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) 2013,2014,2015,2016,2017 by the GROMACS development team.
7 * Copyright (c) 2018,2019,2020,2021, by the GROMACS development team, led by
8 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
9 * and including many others, as listed in the AUTHORS file in the
10 * top-level source directory and at http://www.gromacs.org.
12 * GROMACS is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public License
14 * as published by the Free Software Foundation; either version 2.1
15 * of the License, or (at your option) any later version.
17 * GROMACS is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with GROMACS; if not, see
24 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 * If you want to redistribute modifications to GROMACS, please
28 * consider that scientific software is very special. Version
29 * control is crucial - bugs must be traceable. We will be happy to
30 * consider code for inclusion in the official distribution, but
31 * derived work must not be called official GROMACS. Details are found
32 * in the README & COPYING files - if they are missing, get the
33 * official version at http://www.gromacs.org.
35 * To help us fund GROMACS development, we humbly ask that you cite
36 * the research papers on the package. Check out http://www.gromacs.org.
40 #include "wallcycle.h"
49 #include "gromacs/math/functions.h"
50 #include "gromacs/mdtypes/commrec.h"
51 #include "gromacs/timing/cyclecounter.h"
52 #include "gromacs/timing/gpu_timing.h"
53 #include "gromacs/timing/wallcyclereporting.h"
54 #include "gromacs/utility/cstringutil.h"
55 #include "gromacs/utility/gmxassert.h"
56 #include "gromacs/utility/gmxmpi.h"
57 #include "gromacs/utility/logger.h"
58 #include "gromacs/utility/smalloc.h"
59 #include "gromacs/utility/snprintf.h"
61 static const bool useCycleSubcounters = GMX_CYCLE_SUBCOUNTERS;
64 /*! \brief Enables consistency checking for the counters.
66 * If the macro is set to 1, code checks if you stop a counter different from the last
67 * one that was opened and if you do nest too deep.
69 # define DEBUG_WCYCLE 0
71 //! Whether wallcycle debugging is enabled
72 constexpr bool gmx_unused enableWallcycleDebug = (DEBUG_WCYCLE != 0);
73 //! True if only the master rank should print debugging output
74 constexpr bool gmx_unused onlyMasterDebugPrints = true;
75 //! True if cycle counter nesting depth debuggin prints are enabled
76 constexpr bool gmx_unused debugPrintDepth = false /* enableWallcycleDebug */;
79 # include "gromacs/utility/fatalerror.h"
92 /* did we detect one or more invalid cycle counts */
93 gmx_bool haveInvalidCount;
94 /* variables for testing/debugging */
100 int counterlist[DEPTH_MAX];
105 gmx_cycles_t cycle_prev;
106 int64_t reset_counters;
108 MPI_Comm mpi_comm_mygroup;
113 /* Each name should not exceed 19 printing characters
114 (ie. terminating null can be twentieth) */
115 static const char* wcn[ewcNR] = { "Run",
137 "Wait + Recv. PME F",
138 "Wait PME GPU spread",
140 "PME solve", /* the strings for FFT/solve are repeated here for mixed mode counters */
141 "Wait PME GPU gather",
144 "Wait GPU NB nonloc.",
146 "Wait GPU state copy",
147 "NB X/F buffer ops.",
161 static const char* wcsn[ewcsNR] = {
172 "NS search non-loc.",
176 "Listed buffer ops.",
178 "Nonbonded F kernel",
181 "Launch NB GPU tasks",
182 "Launch Bonded GPU tasks",
183 "Launch PME GPU tasks",
185 "Ewald F correction",
188 "Clear force buffer",
189 "Launch GPU NB X buffer ops.",
190 "Launch GPU NB F buffer ops.",
191 "Launch GPU Comm. coord.",
192 "Launch GPU Comm. force.",
197 /* PME GPU timing events' names - correspond to the enum in the gpu_timing.h */
198 static const char* enumValuetoString(PmeStage enumValue)
200 constexpr gmx::EnumerationArray<PmeStage, const char*> pmeStageNames = {
201 "PME spline", "PME spread", "PME spline + spread", "PME 3D-FFT r2c",
202 "PME solve", "PME 3D-FFT c2r", "PME gather"
204 return pmeStageNames[enumValue];
207 gmx_bool wallcycle_have_counter()
209 return gmx_cycles_have_counter();
212 gmx_wallcycle_t wallcycle_init(FILE* fplog, int resetstep, t_commrec gmx_unused* cr)
217 if (!wallcycle_have_counter())
224 wc->haveInvalidCount = FALSE;
225 wc->wc_barrier = FALSE;
226 wc->wcc_all = nullptr;
229 wc->reset_counters = resetstep;
233 if (PAR(cr) && getenv("GMX_CYCLE_BARRIER") != nullptr)
237 fprintf(fplog, "\nWill call MPI_Barrier before each cycle start/stop call\n\n");
239 wc->wc_barrier = TRUE;
240 wc->mpi_comm_mygroup = cr->mpi_comm_mygroup;
244 snew(wc->wcc, ewcNR);
245 if (getenv("GMX_CYCLE_ALL") != nullptr)
249 fprintf(fplog, "\nWill time all the code during the run\n\n");
251 snew(wc->wcc_all, ewcNR * ewcNR);
254 if (useCycleSubcounters)
256 snew(wc->wcsc, ewcsNR);
261 wc->isMasterRank = MASTER(cr);
267 void wallcycle_destroy(gmx_wallcycle_t wc)
274 if (wc->wcc != nullptr)
278 if (wc->wcc_all != nullptr)
282 if (wc->wcsc != nullptr)
289 static void wallcycle_all_start(gmx_wallcycle_t wc, int ewc, gmx_cycles_t cycle)
292 wc->cycle_prev = cycle;
295 static void wallcycle_all_stop(gmx_wallcycle_t wc, int ewc, gmx_cycles_t cycle)
297 wc->wcc_all[wc->ewc_prev * ewcNR + ewc].n += 1;
298 wc->wcc_all[wc->ewc_prev * ewcNR + ewc].c += cycle - wc->cycle_prev;
303 static void debug_start_check(gmx_wallcycle_t wc, int ewc)
305 if (wc->count_depth < 0 || wc->count_depth >= DEPTH_MAX)
307 gmx_fatal(FARGS, "wallcycle counter depth out of range: %d", wc->count_depth + 1);
309 wc->counterlist[wc->count_depth] = ewc;
312 if (debugPrintDepth && (!onlyMasterDebugPrints || wc->isMasterRank))
314 std::string indentStr(4 * wc->count_depth, ' ');
315 fprintf(stderr, "%swcycle_start depth %d, %s\n", indentStr.c_str(), wc->count_depth, wcn[ewc]);
319 static void debug_stop_check(gmx_wallcycle_t wc, int ewc)
321 if (debugPrintDepth && (!onlyMasterDebugPrints || wc->isMasterRank))
323 std::string indentStr(4 * wc->count_depth, ' ');
324 fprintf(stderr, "%swcycle_stop depth %d, %s\n", indentStr.c_str(), wc->count_depth, wcn[ewc]);
329 if (wc->count_depth < 0)
331 gmx_fatal(FARGS, "wallcycle counter depth out of range when stopping %s: %d", wcn[ewc], wc->count_depth);
333 if (wc->counterlist[wc->count_depth] != ewc)
336 "wallcycle mismatch at stop, start %s, stop %s",
337 wcn[wc->counterlist[wc->count_depth]],
343 void wallcycle_start(gmx_wallcycle_t wc, int ewc)
355 MPI_Barrier(wc->mpi_comm_mygroup);
360 debug_start_check(wc, ewc);
363 cycle = gmx_cycles_read();
364 wc->wcc[ewc].start = cycle;
365 if (wc->wcc_all != nullptr)
370 wallcycle_all_start(wc, ewc, cycle);
372 else if (wc->wc_depth == 3)
374 wallcycle_all_stop(wc, ewc, cycle);
379 void wallcycle_increment_event_count(gmx_wallcycle_t wc, int ewc)
388 void wallcycle_start_nocount(gmx_wallcycle_t wc, int ewc)
395 wallcycle_start(wc, ewc);
399 double wallcycle_stop(gmx_wallcycle_t wc, int ewc)
401 gmx_cycles_t cycle, last;
411 MPI_Barrier(wc->mpi_comm_mygroup);
416 debug_stop_check(wc, ewc);
419 /* When processes or threads migrate between cores, the cycle counting
420 * can get messed up if the cycle counter on different cores are not
421 * synchronized. When this happens we expect both large negative and
422 * positive cycle differences. We can detect negative cycle differences.
423 * Detecting too large positive counts if difficult, since count can be
424 * large, especially for ewcRUN. If we detect a negative count,
425 * we will not print the cycle accounting table.
427 cycle = gmx_cycles_read();
428 if (cycle >= wc->wcc[ewc].start)
430 last = cycle - wc->wcc[ewc].start;
435 wc->haveInvalidCount = TRUE;
437 wc->wcc[ewc].c += last;
444 wallcycle_all_stop(wc, ewc, cycle);
446 else if (wc->wc_depth == 2)
448 wallcycle_all_start(wc, ewc, cycle);
455 void wallcycle_get(gmx_wallcycle_t wc, int ewc, int* n, double* c)
458 *c = static_cast<double>(wc->wcc[ewc].c);
461 void wallcycle_reset_all(gmx_wallcycle_t wc)
470 for (i = 0; i < ewcNR; i++)
475 wc->haveInvalidCount = FALSE;
479 for (i = 0; i < ewcNR * ewcNR; i++)
481 wc->wcc_all[i].n = 0;
482 wc->wcc_all[i].c = 0;
487 for (i = 0; i < ewcsNR; i++)
495 static gmx_bool is_pme_counter(int ewc)
497 return (ewc >= ewcPMEMESH && ewc <= ewcPMEWAITCOMM);
500 static gmx_bool is_pme_subcounter(int ewc)
502 return (ewc >= ewcPME_REDISTXF && ewc < ewcPMEWAITCOMM);
505 /* Subtract counter ewc_sub timed inside a timing block for ewc_main */
506 static void subtract_cycles(wallcc_t* wcc, int ewc_main, int ewc_sub)
508 if (wcc[ewc_sub].n > 0)
510 if (wcc[ewc_main].c >= wcc[ewc_sub].c)
512 wcc[ewc_main].c -= wcc[ewc_sub].c;
516 /* Something is wrong with the cycle counting */
522 void wallcycle_scale_by_num_threads(gmx_wallcycle_t wc, bool isPmeRank, int nthreads_pp, int nthreads_pme)
529 for (int i = 0; i < ewcNR; i++)
531 if (is_pme_counter(i) || (i == ewcRUN && isPmeRank))
533 wc->wcc[i].c *= nthreads_pme;
537 for (int j = 0; j < ewcNR; j++)
539 wc->wcc_all[i * ewcNR + j].c *= nthreads_pme;
545 wc->wcc[i].c *= nthreads_pp;
549 for (int j = 0; j < ewcNR; j++)
551 wc->wcc_all[i * ewcNR + j].c *= nthreads_pp;
556 if (useCycleSubcounters && wc->wcsc && !isPmeRank)
558 for (int i = 0; i < ewcsNR; i++)
560 wc->wcsc[i].c *= nthreads_pp;
565 /* TODO Make an object for this function to return, containing some
566 * vectors of something like wallcc_t for the summed wcc, wcc_all and
567 * wcsc, AND the original wcc for rank 0.
569 * The GPU timing is reported only for rank 0, so we want to preserve
570 * the original wcycle on that rank. Rank 0 also reports the global
571 * counts before that, so needs something to contain the global data
572 * without over-writing the rank-0 data. The current implementation
573 * uses cycles_sum to manage this, which works OK now because wcsc and
574 * wcc_all are unused by the GPU reporting, but it is not satisfactory
575 * for the future. Also, there's no need for MPI_Allreduce, since
576 * only MASTERRANK uses any of the results. */
577 WallcycleCounts wallcycle_sum(const t_commrec* cr, gmx_wallcycle_t wc)
579 WallcycleCounts cycles_sum;
581 double cycles[int(ewcNR) + int(ewcsNR)];
583 double cycles_n[int(ewcNR) + int(ewcsNR) + 1];
590 /* Default construction of std::array of non-class T can leave
591 the values indeterminate, just like a C array */
598 subtract_cycles(wcc, ewcDOMDEC, ewcDDCOMMLOAD);
599 subtract_cycles(wcc, ewcDOMDEC, ewcDDCOMMBOUND);
601 subtract_cycles(wcc, ewcPME_FFT, ewcPME_FFTCOMM);
603 if (cr->npmenodes == 0)
605 /* All nodes do PME (or no PME at all) */
606 subtract_cycles(wcc, ewcFORCE, ewcPMEMESH);
610 /* The are PME-only nodes */
611 if (wcc[ewcPMEMESH].n > 0)
613 /* This must be a PME only node, calculate the Wait + Comm. time */
614 GMX_ASSERT(wcc[ewcRUN].c >= wcc[ewcPMEMESH].c,
615 "Total run ticks must be greater than PME-only ticks");
616 wcc[ewcPMEWAITCOMM].c = wcc[ewcRUN].c - wcc[ewcPMEMESH].c;
620 /* Store the cycles in a double buffer for summing */
621 for (i = 0; i < ewcNR; i++)
624 cycles_n[i] = static_cast<double>(wcc[i].n);
626 cycles[i] = static_cast<double>(wcc[i].c);
631 for (i = 0; i < ewcsNR; i++)
634 cycles_n[ewcNR + i] = static_cast<double>(wc->wcsc[i].n);
636 cycles[ewcNR + i] = static_cast<double>(wc->wcsc[i].c);
644 double buf[int(ewcNR) + int(ewcsNR) + 1];
646 // TODO this code is used only at the end of the run, so we
647 // can just do a simple reduce of haveInvalidCount in
648 // wallcycle_print, and avoid bugs
649 cycles_n[nsum] = (wc->haveInvalidCount ? 1 : 0);
650 // TODO Use MPI_Reduce
651 MPI_Allreduce(cycles_n, buf, nsum + 1, MPI_DOUBLE, MPI_MAX, cr->mpi_comm_mysim);
652 for (i = 0; i < ewcNR; i++)
654 wcc[i].n = gmx::roundToInt(buf[i]);
656 wc->haveInvalidCount = (buf[nsum] > 0);
659 for (i = 0; i < ewcsNR; i++)
661 wc->wcsc[i].n = gmx::roundToInt(buf[ewcNR + i]);
665 // TODO Use MPI_Reduce
666 MPI_Allreduce(cycles, cycles_sum.data(), nsum, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
668 if (wc->wcc_all != nullptr)
670 double *buf_all, *cyc_all;
672 snew(cyc_all, ewcNR * ewcNR);
673 snew(buf_all, ewcNR * ewcNR);
674 for (i = 0; i < ewcNR * ewcNR; i++)
676 cyc_all[i] = wc->wcc_all[i].c;
678 // TODO Use MPI_Reduce
679 MPI_Allreduce(cyc_all, buf_all, ewcNR * ewcNR, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
680 for (i = 0; i < ewcNR * ewcNR; i++)
682 wc->wcc_all[i].c = static_cast<gmx_cycles_t>(buf_all[i]);
691 for (i = 0; i < nsum; i++)
693 cycles_sum[i] = cycles[i];
701 print_cycles(FILE* fplog, double c2t, const char* name, int nnodes, int nthreads, int ncalls, double c_sum, double tot)
703 char nnodes_str[STRLEN];
704 char nthreads_str[STRLEN];
705 char ncalls_str[STRLEN];
707 double percentage = (tot > 0.) ? (100. * c_sum / tot) : 0.;
713 snprintf(ncalls_str, sizeof(ncalls_str), "%10d", ncalls);
716 snprintf(nnodes_str, sizeof(nnodes_str), "N/A");
720 snprintf(nnodes_str, sizeof(nnodes_str), "%4d", nnodes);
724 snprintf(nthreads_str, sizeof(nthreads_str), "N/A");
728 snprintf(nthreads_str, sizeof(nthreads_str), "%4d", nthreads);
737 /* Convert the cycle count to wallclock time for this task */
741 " %-19.19s %4s %4s %10s %10.3f %14.3f %5.1f\n",
752 static void print_gputimes(FILE* fplog, const char* name, int n, double t, double tot_t)
759 snprintf(num, sizeof(num), "%10d", n);
760 snprintf(avg_perf, sizeof(avg_perf), "%10.3f", t / n);
765 sprintf(avg_perf, " ");
767 if (t != tot_t && tot_t > 0)
769 fprintf(fplog, " %-29s %10s%12.3f %s %5.1f\n", name, num, t / 1000, avg_perf, 100 * t / tot_t);
773 fprintf(fplog, " %-29s %10s%12.3f %s %5.1f\n", name, "", t / 1000, avg_perf, 100.0);
777 static void print_header(FILE* fplog, int nrank_pp, int nth_pp, int nrank_pme, int nth_pme)
779 int nrank_tot = nrank_pp + nrank_pme;
782 fprintf(fplog, "On %d MPI rank%s", nrank_tot, nrank_tot == 1 ? "" : "s");
785 fprintf(fplog, ", each using %d OpenMP threads", nth_pp);
787 /* Don't report doing PP+PME, because we can't tell here if
788 * this is RF, etc. */
792 fprintf(fplog, "On %d MPI rank%s doing PP", nrank_pp, nrank_pp == 1 ? "" : "s");
795 fprintf(fplog, ",%s using %d OpenMP threads", nrank_pp > 1 ? " each" : "", nth_pp);
797 fprintf(fplog, ", and\non %d MPI rank%s doing PME", nrank_pme, nrank_pme == 1 ? "" : "s");
800 fprintf(fplog, ",%s using %d OpenMP threads", nrank_pme > 1 ? " each" : "", nth_pme);
804 fprintf(fplog, "\n\n");
805 fprintf(fplog, " Computing: Num Num Call Wall time Giga-Cycles\n");
806 fprintf(fplog, " Ranks Threads Count (s) total sum %%\n");
810 void wallcycle_print(FILE* fplog,
811 const gmx::MDLogger& mdlog,
818 const WallcycleCounts& cyc_sum,
819 const gmx_wallclock_gpu_nbnxn_t* gpu_nbnxn_t,
820 const gmx_wallclock_gpu_pme_t* gpu_pme_t)
822 double tot, tot_for_pp, tot_for_rest, tot_cpu_overlap, gpu_cpu_ratio;
823 double c2t, c2t_pp, c2t_pme = 0;
824 int i, j, npp, nth_tot;
827 "-----------------------------------------------------------------------------";
834 GMX_ASSERT(nth_pp > 0, "Number of particle-particle threads must be >0");
835 GMX_ASSERT(nth_pme > 0, "Number of PME threads must be >0");
836 GMX_ASSERT(nnodes > 0, "Number of nodes must be >0");
837 GMX_ASSERT(npme >= 0, "Number of PME nodes cannot be negative");
839 /* npme is the number of PME-only ranks used, and we always do PP work */
840 GMX_ASSERT(npp > 0, "Number of particle-particle nodes must be >0");
842 nth_tot = npp * nth_pp + npme * nth_pme;
844 /* When using PME-only nodes, the next line is valid for both
845 PP-only and PME-only nodes because they started ewcRUN at the
847 tot = cyc_sum[ewcRUN];
852 /* TODO This is heavy handed, but until someone reworks the
853 code so that it is provably robust with respect to
854 non-positive values for all possible timer and cycle
855 counters, there is less value gained from printing whatever
856 timing data might still be sensible for some non-Jenkins
857 run, than is lost from diagnosing Jenkins FP exceptions on
858 runs about whose execution time we don't care. */
859 GMX_LOG(mdlog.warning)
861 .appendTextFormatted(
862 "WARNING: A total of %f CPU cycles was recorded, so mdrun cannot print a "
868 if (wc->haveInvalidCount)
870 GMX_LOG(mdlog.warning)
873 "NOTE: Detected invalid cycle counts, probably because threads moved "
874 "between CPU cores that do not have synchronized cycle counters. Will not "
875 "print the cycle accounting.");
880 /* Conversion factor from cycles to seconds */
881 c2t = realtime / tot;
882 c2t_pp = c2t * nth_tot / static_cast<double>(npp * nth_pp);
885 c2t_pme = c2t * nth_tot / static_cast<double>(npme * nth_pme);
892 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");
894 print_header(fplog, npp, nth_pp, npme, nth_pme);
896 fprintf(fplog, "%s\n", hline);
897 for (i = ewcPPDURINGPME + 1; i < ewcNR; i++)
899 if (is_pme_subcounter(i))
901 /* Do not count these at all */
903 else if (npme > 0 && is_pme_counter(i))
905 /* Print timing information for PME-only nodes, but add an
906 * asterisk so the reader of the table can know that the
907 * walltimes are not meant to add up. The asterisk still
908 * fits in the required maximum of 19 characters. */
910 snprintf(buffer, STRLEN, "%s *", wcn[i]);
911 print_cycles(fplog, c2t_pme, buffer, npme, nth_pme, wc->wcc[i].n, cyc_sum[i], tot);
915 /* Print timing information when it is for a PP or PP+PME
917 print_cycles(fplog, c2t_pp, wcn[i], npp, nth_pp, wc->wcc[i].n, cyc_sum[i], tot);
918 tot_for_pp += cyc_sum[i];
921 if (wc->wcc_all != nullptr)
923 for (i = 0; i < ewcNR; i++)
925 for (j = 0; j < ewcNR; j++)
927 snprintf(buf, 20, "%-9.9s %-9.9s", wcn[i], wcn[j]);
933 wc->wcc_all[i * ewcNR + j].n,
934 wc->wcc_all[i * ewcNR + j].c,
939 tot_for_rest = tot * npp * nth_pp / static_cast<double>(nth_tot);
940 print_cycles(fplog, c2t_pp, "Rest", npp, nth_pp, -1, tot_for_rest - tot_for_pp, tot);
941 fprintf(fplog, "%s\n", hline);
942 print_cycles(fplog, c2t, "Total", npp, nth_pp, -1, tot, tot);
943 fprintf(fplog, "%s\n", hline);
948 "(*) Note that with separate PME ranks, the walltime column actually sums to\n"
949 " twice the total reported, but the cycle count total and %% are correct.\n"
954 if (wc->wcc[ewcPMEMESH].n > 0)
956 // A workaround to not print breakdown when no subcounters were recorded.
957 // TODO: figure out and record PME GPU counters (what to do with the waiting ones?)
958 std::vector<int> validPmeSubcounterIndices;
959 for (i = ewcPPDURINGPME + 1; i < ewcNR; i++)
961 if (is_pme_subcounter(i) && wc->wcc[i].n > 0)
963 validPmeSubcounterIndices.push_back(i);
967 if (!validPmeSubcounterIndices.empty())
969 fprintf(fplog, " Breakdown of PME mesh computation\n");
970 fprintf(fplog, "%s\n", hline);
971 for (auto i : validPmeSubcounterIndices)
974 npme > 0 ? c2t_pme : c2t_pp,
976 npme > 0 ? npme : npp,
982 fprintf(fplog, "%s\n", hline);
986 if (useCycleSubcounters && wc->wcsc)
988 fprintf(fplog, " Breakdown of PP computation\n");
989 fprintf(fplog, "%s\n", hline);
990 for (i = 0; i < ewcsNR; i++)
992 print_cycles(fplog, c2t_pp, wcsn[i], npp, nth_pp, wc->wcsc[i].n, cyc_sum[ewcNR + i], tot);
994 fprintf(fplog, "%s\n", hline);
997 /* print GPU timing summary */
998 double tot_gpu = 0.0;
1001 for (auto key : keysOf(gpu_pme_t->timing))
1003 tot_gpu += gpu_pme_t->timing[key].t;
1008 const char* k_log_str[2][2] = { { "Nonbonded F kernel", "Nonbonded F+ene k." },
1009 { "Nonbonded F+prune k.", "Nonbonded F+ene+prune k." } };
1010 tot_gpu += gpu_nbnxn_t->pl_h2d_t + gpu_nbnxn_t->nb_h2d_t + gpu_nbnxn_t->nb_d2h_t;
1012 /* add up the kernel timings */
1013 for (i = 0; i < 2; i++)
1015 for (j = 0; j < 2; j++)
1017 tot_gpu += gpu_nbnxn_t->ktime[i][j].t;
1020 tot_gpu += gpu_nbnxn_t->pruneTime.t;
1022 tot_cpu_overlap = wc->wcc[ewcFORCE].c;
1023 if (wc->wcc[ewcPMEMESH].n > 0)
1025 tot_cpu_overlap += wc->wcc[ewcPMEMESH].c;
1027 tot_cpu_overlap *= realtime * 1000 / tot; /* convert s to ms */
1029 fprintf(fplog, "\n GPU timings\n%s\n", hline);
1031 " Computing: Count Wall t (s) ms/step %c\n",
1033 fprintf(fplog, "%s\n", hline);
1034 print_gputimes(fplog, "Pair list H2D", gpu_nbnxn_t->pl_h2d_c, gpu_nbnxn_t->pl_h2d_t, tot_gpu);
1035 print_gputimes(fplog, "X / q H2D", gpu_nbnxn_t->nb_c, gpu_nbnxn_t->nb_h2d_t, tot_gpu);
1037 for (i = 0; i < 2; i++)
1039 for (j = 0; j < 2; j++)
1041 if (gpu_nbnxn_t->ktime[i][j].c)
1043 print_gputimes(fplog,
1045 gpu_nbnxn_t->ktime[i][j].c,
1046 gpu_nbnxn_t->ktime[i][j].t,
1053 for (auto key : keysOf(gpu_pme_t->timing))
1055 if (gpu_pme_t->timing[key].c)
1057 print_gputimes(fplog,
1058 enumValuetoString(key),
1059 gpu_pme_t->timing[key].c,
1060 gpu_pme_t->timing[key].t,
1065 if (gpu_nbnxn_t->pruneTime.c)
1067 print_gputimes(fplog, "Pruning kernel", gpu_nbnxn_t->pruneTime.c, gpu_nbnxn_t->pruneTime.t, tot_gpu);
1069 print_gputimes(fplog, "F D2H", gpu_nbnxn_t->nb_c, gpu_nbnxn_t->nb_d2h_t, tot_gpu);
1070 fprintf(fplog, "%s\n", hline);
1071 print_gputimes(fplog, "Total ", gpu_nbnxn_t->nb_c, tot_gpu, tot_gpu);
1072 fprintf(fplog, "%s\n", hline);
1073 if (gpu_nbnxn_t->dynamicPruneTime.c)
1075 /* We print the dynamic pruning kernel timings after a separator
1076 * and avoid adding it to tot_gpu as this is not in the force
1077 * overlap. We print the fraction as relative to the rest.
1079 print_gputimes(fplog,
1081 gpu_nbnxn_t->dynamicPruneTime.c,
1082 gpu_nbnxn_t->dynamicPruneTime.t,
1084 fprintf(fplog, "%s\n", hline);
1086 gpu_cpu_ratio = tot_gpu / tot_cpu_overlap;
1087 if (gpu_nbnxn_t->nb_c > 0 && wc->wcc[ewcFORCE].n > 0)
1090 "\nAverage per-step force GPU/CPU evaluation time ratio: %.3f ms/%.3f ms = "
1092 tot_gpu / gpu_nbnxn_t->nb_c,
1093 tot_cpu_overlap / wc->wcc[ewcFORCE].n,
1097 /* only print notes related to CPU-GPU load balance with PME */
1098 if (wc->wcc[ewcPMEMESH].n > 0)
1100 fprintf(fplog, "For optimal resource utilization this ratio should be close to 1\n");
1102 /* print note if the imbalance is high with PME case in which
1103 * CPU-GPU load balancing is possible */
1104 if (gpu_cpu_ratio < 0.8 || gpu_cpu_ratio > 1.25)
1106 /* Only the sim master calls this function, so always print to stderr */
1107 if (gpu_cpu_ratio < 0.8)
1111 /* The user could have used -notunepme,
1112 * but we currently can't check that here.
1114 GMX_LOG(mdlog.warning)
1117 "NOTE: The CPU has >25% more load than the GPU. This "
1118 "imbalance wastes\n"
1119 " GPU resources. Maybe the domain decomposition "
1120 "limits the PME tuning.\n"
1121 " In that case, try setting the DD grid manually "
1122 "(-dd) or lowering -dds.");
1126 /* We should not end up here, unless the box is
1127 * too small for increasing the cut-off for PME tuning.
1129 GMX_LOG(mdlog.warning)
1132 "NOTE: The CPU has >25% more load than the GPU. This "
1133 "imbalance wastes\n"
1137 if (gpu_cpu_ratio > 1.25)
1139 GMX_LOG(mdlog.warning)
1142 "NOTE: The GPU has >25% more load than the CPU. This imbalance "
1152 GMX_LOG(mdlog.warning)
1155 "MPI_Barrier was called before each cycle start/stop\n"
1156 "call, so timings are not those of real runs.");
1159 if (wc->wcc[ewcNB_XF_BUF_OPS].n > 0 && (cyc_sum[ewcDOMDEC] > tot * 0.1 || cyc_sum[ewcNS] > tot * 0.1))
1161 /* Only the sim master calls this function, so always print to stderr */
1162 if (wc->wcc[ewcDOMDEC].n == 0)
1164 GMX_LOG(mdlog.warning)
1166 .appendTextFormatted(
1167 "NOTE: %d %% of the run time was spent in pair search,\n"
1168 " you might want to increase nstlist (this has no effect on "
1170 gmx::roundToInt(100 * cyc_sum[ewcNS] / tot));
1174 GMX_LOG(mdlog.warning)
1176 .appendTextFormatted(
1177 "NOTE: %d %% of the run time was spent in domain decomposition,\n"
1178 " %d %% of the run time was spent in pair search,\n"
1179 " you might want to increase nstlist (this has no effect on "
1181 gmx::roundToInt(100 * cyc_sum[ewcDOMDEC] / tot),
1182 gmx::roundToInt(100 * cyc_sum[ewcNS] / tot));
1186 if (cyc_sum[ewcMoveE] > tot * 0.05)
1188 GMX_LOG(mdlog.warning)
1190 .appendTextFormatted(
1191 "NOTE: %d %% of the run time was spent communicating energies,\n"
1192 " you might want to increase some nst* mdp options\n",
1193 gmx::roundToInt(100 * cyc_sum[ewcMoveE] / tot));
1197 extern int64_t wcycle_get_reset_counters(gmx_wallcycle_t wc)
1204 return wc->reset_counters;
1207 extern void wcycle_set_reset_counters(gmx_wallcycle_t wc, int64_t reset_counters)
1214 wc->reset_counters = reset_counters;
1217 void wallcycle_sub_start(gmx_wallcycle_t wc, int ewcs)
1219 if (useCycleSubcounters && wc != nullptr)
1221 wc->wcsc[ewcs].start = gmx_cycles_read();
1225 void wallcycle_sub_start_nocount(gmx_wallcycle_t wc, int ewcs)
1227 if (useCycleSubcounters && wc != nullptr)
1229 wallcycle_sub_start(wc, ewcs);
1234 void wallcycle_sub_stop(gmx_wallcycle_t wc, int ewcs)
1236 if (useCycleSubcounters && wc != nullptr)
1238 wc->wcsc[ewcs].c += gmx_cycles_read() - wc->wcsc[ewcs].start;