_NB_L_EST].n = 0;
411 wcc[ewcWAIT_GPU_NB_L_EST].c = 0;
413 for (i = 0; i < ewcNR; i++)
415 if (is_pme_counter(i) || (i == ewcRUN && cr->duty == DUTY_PME))
417 wcc[i].c *= wc->nthreads_pme;
421 for (j = 0; j < ewcNR; j++)
423 wc->wcc_all[i*ewcNR+j].c *= wc->nthreads_pme;
429 wcc[i].c *= wc->nthreads_pp;
433 for (j = 0; j < ewcNR; j++)
435 wc->wcc_all[i*ewcNR+j].c *= wc->nthreads_pp;
441 if (wcc[ewcDDCOMMLOAD].n > 0)
443 wcc[ewcDOMDEC].c -= wcc[ewcDDCOMMLOAD].c;
445 if (wcc[ewcDDCOMMBOUND].n > 0)
447 wcc[ewcDOMDEC].c -= wcc[ewcDDCOMMBOUND].c;
449 if (wcc[ewcPME_FFTCOMM].n > 0)
451 wcc[ewcPME_FFT].c -= wcc[ewcPME_FFTCOMM].c;
454 if (cr->npmenodes == 0)
456 /* All nodes do PME (or no PME at all) */
457 if (wcc[ewcPMEMESH].n > 0)
459 wcc[ewcFORCE].c -= wcc[ewcPMEMESH].c;
464 /* The are PME-only nodes */
465 if (wcc[ewcPMEMESH].n > 0)
467 /* This must be a PME only node, calculate the Wait + Comm. time */
468 wcc[ewcPMEWAITCOMM].c = wcc[ewcRUN].c - wcc[ewcPMEMESH].c;
472 /* Store the cycles in a double buffer for summing */
473 for (i = 0; i < ewcNR; i++)
475 cycles_n[i] = (double)wcc[i].n;
476 cycles[i] = (double)wcc[i].c;
479 #ifdef GMX_CYCLE_SUBCOUNTERS
480 for (i = 0; i < ewcsNR; i++)
482 wc->wcsc[i].c *= wc->nthreads_pp;
483 cycles_n[ewcNR+i] = (double)wc->wcsc[i].n;
484 cycles[ewcNR+i] = (double)wc->wcsc[i].c;
492 MPI_Allreduce(cycles_n, buf, nsum, MPI_DOUBLE, MPI_MAX,
494 for (i = 0; i < ewcNR; i++)
496 wcc[i].n = (int)(buf[i] + 0.5);
498 #ifdef GMX_CYCLE_SUBCOUNTERS
499 for (i = 0; i < ewcsNR; i++)
501 wc->wcsc[i].n = (int)(buf[ewcNR+i] + 0.5);
505 MPI_Allreduce(cycles, wc->cycles_sum, nsum, MPI_DOUBLE, MPI_SUM,
508 if (wc->wcc_all != NULL)
510 snew(cyc_all, ewcNR*ewcNR);
511 snew(buf_all, ewcNR*ewcNR);
512 for (i = 0; i < ewcNR*ewcNR; i++)
514 cyc_all[i] = wc->wcc_all[i].c;
516 MPI_Allreduce(cyc_all, buf_all, ewcNR*ewcNR, MPI_DOUBLE, MPI_SUM,
518 for (i = 0; i < ewcNR*ewcNR; i++)
520 wc->wcc_all[i].c = buf_all[i];
529 for (i = 0; i < nsum; i++)
531 wc->cycles_sum[i] = cycles[i];
536 static void print_cycles(FILE *fplog, double c2t, const char *name,
537 int nnodes, int nthreads,
538 int ncalls, double c_sum, double tot)
541 char nthreads_str[6];
544 double percentage = (tot > 0.) ? (100. * c_sum / tot) : 0.;
550 snprintf(ncalls_str, sizeof(ncalls_str), "%10d", ncalls);
553 snprintf(nnodes_str, sizeof(nnodes_str), "N/A");
557 snprintf(nnodes_str, sizeof(nnodes_str), "%4d", nnodes);
561 snprintf(nthreads_str, sizeof(nthreads_str), "N/A");
565 snprintf(nthreads_str, sizeof(nthreads_str), "%4d", nthreads);
574 /* Convert the cycle count to wallclock time for this task */
577 fprintf(fplog, " %-19.19s %4s %4s %10s %10.3f %14.3f %5.1f\n",
578 name, nnodes_str, nthreads_str, ncalls_str, wallt,
579 c_sum*1e-9, percentage);
583 static void print_gputimes(FILE *fplog, const char *name,
584 int n, double t, double tot_t)
591 snprintf(num, sizeof(num), "%10d", n);
592 snprintf(avg_perf, sizeof(avg_perf), "%10.3f", t/n);
597 sprintf(avg_perf, " ");
601 fprintf(fplog, " %-29s %10s%12.3f %s %5.1f\n",
602 name, num, t/1000, avg_perf, 100 * t/tot_t);
606 fprintf(fplog, " %-29s %10s%12.3f %s %5.1f\n",
607 name, "", t/1000, avg_perf, 100.0);
611 static void print_header(FILE *fplog, int nrank_pp, int nth_pp, int nrank_pme, int nth_pme)
613 int nrank_tot = nrank_pp + nrank_pme;
616 fprintf(fplog, "On %d MPI rank%s", nrank_tot, nrank_tot == 1 ? "" : "s");
619 fprintf(fplog, ", each using %d OpenMP threads", nth_pp);
621 /* Don't report doing PP+PME, because we can't tell here if
622 * this is RF, etc. */
626 fprintf(fplog, "On %d MPI rank%s doing PP", nrank_pp, nrank_pp == 1 ? "" : "s");
629 fprintf(fplog, ",%s using %d OpenMP threads", nrank_pp > 1 ? " each" : "", nth_pp);
631 fprintf(fplog, ", and\non %d MPI rank%s doing PME", nrank_pme, nrank_pme == 1 ? "" : "s");
634 fprintf(fplog, ",%s using %d OpenMP threads", nrank_pme > 1 ? " each" : "", nth_pme);
638 fprintf(fplog, "\n\n");
639 fprintf(fplog, " Computing: Num Num Call Wall time Giga-Cycles\n");
640 fprintf(fplog, " Ranks Threads Count (s) total sum %%\n");
643 void wallcycle_print(FILE *fplog, int nnodes, int npme, double realtime,
644 gmx_wallcycle_t wc, struct gmx_wallclock_gpu_t *gpu_t)
647 double tot, tot_for_pp, tot_for_rest, tot_gpu, tot_cpu_overlap, gpu_cpu_ratio, tot_k;
648 double c2t, c2t_pp, c2t_pme = 0;
649 int i, j, npp, nth_pp, nth_pme, nth_tot;
651 const char *hline = "-----------------------------------------------------------------------------";
658 nth_pp = wc->nthreads_pp;
659 nth_pme = wc->nthreads_pme;
661 cyc_sum = wc->cycles_sum;
664 nth_tot = npp*nth_pp + npme*nth_pme;
666 /* When using PME-only nodes, the next line is valid for both
667 PP-only and PME-only nodes because they started ewcRUN at the
669 tot = cyc_sum[ewcRUN];
672 /* Conversion factor from cycles to seconds */
676 c2t_pp = c2t * nth_tot / (double) (npp*nth_pp);
679 c2t_pme = c2t * nth_tot / (double) (npme*nth_pme);
689 fprintf(fplog, "\n R E A L C Y C L E A N D T I M E A C C O U N T I N G\n\n");
691 print_header(fplog, npp, nth_pp, npme, nth_pme);
693 fprintf(fplog, "%s\n", hline);
694 for (i = ewcPPDURINGPME+1; i < ewcNR; i++)
696 if (is_pme_subcounter(i))
698 /* Do not count these at all */
700 else if (npme > 0 && is_pme_counter(i))
702 /* Print timing information for PME-only nodes, but add an
703 * asterisk so the reader of the table can know that the
704 * walltimes are not meant to add up. The asterisk still
705 * fits in the required maximum of 19 characters. */
707 snprintf(buffer, STRLEN, "%s *", wcn[i]);
708 print_cycles(fplog, c2t_pme, buffer,
710 wc->wcc[i].n, cyc_sum[i], tot);
714 /* Print timing information when it is for a PP or PP+PME
716 print_cycles(fplog, c2t_pp, wcn[i],
718 wc->wcc[i].n, cyc_sum[i], tot);
719 tot_for_pp += cyc_sum[i];
722 if (wc->wcc_all != NULL)
724 for (i = 0; i < ewcNR; i++)
726 for (j = 0; j < ewcNR; j++)
728 snprintf(buf, 20, "%-9.9s %-9.9s", wcn[i], wcn[j]);
729 print_cycles(fplog, c2t_pp, buf,
731 wc->wcc_all[i*ewcNR+j].n,
732 wc->wcc_all[i*ewcNR+j].c,
737 tot_for_rest = tot * (npp * nth_pp) / (double) nth_tot;
738 print_cycles(fplog, c2t_pp, "Rest",
740 -1, tot_for_rest - tot_for_pp, tot);
741 fprintf(fplog, "%s\n", hline);
742 print_cycles(fplog, c2t, "Total",
745 fprintf(fplog, "%s\n", hline);
750 "(*) Note that with separate PME ranks, the walltime column actually sums to\n"
751 " twice the total reported, but the cycle count total and %% are correct.\n"
755 if (wc->wcc[ewcPMEMESH].n > 0)
757 fprintf(fplog, " Breakdown of PME mesh computation\n");
758 fprintf(fplog, "%s\n", hline);
759 for (i = ewcPPDURINGPME+1; i < ewcNR; i++)
761 if (is_pme_subcounter(i))
763 print_cycles(fplog, npme > 0 ? c2t_pme : c2t_pp, wcn[i],
764 npme > 0 ? npme : npp, nth_pme,
765 wc->wcc[i].n, cyc_sum[i], tot);
768 fprintf(fplog, "%s\n", hline);
771 #ifdef GMX_CYCLE_SUBCOUNTERS
772 fprintf(fplog, " Breakdown of PP computation\n");
773 fprintf(fplog, "%s\n", hline);
774 for (i = 0; i < ewcsNR; i++)
776 print_cycles(fplog, c2t_pp, wcsn[i],
778 wc->wcsc[i].n, cyc_sum[ewcNR+i], tot);
780 fprintf(fplog, "%s\n", hline);
783 /* print GPU timing summary */
786 const char *k_log_str[2][2] = {
787 {"Nonbonded F kernel", "Nonbonded F+ene k."},
788 {"Nonbonded F+prune k.", "Nonbonded F+ene+prune k."}
791 tot_gpu = gpu_t->pl_h2d_t + gpu_t->nb_h2d_t + gpu_t->nb_d2h_t;
793 /* add up the kernel timings */
795 for (i = 0; i < 2; i++)
797 for (j = 0; j < 2; j++)
799 tot_k += gpu_t->ktime[i][j].t;
804 tot_cpu_overlap = wc->wcc[ewcFORCE].c;
805 if (wc->wcc[ewcPMEMESH].n > 0)
807 tot_cpu_overlap += wc->wcc[ewcPMEMESH].c;
809 tot_cpu_overlap *= realtime*1000/tot; /* convert s to ms */
811 fprintf(fplog, "\n GPU timings\n%s\n", hline);
812 fprintf(fplog, " Computing: Count Wall t (s) ms/step %c\n", '%');
813 fprintf(fplog, "%s\n", hline);
814 print_gputimes(fplog, "Pair list H2D",
815 gpu_t->pl_h2d_c, gpu_t->pl_h2d_t, tot_gpu);
816 print_gputimes(fplog, "X / q H2D",
817 gpu_t->nb_c, gpu_t->nb_h2d_t, tot_gpu);
819 for (i = 0; i < 2; i++)
821 for (j = 0; j < 2; j++)
823 if (gpu_t->ktime[i][j].c)
825 print_gputimes(fplog, k_log_str[i][j],
826 gpu_t->ktime[i][j].c, gpu_t->ktime[i][j].t, tot_gpu);
831 print_gputimes(fplog, "F D2H", gpu_t->nb_c, gpu_t->nb_d2h_t, tot_gpu);
832 fprintf(fplog, "%s\n", hline);
833 print_gputimes(fplog, "Total ", gpu_t->nb_c, tot_gpu, tot_gpu);
834 fprintf(fplog, "%s\n", hline);
836 gpu_cpu_ratio = tot_gpu/tot_cpu_overlap;
837 fprintf(fplog, "\nForce evaluation time GPU/CPU: %.3f ms/%.3f ms = %.3f\n",
838 tot_gpu/gpu_t->nb_c, tot_cpu_overlap/wc->wcc[ewcFORCE].n,
841 /* only print notes related to CPU-GPU load balance with PME */
842 if (wc->wcc[ewcPMEMESH].n > 0)
844 fprintf(fplog, "For optimal performance this ratio should be close to 1!\n");
846 /* print note if the imbalance is high with PME case in which
847 * CPU-GPU load balancing is possible */
848 if (gpu_cpu_ratio < 0.75 || gpu_cpu_ratio > 1.2)
850 /* Only the sim master calls this function, so always print to stderr */
851 if (gpu_cpu_ratio < 0.75)
855 /* The user could have used -notunepme,
856 * but we currently can't check that here.
858 md_print_warn(NULL, fplog,
859 "\nNOTE: The GPU has >25%% less load than the CPU. This imbalance causes\n"
860 " performance loss. Maybe the domain decomposition limits the PME tuning.\n"
861 " In that case, try setting the DD grid manually (-dd) or lowering -dds.");
865 /* We should not end up here, unless the box is
866 * too small for increasing the cut-off for PME tuning.
868 md_print_warn(NULL, fplog,
869 "\nNOTE: The GPU has >25%% less load than the CPU. This imbalance causes\n"
870 " performance loss.");
873 if (gpu_cpu_ratio > 1.2)
875 md_print_warn(NULL, fplog,
876 "\nNOTE: The GPU has >20%% more load than the CPU. This imbalance causes\n"
877 " performance loss, consider using a shorter cut-off and a finer PME grid.");
885 md_print_warn(NULL, fplog,
886 "MPI_Barrier was called before each cycle start/stop\n"
887 "call, so timings are not those of real runs.\n");
890 if (wc->wcc[ewcNB_XF_BUF_OPS].n > 0 &&
891 (cyc_sum[ewcDOMDEC] > tot*0.1 ||
892 cyc_sum[ewcNS] > tot*0.1))
894 /* Only the sim master calls this function, so always print to stderr */
895 if (wc->wcc[ewcDOMDEC].n == 0)
897 md_print_warn(NULL, fplog,
898 "NOTE: %d %% of the run time was spent in pair search,\n"
899 " you might want to increase nstlist (this has no effect on accuracy)\n",
900 (int)(100*cyc_sum[ewcNS]/tot+0.5));
904 md_print_warn(NULL, fplog,
905 "NOTE: %d %% of the run time was spent in domain decomposition,\n"
906 " %d %% of the run time was spent in pair search,\n"
907 " you might want to increase nstlist (this has no effect on accuracy)\n",
908 (int)(100*cyc_sum[ewcDOMDEC]/tot+0.5),
909 (int)(100*cyc_sum[ewcNS]/tot+0.5));
913 if (cyc_sum[ewcMoveE] > tot*0.05 && tot > 0)
915 /* Only the sim master calls this function, so always print to stderr */
916 md_print_warn(NULL, fplog,
917 "NOTE: %d %% of the run time was spent communicating energies,\n"
918 " you might want to use the -gcom option of mdrun\n",
919 (int)(100*cyc_sum[ewcMoveE]/tot+0.5));
923 extern gmx_int64_t wcycle_get_reset_counters(gmx_wallcycle_t wc)
930 return wc->reset_counters;
933 extern void wcycle_set_reset_counters(gmx_wallcycle_t wc, gmx_int64_t reset_counters)
940 wc->reset_counters = reset_counters;
943 #ifdef GMX_CYCLE_SUBCOUNTERS
945 void wallcycle_sub_start(gmx_wallcycle_t wc, int ewcs)
949 wc->wcsc[ewcs].start = gmx_cycles_read();
953 void wallcycle_sub_start_nocount(gmx_wallcycle_t wc, int ewcs)
960 wallcycle_sub_start(wc, ewcs);
964 void wallcycle_sub_stop(gmx_wallcycle_t wc, int ewcs)
968 wc->wcsc[ewcs].c += gmx_cycles_read() - wc->wcsc[ewcs].start;
975 void wallcycle_sub_start(gmx_wallcycle_t gmx_unused wc, int gmx_unused ewcs)
978 void wallcycle_sub_start_nocount(gmx_wallcycle_t gmx_unused wc, int gmx_unused ewcs)
981 void wallcycle_sub_stop(gmx_wallcycle_t gmx_unused wc, int gmx_unused ewcs)
985 #endif /* GMX_CYCLE_SUBCOUNTERS */