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"
46 #include "gromacs/legacyheaders/md_logging.h"
47 #include "gromacs/legacyheaders/types/commrec.h"
48 #include "gromacs/timing/cyclecounter.h"
49 #include "gromacs/timing/gpu_timing.h"
50 #include "gromacs/utility/cstringutil.h"
51 #include "gromacs/utility/gmxmpi.h"
52 #include "gromacs/utility/smalloc.h"
53 #include "gromacs/utility/snprintf.h"
55 /* DEBUG_WCYCLE adds consistency checking for the counters.
56 * It checks if you stop a counter different from the last
57 * one that was opened and if you do nest too deep.
59 /* #define DEBUG_WCYCLE */
62 #include "gromacs/utility/fatalerror.h"
73 typedef struct gmx_wallcycle
76 /* variables for testing/debugging */
82 int counterlist[DEPTH_MAX];
86 gmx_cycles_t cycle_prev;
87 gmx_int64_t reset_counters;
89 MPI_Comm mpi_comm_mygroup;
93 #ifdef GMX_CYCLE_SUBCOUNTERS
99 /* Each name should not exceed 19 printing characters
100 (ie. terminating null can be twentieth) */
101 static const char *wcn[ewcNR] =
103 "Run", "Step", "PP during PME", "Domain decomp.", "DD comm. load",
104 "DD comm. bounds", "Vsite constr.", "Send X to PME", "Neighbor search", "Launch GPU ops.",
105 "Comm. coord.", "Born radii", "Force", "Wait + Comm. F", "PME mesh",
106 "PME redist. X/F", "PME spread/gather", "PME 3D-FFT", "PME 3D-FFT Comm.", "PME solve LJ", "PME solve Elec",
107 "PME wait for PP", "Wait + Recv. PME F", "Wait GPU nonlocal", "Wait GPU local", "Wait GPU loc. est.", "NB X/F buffer ops.",
108 "Vsite spread", "COM pull force",
109 "Write traj.", "Update", "Constraints", "Comm. energies",
110 "Enforced rotation", "Add rot. forces", "Coordinate swapping", "IMD", "Test"
113 static const char *wcsn[ewcsNR] =
115 "DD redist.", "DD NS grid + sort", "DD setup comm.",
116 "DD make top.", "DD make constr.", "DD top. other",
117 "NS grid local", "NS grid non-loc.", "NS search local", "NS search non-loc.",
121 "Listed buffer ops.",
123 "Ewald F correction",
128 gmx_bool wallcycle_have_counter(void)
130 return gmx_cycles_have_counter();
133 gmx_wallcycle_t wallcycle_init(FILE *fplog, int resetstep, t_commrec gmx_unused *cr,
134 int nthreads_pp, int nthreads_pme)
139 if (!wallcycle_have_counter())
146 wc->wc_barrier = FALSE;
150 wc->reset_counters = resetstep;
151 wc->nthreads_pp = nthreads_pp;
152 wc->nthreads_pme = nthreads_pme;
153 wc->cycles_sum = NULL;
156 if (PAR(cr) && getenv("GMX_CYCLE_BARRIER") != NULL)
160 fprintf(fplog, "\nWill call MPI_Barrier before each cycle start/stop call\n\n");
162 wc->wc_barrier = TRUE;
163 wc->mpi_comm_mygroup = cr->mpi_comm_mygroup;
167 snew(wc->wcc, ewcNR);
168 if (getenv("GMX_CYCLE_ALL") != NULL)
172 fprintf(fplog, "\nWill time all the code during the run\n\n");
174 snew(wc->wcc_all, ewcNR*ewcNR);
177 #ifdef GMX_CYCLE_SUBCOUNTERS
178 snew(wc->wcsc, ewcsNR);
188 void wallcycle_destroy(gmx_wallcycle_t wc)
199 if (wc->wcc_all != NULL)
203 #ifdef GMX_CYCLE_SUBCOUNTERS
204 if (wc->wcsc != NULL)
212 static void wallcycle_all_start(gmx_wallcycle_t wc, int ewc, gmx_cycles_t cycle)
215 wc->cycle_prev = cycle;
218 static void wallcycle_all_stop(gmx_wallcycle_t wc, int ewc, gmx_cycles_t cycle)
220 wc->wcc_all[wc->ewc_prev*ewcNR+ewc].n += 1;
221 wc->wcc_all[wc->ewc_prev*ewcNR+ewc].c += cycle - wc->cycle_prev;
226 static void debug_start_check(gmx_wallcycle_t wc, int ewc)
228 /* fprintf(stderr,"wcycle_start depth %d, %s\n",wc->count_depth,wcn[ewc]); */
230 if (wc->count_depth < 0 || wc->count_depth >= DEPTH_MAX)
232 gmx_fatal(FARGS, "wallcycle counter depth out of range: %d",
235 wc->counterlist[wc->count_depth] = ewc;
239 static void debug_stop_check(gmx_wallcycle_t wc, int ewc)
243 /* fprintf(stderr,"wcycle_stop depth %d, %s\n",wc->count_depth,wcn[ewc]); */
245 if (wc->count_depth < 0)
247 gmx_fatal(FARGS, "wallcycle counter depth out of range when stopping %s: %d", wcn[ewc], wc->count_depth);
249 if (wc->counterlist[wc->count_depth] != ewc)
251 gmx_fatal(FARGS, "wallcycle mismatch at stop, start %s, stop %s",
252 wcn[wc->counterlist[wc->count_depth]], wcn[ewc]);
257 void wallcycle_start(gmx_wallcycle_t wc, int ewc)
269 MPI_Barrier(wc->mpi_comm_mygroup);
274 debug_start_check(wc, ewc);
277 cycle = gmx_cycles_read();
278 wc->wcc[ewc].start = cycle;
279 if (wc->wcc_all != NULL)
284 wallcycle_all_start(wc, ewc, cycle);
286 else if (wc->wc_depth == 3)
288 wallcycle_all_stop(wc, ewc, cycle);
293 void wallcycle_start_nocount(gmx_wallcycle_t wc, int ewc)
300 wallcycle_start(wc, ewc);
304 double wallcycle_stop(gmx_wallcycle_t wc, int ewc)
306 gmx_cycles_t cycle, last;
316 MPI_Barrier(wc->mpi_comm_mygroup);
321 debug_stop_check(wc, ewc);
324 cycle = gmx_cycles_read();
325 last = cycle - wc->wcc[ewc].start;
326 wc->wcc[ewc].c += last;
333 wallcycle_all_stop(wc, ewc, cycle);
335 else if (wc->wc_depth == 2)
337 wallcycle_all_start(wc, ewc, cycle);
344 void wallcycle_get(gmx_wallcycle_t wc, int ewc, int *n, double *c)
347 *c = (double)wc->wcc[ewc].c;
350 void wallcycle_reset_all(gmx_wallcycle_t wc)
359 for (i = 0; i < ewcNR; i++)
366 for (i = 0; i < ewcNR*ewcNR; i++)
368 wc->wcc_all[i].n = 0;
369 wc->wcc_all[i].c = 0;
372 #ifdef GMX_CYCLE_SUBCOUNTERS
373 for (i = 0; i < ewcsNR; i++)
381 static gmx_bool is_pme_counter(int ewc)
383 return (ewc >= ewcPMEMESH && ewc <= ewcPMEWAITCOMM);
386 static gmx_bool is_pme_subcounter(int ewc)
388 return (ewc >= ewcPME_REDISTXF && ewc < ewcPMEWAITCOMM);
391 /* Subtract counter ewc_sub timed inside a timing block for ewc_main */
392 static void subtract_cycles(wallcc_t *wcc, int ewc_main, int ewc_sub)
394 if (wcc[ewc_sub].n > 0)
396 assert(wcc[ewc_main].c >= wcc[ewc_sub].c);
398 wcc[ewc_main].c -= wcc[ewc_sub].c;
402 void wallcycle_sum(t_commrec *cr, gmx_wallcycle_t wc)
405 double cycles[ewcNR+ewcsNR];
406 double cycles_n[ewcNR+ewcsNR], buf[ewcNR+ewcsNR], *cyc_all, *buf_all;
415 snew(wc->cycles_sum, ewcNR+ewcsNR);
419 /* The GPU wait estimate counter is used for load balancing only
420 * and will mess up the total due to double counting: clear it.
422 wcc[ewcWAIT_GPU_NB_L_EST].n = 0;
423 wcc[ewcWAIT_GPU_NB_L_EST].c = 0;
425 for (i = 0; i < ewcNR; i++)
427 if (is_pme_counter(i) || (i == ewcRUN && cr->duty == DUTY_PME))
429 wcc[i].c *= wc->nthreads_pme;
433 for (j = 0; j < ewcNR; j++)
435 wc->wcc_all[i*ewcNR+j].c *= wc->nthreads_pme;
441 wcc[i].c *= wc->nthreads_pp;
445 for (j = 0; j < ewcNR; j++)
447 wc->wcc_all[i*ewcNR+j].c *= wc->nthreads_pp;
453 subtract_cycles(wcc, ewcDOMDEC, ewcDDCOMMLOAD);
454 subtract_cycles(wcc, ewcDOMDEC, ewcDDCOMMBOUND);
456 subtract_cycles(wcc, ewcPME_FFT, ewcPME_FFTCOMM);
458 if (cr->npmenodes == 0)
460 /* All nodes do PME (or no PME at all) */
461 subtract_cycles(wcc, ewcFORCE, ewcPMEMESH);
465 /* The are PME-only nodes */
466 if (wcc[ewcPMEMESH].n > 0)
468 /* This must be a PME only node, calculate the Wait + Comm. time */
469 assert(wcc[ewcRUN].c >= wcc[ewcPMEMESH].c);
470 wcc[ewcPMEWAITCOMM].c = wcc[ewcRUN].c - wcc[ewcPMEMESH].c;
474 /* Store the cycles in a double buffer for summing */
475 for (i = 0; i < ewcNR; i++)
477 cycles_n[i] = (double)wcc[i].n;
478 cycles[i] = (double)wcc[i].c;
481 #ifdef GMX_CYCLE_SUBCOUNTERS
482 for (i = 0; i < ewcsNR; i++)
484 wc->wcsc[i].c *= wc->nthreads_pp;
485 cycles_n[ewcNR+i] = (double)wc->wcsc[i].n;
486 cycles[ewcNR+i] = (double)wc->wcsc[i].c;
494 MPI_Allreduce(cycles_n, buf, nsum, MPI_DOUBLE, MPI_MAX,
496 for (i = 0; i < ewcNR; i++)
498 wcc[i].n = (int)(buf[i] + 0.5);
500 #ifdef GMX_CYCLE_SUBCOUNTERS
501 for (i = 0; i < ewcsNR; i++)
503 wc->wcsc[i].n = (int)(buf[ewcNR+i] + 0.5);
507 MPI_Allreduce(cycles, wc->cycles_sum, nsum, MPI_DOUBLE, MPI_SUM,
510 if (wc->wcc_all != NULL)
512 snew(cyc_all, ewcNR*ewcNR);
513 snew(buf_all, ewcNR*ewcNR);
514 for (i = 0; i < ewcNR*ewcNR; i++)
516 cyc_all[i] = wc->wcc_all[i].c;
518 MPI_Allreduce(cyc_all, buf_all, ewcNR*ewcNR, MPI_DOUBLE, MPI_SUM,
520 for (i = 0; i < ewcNR*ewcNR; i++)
522 wc->wcc_all[i].c = buf_all[i];
531 for (i = 0; i < nsum; i++)
533 wc->cycles_sum[i] = cycles[i];
538 static void print_cycles(FILE *fplog, double c2t, const char *name,
539 int nnodes, int nthreads,
540 int ncalls, double c_sum, double tot)
543 char nthreads_str[6];
546 double percentage = (tot > 0.) ? (100. * c_sum / tot) : 0.;
552 snprintf(ncalls_str, sizeof(ncalls_str), "%10d", ncalls);
555 snprintf(nnodes_str, sizeof(nnodes_str), "N/A");
559 snprintf(nnodes_str, sizeof(nnodes_str), "%4d", nnodes);
563 snprintf(nthreads_str, sizeof(nthreads_str), "N/A");
567 snprintf(nthreads_str, sizeof(nthreads_str), "%4d", nthreads);
576 /* Convert the cycle count to wallclock time for this task */
579 fprintf(fplog, " %-19.19s %4s %4s %10s %10.3f %14.3f %5.1f\n",
580 name, nnodes_str, nthreads_str, ncalls_str, wallt,
581 c_sum*1e-9, percentage);
585 static void print_gputimes(FILE *fplog, const char *name,
586 int n, double t, double tot_t)
593 snprintf(num, sizeof(num), "%10d", n);
594 snprintf(avg_perf, sizeof(avg_perf), "%10.3f", t/n);
599 sprintf(avg_perf, " ");
601 if (t != tot_t && tot_t > 0)
603 fprintf(fplog, " %-29s %10s%12.3f %s %5.1f\n",
604 name, num, t/1000, avg_perf, 100 * t/tot_t);
608 fprintf(fplog, " %-29s %10s%12.3f %s %5.1f\n",
609 name, "", t/1000, avg_perf, 100.0);
613 static void print_header(FILE *fplog, int nrank_pp, int nth_pp, int nrank_pme, int nth_pme)
615 int nrank_tot = nrank_pp + nrank_pme;
618 fprintf(fplog, "On %d MPI rank%s", nrank_tot, nrank_tot == 1 ? "" : "s");
621 fprintf(fplog, ", each using %d OpenMP threads", nth_pp);
623 /* Don't report doing PP+PME, because we can't tell here if
624 * this is RF, etc. */
628 fprintf(fplog, "On %d MPI rank%s doing PP", nrank_pp, nrank_pp == 1 ? "" : "s");
631 fprintf(fplog, ",%s using %d OpenMP threads", nrank_pp > 1 ? " each" : "", nth_pp);
633 fprintf(fplog, ", and\non %d MPI rank%s doing PME", nrank_pme, nrank_pme == 1 ? "" : "s");
636 fprintf(fplog, ",%s using %d OpenMP threads", nrank_pme > 1 ? " each" : "", nth_pme);
640 fprintf(fplog, "\n\n");
641 fprintf(fplog, " Computing: Num Num Call Wall time Giga-Cycles\n");
642 fprintf(fplog, " Ranks Threads Count (s) total sum %%\n");
645 void wallcycle_print(FILE *fplog, int nnodes, int npme, double realtime,
646 gmx_wallcycle_t wc, struct gmx_wallclock_gpu_t *gpu_t)
649 double tot, tot_for_pp, tot_for_rest, tot_gpu, tot_cpu_overlap, gpu_cpu_ratio, tot_k;
650 double c2t, c2t_pp, c2t_pme = 0;
651 int i, j, npp, nth_pp, nth_pme, nth_tot;
653 const char *hline = "-----------------------------------------------------------------------------";
660 nth_pp = wc->nthreads_pp;
663 nth_pme = wc->nthreads_pme;
666 cyc_sum = wc->cycles_sum;
671 /* npme is the number of PME-only ranks used, and we always do PP work */
674 nth_tot = npp*nth_pp + npme*nth_pme;
676 /* When using PME-only nodes, the next line is valid for both
677 PP-only and PME-only nodes because they started ewcRUN at the
679 tot = cyc_sum[ewcRUN];
684 /* TODO This is heavy handed, but until someone reworks the
685 code so that it is provably robust with respect to
686 non-positive values for all possible timer and cycle
687 counters, there is less value gained from printing whatever
688 timing data might still be sensible for some non-Jenkins
689 run, than is lost from diagnosing Jenkins FP exceptions on
690 runs about whose execution time we don't care. */
691 md_print_warn(NULL, fplog, "WARNING: A total of %f CPU cycles was recorded, so mdrun cannot print a time accounting\n", tot);
695 /* Conversion factor from cycles to seconds */
697 c2t_pp = c2t * nth_tot / (double) (npp*nth_pp);
700 c2t_pme = c2t * nth_tot / (double) (npme*nth_pme);
707 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");
709 print_header(fplog, npp, nth_pp, npme, nth_pme);
711 fprintf(fplog, "%s\n", hline);
712 for (i = ewcPPDURINGPME+1; i < ewcNR; i++)
714 if (is_pme_subcounter(i))
716 /* Do not count these at all */
718 else if (npme > 0 && is_pme_counter(i))
720 /* Print timing information for PME-only nodes, but add an
721 * asterisk so the reader of the table can know that the
722 * walltimes are not meant to add up. The asterisk still
723 * fits in the required maximum of 19 characters. */
725 snprintf(buffer, STRLEN, "%s *", wcn[i]);
726 print_cycles(fplog, c2t_pme, buffer,
728 wc->wcc[i].n, cyc_sum[i], tot);
732 /* Print timing information when it is for a PP or PP+PME
734 print_cycles(fplog, c2t_pp, wcn[i],
736 wc->wcc[i].n, cyc_sum[i], tot);
737 tot_for_pp += cyc_sum[i];
740 if (wc->wcc_all != NULL)
742 for (i = 0; i < ewcNR; i++)
744 for (j = 0; j < ewcNR; j++)
746 snprintf(buf, 20, "%-9.9s %-9.9s", wcn[i], wcn[j]);
747 print_cycles(fplog, c2t_pp, buf,
749 wc->wcc_all[i*ewcNR+j].n,
750 wc->wcc_all[i*ewcNR+j].c,
755 tot_for_rest = tot * (npp * nth_pp) / (double) nth_tot;
756 print_cycles(fplog, c2t_pp, "Rest",
758 -1, tot_for_rest - tot_for_pp, tot);
759 fprintf(fplog, "%s\n", hline);
760 print_cycles(fplog, c2t, "Total",
763 fprintf(fplog, "%s\n", hline);
768 "(*) Note that with separate PME ranks, the walltime column actually sums to\n"
769 " twice the total reported, but the cycle count total and %% are correct.\n"
773 if (wc->wcc[ewcPMEMESH].n > 0)
775 fprintf(fplog, " Breakdown of PME mesh computation\n");
776 fprintf(fplog, "%s\n", hline);
777 for (i = ewcPPDURINGPME+1; i < ewcNR; i++)
779 if (is_pme_subcounter(i))
781 print_cycles(fplog, npme > 0 ? c2t_pme : c2t_pp, wcn[i],
782 npme > 0 ? npme : npp, nth_pme,
783 wc->wcc[i].n, cyc_sum[i], tot);
786 fprintf(fplog, "%s\n", hline);
789 #ifdef GMX_CYCLE_SUBCOUNTERS
790 fprintf(fplog, " Breakdown of PP computation\n");
791 fprintf(fplog, "%s\n", hline);
792 for (i = 0; i < ewcsNR; i++)
794 print_cycles(fplog, c2t_pp, wcsn[i],
796 wc->wcsc[i].n, cyc_sum[ewcNR+i], tot);
798 fprintf(fplog, "%s\n", hline);
801 /* print GPU timing summary */
804 const char *k_log_str[2][2] = {
805 {"Nonbonded F kernel", "Nonbonded F+ene k."},
806 {"Nonbonded F+prune k.", "Nonbonded F+ene+prune k."}
809 tot_gpu = gpu_t->pl_h2d_t + gpu_t->nb_h2d_t + gpu_t->nb_d2h_t;
811 /* add up the kernel timings */
813 for (i = 0; i < 2; i++)
815 for (j = 0; j < 2; j++)
817 tot_k += gpu_t->ktime[i][j].t;
822 tot_cpu_overlap = wc->wcc[ewcFORCE].c;
823 if (wc->wcc[ewcPMEMESH].n > 0)
825 tot_cpu_overlap += wc->wcc[ewcPMEMESH].c;
827 tot_cpu_overlap *= realtime*1000/tot; /* convert s to ms */
829 fprintf(fplog, "\n GPU timings\n%s\n", hline);
830 fprintf(fplog, " Computing: Count Wall t (s) ms/step %c\n", '%');
831 fprintf(fplog, "%s\n", hline);
832 print_gputimes(fplog, "Pair list H2D",
833 gpu_t->pl_h2d_c, gpu_t->pl_h2d_t, tot_gpu);
834 print_gputimes(fplog, "X / q H2D",
835 gpu_t->nb_c, gpu_t->nb_h2d_t, tot_gpu);
837 for (i = 0; i < 2; i++)
839 for (j = 0; j < 2; j++)
841 if (gpu_t->ktime[i][j].c)
843 print_gputimes(fplog, k_log_str[i][j],
844 gpu_t->ktime[i][j].c, gpu_t->ktime[i][j].t, tot_gpu);
849 print_gputimes(fplog, "F D2H", gpu_t->nb_c, gpu_t->nb_d2h_t, tot_gpu);
850 fprintf(fplog, "%s\n", hline);
851 print_gputimes(fplog, "Total ", gpu_t->nb_c, tot_gpu, tot_gpu);
852 fprintf(fplog, "%s\n", hline);
854 gpu_cpu_ratio = tot_gpu/tot_cpu_overlap;
855 if (gpu_t->nb_c > 0 && wc->wcc[ewcFORCE].n > 0)
857 fprintf(fplog, "\nForce evaluation time GPU/CPU: %.3f ms/%.3f ms = %.3f\n",
858 tot_gpu/gpu_t->nb_c, tot_cpu_overlap/wc->wcc[ewcFORCE].n,
862 /* only print notes related to CPU-GPU load balance with PME */
863 if (wc->wcc[ewcPMEMESH].n > 0)
865 fprintf(fplog, "For optimal performance this ratio should be close to 1!\n");
867 /* print note if the imbalance is high with PME case in which
868 * CPU-GPU load balancing is possible */
869 if (gpu_cpu_ratio < 0.75 || gpu_cpu_ratio > 1.2)
871 /* Only the sim master calls this function, so always print to stderr */
872 if (gpu_cpu_ratio < 0.75)
876 /* The user could have used -notunepme,
877 * but we currently can't check that here.
879 md_print_warn(NULL, fplog,
880 "\nNOTE: The GPU has >25%% less load than the CPU. This imbalance causes\n"
881 " performance loss. Maybe the domain decomposition limits the PME tuning.\n"
882 " In that case, try setting the DD grid manually (-dd) or lowering -dds.");
886 /* We should not end up here, unless the box is
887 * too small for increasing the cut-off for PME tuning.
889 md_print_warn(NULL, fplog,
890 "\nNOTE: The GPU has >25%% less load than the CPU. This imbalance causes\n"
891 " performance loss.");
894 if (gpu_cpu_ratio > 1.2)
896 md_print_warn(NULL, fplog,
897 "\nNOTE: The GPU has >20%% more load than the CPU. This imbalance causes\n"
898 " performance loss, consider using a shorter cut-off and a finer PME grid.");
906 md_print_warn(NULL, fplog,
907 "MPI_Barrier was called before each cycle start/stop\n"
908 "call, so timings are not those of real runs.\n");
911 if (wc->wcc[ewcNB_XF_BUF_OPS].n > 0 &&
912 (cyc_sum[ewcDOMDEC] > tot*0.1 ||
913 cyc_sum[ewcNS] > tot*0.1))
915 /* Only the sim master calls this function, so always print to stderr */
916 if (wc->wcc[ewcDOMDEC].n == 0)
918 md_print_warn(NULL, fplog,
919 "NOTE: %d %% of the run time was spent in pair search,\n"
920 " you might want to increase nstlist (this has no effect on accuracy)\n",
921 (int)(100*cyc_sum[ewcNS]/tot+0.5));
925 md_print_warn(NULL, fplog,
926 "NOTE: %d %% of the run time was spent in domain decomposition,\n"
927 " %d %% of the run time was spent in pair search,\n"
928 " you might want to increase nstlist (this has no effect on accuracy)\n",
929 (int)(100*cyc_sum[ewcDOMDEC]/tot+0.5),
930 (int)(100*cyc_sum[ewcNS]/tot+0.5));
934 if (cyc_sum[ewcMoveE] > tot*0.05)
936 /* Only the sim master calls this function, so always print to stderr */
937 md_print_warn(NULL, fplog,
938 "NOTE: %d %% of the run time was spent communicating energies,\n"
939 " you might want to use the -gcom option of mdrun\n",
940 (int)(100*cyc_sum[ewcMoveE]/tot+0.5));
944 extern gmx_int64_t wcycle_get_reset_counters(gmx_wallcycle_t wc)
951 return wc->reset_counters;
954 extern void wcycle_set_reset_counters(gmx_wallcycle_t wc, gmx_int64_t reset_counters)
961 wc->reset_counters = reset_counters;
964 #ifdef GMX_CYCLE_SUBCOUNTERS
966 void wallcycle_sub_start(gmx_wallcycle_t wc, int ewcs)
970 wc->wcsc[ewcs].start = gmx_cycles_read();
974 void wallcycle_sub_start_nocount(gmx_wallcycle_t wc, int ewcs)
981 wallcycle_sub_start(wc, ewcs);
985 void wallcycle_sub_stop(gmx_wallcycle_t wc, int ewcs)
989 wc->wcsc[ewcs].c += gmx_cycles_read() - wc->wcsc[ewcs].start;
996 void wallcycle_sub_start(gmx_wallcycle_t gmx_unused wc, int gmx_unused ewcs)
999 void wallcycle_sub_start_nocount(gmx_wallcycle_t gmx_unused wc, int gmx_unused ewcs)
1002 void wallcycle_sub_stop(gmx_wallcycle_t gmx_unused wc, int gmx_unused ewcs)
1006 #endif /* GMX_CYCLE_SUBCOUNTERS */