Separate CPU NB kernel and buffer clearing subcounters

[alexxy/gromacs.git] / src / gromacs / timing / wallcycle.cpp

/*
 * This file is part of the GROMACS molecular simulation package.
 *
 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
 * Copyright (c) 2001-2008, The GROMACS development team.
 * Copyright (c) 2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
 * and including many others, as listed in the AUTHORS file in the
 * top-level source directory and at http://www.gromacs.org.
 *
 * GROMACS is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public License
 * as published by the Free Software Foundation; either version 2.1
 * of the License, or (at your option) any later version.
 *
 * GROMACS is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with GROMACS; if not, see
 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
 *
 * If you want to redistribute modifications to GROMACS, please
 * consider that scientific software is very special. Version
 * control is crucial - bugs must be traceable. We will be happy to
 * consider code for inclusion in the official distribution, but
 * derived work must not be called official GROMACS. Details are found
 * in the README & COPYING files - if they are missing, get the
 * official version at http://www.gromacs.org.
 *
 * To help us fund GROMACS development, we humbly ask that you cite
 * the research papers on the package. Check out http://www.gromacs.org.
 */
#include "gmxpre.h"

#include "wallcycle.h"

#include "config.h"

#include <cstdlib>

#include <array>
#include <vector>

#include "gromacs/math/functions.h"
#include "gromacs/mdtypes/commrec.h"
#include "gromacs/timing/cyclecounter.h"
#include "gromacs/timing/gpu_timing.h"
#include "gromacs/timing/wallcyclereporting.h"
#include "gromacs/utility/cstringutil.h"
#include "gromacs/utility/gmxassert.h"
#include "gromacs/utility/gmxmpi.h"
#include "gromacs/utility/logger.h"
#include "gromacs/utility/smalloc.h"
#include "gromacs/utility/snprintf.h"

static const bool useCycleSubcounters = GMX_CYCLE_SUBCOUNTERS;

/* DEBUG_WCYCLE adds consistency checking for the counters.
 * It checks if you stop a counter different from the last
 * one that was opened and if you do nest too deep.
 */
/* #define DEBUG_WCYCLE */

#ifdef DEBUG_WCYCLE
#include "gromacs/utility/fatalerror.h"
#endif

typedef struct
{
    int          n;
    gmx_cycles_t c;
    gmx_cycles_t start;
} wallcc_t;

struct gmx_wallcycle
{
    wallcc_t        *wcc;
    /* did we detect one or more invalid cycle counts */
    gmx_bool         haveInvalidCount;
    /* variables for testing/debugging */
    gmx_bool         wc_barrier;
    wallcc_t        *wcc_all;
    int              wc_depth;
#ifdef DEBUG_WCYCLE
#define DEPTH_MAX 6
    int               counterlist[DEPTH_MAX];
    int               count_depth;
#endif
    int               ewc_prev;
    gmx_cycles_t      cycle_prev;
    int64_t           reset_counters;
#if GMX_MPI
    MPI_Comm          mpi_comm_mygroup;
#endif
    wallcc_t         *wcsc;
};

/* Each name should not exceed 19 printing characters
   (ie. terminating null can be twentieth) */
static const char *wcn[ewcNR] =
{
    "Run", "Step", "PP during PME", "Domain decomp.", "DD comm. load",
    "DD comm. bounds", "Vsite constr.", "Send X to PME", "Neighbor search", "Launch GPU ops.",
    "Comm. coord.", "Force", "Wait + Comm. F", "PME mesh",
    "PME redist. X/F", "PME spread", "PME gather", "PME 3D-FFT", "PME 3D-FFT Comm.", "PME solve LJ", "PME solve Elec",
    "PME wait for PP", "Wait + Recv. PME F",
    "Wait PME GPU spread", "PME 3D-FFT", "PME solve", /* the strings for FFT/solve are repeated here for mixed mode counters */
    "Wait PME GPU gather", "Wait Bonded GPU", "Reduce GPU PME F",
    "Wait GPU NB nonloc.", "Wait GPU NB local", "NB X/F buffer ops.",
    "Vsite spread", "COM pull force", "AWH",
    "Write traj.", "Update", "Constraints", "Comm. energies",
    "Enforced rotation", "Add rot. forces", "Position swapping", "IMD", "Test"
};

static const char *wcsn[ewcsNR] =
{
    "DD redist.", "DD NS grid + sort", "DD setup comm.",
    "DD make top.", "DD make constr.", "DD top. other",
    "NS grid local", "NS grid non-loc.", "NS search local", "NS search non-loc.",
    "Bonded F",
    "Bonded-FEP F",
    "Restraints F",
    "Listed buffer ops.",
    "Nonbonded pruning",
    "Nonbonded F",
    "NB F kernel", "NB F clear buf",
    "Launch NB GPU tasks",
    "Launch Bonded GPU tasks",
    "Launch PME GPU tasks",
    "Ewald F correction",
    "NB X buffer ops.",
    "NB F buffer ops.",
    "Clear force buffer",
    "Test subcounter",
};

/* PME GPU timing events' names - correspond to the enum in the gpu_timing.h */
static const char *PMEStageNames[] =
{
    "PME spline",
    "PME spread",
    "PME spline + spread",
    "PME 3D-FFT r2c",
    "PME solve",
    "PME 3D-FFT c2r",
    "PME gather",
};

gmx_bool wallcycle_have_counter()
{
    return gmx_cycles_have_counter();
}

gmx_wallcycle_t wallcycle_init(FILE *fplog, int resetstep, t_commrec gmx_unused *cr)
{
    gmx_wallcycle_t wc;


    if (!wallcycle_have_counter())
    {
        return nullptr;
    }

    snew(wc, 1);

    wc->haveInvalidCount    = FALSE;
    wc->wc_barrier          = FALSE;
    wc->wcc_all             = nullptr;
    wc->wc_depth            = 0;
    wc->ewc_prev            = -1;
    wc->reset_counters      = resetstep;

#if GMX_MPI
    if (PAR(cr) && getenv("GMX_CYCLE_BARRIER") != nullptr)
    {
        if (fplog)
        {
            fprintf(fplog, "\nWill call MPI_Barrier before each cycle start/stop call\n\n");
        }
        wc->wc_barrier       = TRUE;
        wc->mpi_comm_mygroup = cr->mpi_comm_mygroup;
    }
#endif

    snew(wc->wcc, ewcNR);
    if (getenv("GMX_CYCLE_ALL") != nullptr)
    {
        if (fplog)
        {
            fprintf(fplog, "\nWill time all the code during the run\n\n");
        }
        snew(wc->wcc_all, ewcNR*ewcNR);
    }

    if (useCycleSubcounters)
    {
        snew(wc->wcsc, ewcsNR);
    }

#ifdef DEBUG_WCYCLE
    wc->count_depth = 0;
#endif

    return wc;
}

void wallcycle_destroy(gmx_wallcycle_t wc)
{
    if (wc == nullptr)
    {
        return;
    }

    if (wc->wcc != nullptr)
    {
        sfree(wc->wcc);
    }
    if (wc->wcc_all != nullptr)
    {
        sfree(wc->wcc_all);
    }
    if (wc->wcsc != nullptr)
    {
        sfree(wc->wcsc);
    }
    sfree(wc);
}

static void wallcycle_all_start(gmx_wallcycle_t wc, int ewc, gmx_cycles_t cycle)
{
    wc->ewc_prev   = ewc;
    wc->cycle_prev = cycle;
}

static void wallcycle_all_stop(gmx_wallcycle_t wc, int ewc, gmx_cycles_t cycle)
{
    wc->wcc_all[wc->ewc_prev*ewcNR+ewc].n += 1;
    wc->wcc_all[wc->ewc_prev*ewcNR+ewc].c += cycle - wc->cycle_prev;
}


#ifdef DEBUG_WCYCLE
static void debug_start_check(gmx_wallcycle_t wc, int ewc)
{
    /* fprintf(stderr,"wcycle_start depth %d, %s\n",wc->count_depth,wcn[ewc]); */

    if (wc->count_depth < 0 || wc->count_depth >= DEPTH_MAX)
    {
        gmx_fatal(FARGS, "wallcycle counter depth out of range: %d",
                  wc->count_depth);
    }
    wc->counterlist[wc->count_depth] = ewc;
    wc->count_depth++;
}

static void debug_stop_check(gmx_wallcycle_t wc, int ewc)
{
    wc->count_depth--;

    /* fprintf(stderr,"wcycle_stop depth %d, %s\n",wc->count_depth,wcn[ewc]); */

    if (wc->count_depth < 0)
    {
        gmx_fatal(FARGS, "wallcycle counter depth out of range when stopping %s: %d", wcn[ewc], wc->count_depth);
    }
    if (wc->counterlist[wc->count_depth] != ewc)
    {
        gmx_fatal(FARGS, "wallcycle mismatch at stop, start %s, stop %s",
                  wcn[wc->counterlist[wc->count_depth]], wcn[ewc]);
    }
}
#endif

void wallcycle_start(gmx_wallcycle_t wc, int ewc)
{
    gmx_cycles_t cycle;

    if (wc == nullptr)
    {
        return;
    }

#if GMX_MPI
    if (wc->wc_barrier)
    {
        MPI_Barrier(wc->mpi_comm_mygroup);
    }
#endif

#ifdef DEBUG_WCYCLE
    debug_start_check(wc, ewc);
#endif

    cycle              = gmx_cycles_read();
    wc->wcc[ewc].start = cycle;
    if (wc->wcc_all != nullptr)
    {
        wc->wc_depth++;
        if (ewc == ewcRUN)
        {
            wallcycle_all_start(wc, ewc, cycle);
        }
        else if (wc->wc_depth == 3)
        {
            wallcycle_all_stop(wc, ewc, cycle);
        }
    }
}

void wallcycle_start_nocount(gmx_wallcycle_t wc, int ewc)
{
    if (wc == nullptr)
    {
        return;
    }

    wallcycle_start(wc, ewc);
    wc->wcc[ewc].n--;
}

double wallcycle_stop(gmx_wallcycle_t wc, int ewc)
{
    gmx_cycles_t cycle, last;

    if (wc == nullptr)
    {
        return 0;
    }

#if GMX_MPI
    if (wc->wc_barrier)
    {
        MPI_Barrier(wc->mpi_comm_mygroup);
    }
#endif

#ifdef DEBUG_WCYCLE
    debug_stop_check(wc, ewc);
#endif

    /* When processes or threads migrate between cores, the cycle counting
     * can get messed up if the cycle counter on different cores are not
     * synchronized. When this happens we expect both large negative and
     * positive cycle differences. We can detect negative cycle differences.
     * Detecting too large positive counts if difficult, since count can be
     * large, especially for ewcRUN. If we detect a negative count,
     * we will not print the cycle accounting table.
     */
    cycle                    = gmx_cycles_read();
    if (cycle >= wc->wcc[ewc].start)
    {
        last                 = cycle - wc->wcc[ewc].start;
    }
    else
    {
        last                 = 0;
        wc->haveInvalidCount = TRUE;
    }
    wc->wcc[ewc].c          += last;
    wc->wcc[ewc].n++;
    if (wc->wcc_all)
    {
        wc->wc_depth--;
        if (ewc == ewcRUN)
        {
            wallcycle_all_stop(wc, ewc, cycle);
        }
        else if (wc->wc_depth == 2)
        {
            wallcycle_all_start(wc, ewc, cycle);
        }
    }

    return last;
}

void wallcycle_get(gmx_wallcycle_t wc, int ewc, int *n, double *c)
{
    *n = wc->wcc[ewc].n;
    *c = static_cast<double>(wc->wcc[ewc].c);
}

void wallcycle_reset_all(gmx_wallcycle_t wc)
{
    int i;

    if (wc == nullptr)
    {
        return;
    }

    for (i = 0; i < ewcNR; i++)
    {
        wc->wcc[i].n = 0;
        wc->wcc[i].c = 0;
    }
    wc->haveInvalidCount = FALSE;

    if (wc->wcc_all)
    {
        for (i = 0; i < ewcNR*ewcNR; i++)
        {
            wc->wcc_all[i].n = 0;
            wc->wcc_all[i].c = 0;
        }
    }
    if (wc->wcsc)
    {
        for (i = 0; i < ewcsNR; i++)
        {
            wc->wcsc[i].n = 0;
            wc->wcsc[i].c = 0;
        }
    }
}

static gmx_bool is_pme_counter(int ewc)
{
    return (ewc >= ewcPMEMESH && ewc <= ewcPMEWAITCOMM);
}

static gmx_bool is_pme_subcounter(int ewc)
{
    return (ewc >= ewcPME_REDISTXF && ewc < ewcPMEWAITCOMM);
}

/* Subtract counter ewc_sub timed inside a timing block for ewc_main */
static void subtract_cycles(wallcc_t *wcc, int ewc_main, int ewc_sub)
{
    if (wcc[ewc_sub].n > 0)
    {
        if (wcc[ewc_main].c >= wcc[ewc_sub].c)
        {
            wcc[ewc_main].c -= wcc[ewc_sub].c;
        }
        else
        {
            /* Something is wrong with the cycle counting */
            wcc[ewc_main].c  = 0;
        }
    }
}

void wallcycle_scale_by_num_threads(gmx_wallcycle_t wc, bool isPmeRank, int nthreads_pp, int nthreads_pme)
{
    if (wc == nullptr)
    {
        return;
    }

    for (int i = 0; i < ewcNR; i++)
    {
        if (is_pme_counter(i) || (i == ewcRUN && isPmeRank))
        {
            wc->wcc[i].c *= nthreads_pme;

            if (wc->wcc_all)
            {
                for (int j = 0; j < ewcNR; j++)
                {
                    wc->wcc_all[i*ewcNR+j].c *= nthreads_pme;
                }
            }
        }
        else
        {
            wc->wcc[i].c *= nthreads_pp;

            if (wc->wcc_all)
            {
                for (int j = 0; j < ewcNR; j++)
                {
                    wc->wcc_all[i*ewcNR+j].c *= nthreads_pp;
                }
            }
        }
    }
    if (useCycleSubcounters && wc->wcsc && !isPmeRank)
    {
        for (int i = 0; i < ewcsNR; i++)
        {
            wc->wcsc[i].c *= nthreads_pp;
        }
    }
}

/* TODO Make an object for this function to return, containing some
 * vectors of something like wallcc_t for the summed wcc, wcc_all and
 * wcsc, AND the original wcc for rank 0.
 *
 * The GPU timing is reported only for rank 0, so we want to preserve
 * the original wcycle on that rank. Rank 0 also reports the global
 * counts before that, so needs something to contain the global data
 * without over-writing the rank-0 data. The current implementation
 * uses cycles_sum to manage this, which works OK now because wcsc and
 * wcc_all are unused by the GPU reporting, but it is not satisfactory
 * for the future. Also, there's no need for MPI_Allreduce, since
 * only MASTERRANK uses any of the results. */
WallcycleCounts wallcycle_sum(const t_commrec *cr, gmx_wallcycle_t wc)
{
    WallcycleCounts cycles_sum;
    wallcc_t       *wcc;
    double          cycles[ewcNR+ewcsNR];
#if GMX_MPI
    double          cycles_n[ewcNR+ewcsNR+1];
#endif
    int             i;
    int             nsum;

    if (wc == nullptr)
    {
        /* Default construction of std::array of non-class T can leave
           the values indeterminate, just like a C array, and icc
           warns about it. */
        cycles_sum.fill(0);
        return cycles_sum;
    }

    wcc = wc->wcc;

    subtract_cycles(wcc, ewcDOMDEC, ewcDDCOMMLOAD);
    subtract_cycles(wcc, ewcDOMDEC, ewcDDCOMMBOUND);

    subtract_cycles(wcc, ewcPME_FFT, ewcPME_FFTCOMM);

    if (cr->npmenodes == 0)
    {
        /* All nodes do PME (or no PME at all) */
        subtract_cycles(wcc, ewcFORCE, ewcPMEMESH);
    }
    else
    {
        /* The are PME-only nodes */
        if (wcc[ewcPMEMESH].n > 0)
        {
            /* This must be a PME only node, calculate the Wait + Comm. time */
            GMX_ASSERT(wcc[ewcRUN].c >= wcc[ewcPMEMESH].c, "Total run ticks must be greater than PME-only ticks");
            wcc[ewcPMEWAITCOMM].c = wcc[ewcRUN].c - wcc[ewcPMEMESH].c;
        }
    }

    /* Store the cycles in a double buffer for summing */
    for (i = 0; i < ewcNR; i++)
    {
#if GMX_MPI
        cycles_n[i] = static_cast<double>(wcc[i].n);
#endif
        cycles[i]   = static_cast<double>(wcc[i].c);
    }
    nsum = ewcNR;
    if (wc->wcsc)
    {
        for (i = 0; i < ewcsNR; i++)
        {
#if GMX_MPI
            cycles_n[ewcNR+i] = static_cast<double>(wc->wcsc[i].n);
#endif
            cycles[ewcNR+i]   = static_cast<double>(wc->wcsc[i].c);
        }
        nsum += ewcsNR;
    }

#if GMX_MPI
    if (cr->nnodes > 1)
    {
        double buf[ewcNR+ewcsNR+1];

        // TODO this code is used only at the end of the run, so we
        // can just do a simple reduce of haveInvalidCount in
        // wallcycle_print, and avoid bugs
        cycles_n[nsum] = (wc->haveInvalidCount ? 1 : 0);
        // TODO Use MPI_Reduce
        MPI_Allreduce(cycles_n, buf, nsum + 1, MPI_DOUBLE, MPI_MAX,
                      cr->mpi_comm_mysim);
        for (i = 0; i < ewcNR; i++)
        {
            wcc[i].n = gmx::roundToInt(buf[i]);
        }
        wc->haveInvalidCount = (buf[nsum] > 0);
        if (wc->wcsc)
        {
            for (i = 0; i < ewcsNR; i++)
            {
                wc->wcsc[i].n = gmx::roundToInt(buf[ewcNR+i]);
            }
        }

        // TODO Use MPI_Reduce
        MPI_Allreduce(cycles, cycles_sum.data(), nsum, MPI_DOUBLE, MPI_SUM,
                      cr->mpi_comm_mysim);

        if (wc->wcc_all != nullptr)
        {
            double *buf_all, *cyc_all;

            snew(cyc_all, ewcNR*ewcNR);
            snew(buf_all, ewcNR*ewcNR);
            for (i = 0; i < ewcNR*ewcNR; i++)
            {
                cyc_all[i] = wc->wcc_all[i].c;
            }
            // TODO Use MPI_Reduce
            MPI_Allreduce(cyc_all, buf_all, ewcNR*ewcNR, MPI_DOUBLE, MPI_SUM,
                          cr->mpi_comm_mysim);
            for (i = 0; i < ewcNR*ewcNR; i++)
            {
                wc->wcc_all[i].c = static_cast<gmx_cycles_t>(buf_all[i]);
            }
            sfree(buf_all);
            sfree(cyc_all);
        }
    }
    else
#endif
    {
        for (i = 0; i < nsum; i++)
        {
            cycles_sum[i] = cycles[i];
        }
    }

    return cycles_sum;
}

static void print_cycles(FILE *fplog, double c2t, const char *name,
                         int nnodes, int nthreads,
                         int ncalls, double c_sum, double tot)
{
    char   nnodes_str[STRLEN];
    char   nthreads_str[STRLEN];
    char   ncalls_str[STRLEN];
    double wallt;
    double percentage = (tot > 0.) ? (100. * c_sum / tot) : 0.;

    if (c_sum > 0)
    {
        if (ncalls > 0)
        {
            snprintf(ncalls_str, sizeof(ncalls_str), "%10d", ncalls);
            if (nnodes < 0)
            {
                snprintf(nnodes_str, sizeof(nnodes_str), "N/A");
            }
            else
            {
                snprintf(nnodes_str, sizeof(nnodes_str), "%4d", nnodes);
            }
            if (nthreads < 0)
            {
                snprintf(nthreads_str, sizeof(nthreads_str), "N/A");
            }
            else
            {
                snprintf(nthreads_str, sizeof(nthreads_str), "%4d", nthreads);
            }
        }
        else
        {
            nnodes_str[0]   = 0;
            nthreads_str[0] = 0;
            ncalls_str[0]   = 0;
        }
        /* Convert the cycle count to wallclock time for this task */
        wallt = c_sum*c2t;

        fprintf(fplog, " %-19.19s %4s %4s %10s  %10.3f %14.3f %5.1f\n",
                name, nnodes_str, nthreads_str, ncalls_str, wallt,
                c_sum*1e-9, percentage);
    }
}

static void print_gputimes(FILE *fplog, const char *name,
                           int n, double t, double tot_t)
{
    char num[11];
    char avg_perf[11];

    if (n > 0)
    {
        snprintf(num, sizeof(num), "%10d", n);
        snprintf(avg_perf, sizeof(avg_perf), "%10.3f", t/n);
    }
    else
    {
        sprintf(num, "          ");
        sprintf(avg_perf, "          ");
    }
    if (t != tot_t && tot_t > 0)
    {
        fprintf(fplog, " %-29s %10s%12.3f   %s   %5.1f\n",
                name, num, t/1000, avg_perf, 100 * t/tot_t);
    }
    else
    {
        fprintf(fplog, " %-29s %10s%12.3f   %s   %5.1f\n",
                name, "", t/1000, avg_perf, 100.0);
    }
}

static void print_header(FILE *fplog, int nrank_pp, int nth_pp, int nrank_pme, int nth_pme)
{
    int nrank_tot = nrank_pp + nrank_pme;
    if (0 == nrank_pme)
    {
        fprintf(fplog, "On %d MPI rank%s", nrank_tot, nrank_tot == 1 ? "" : "s");
        if (nth_pp > 1)
        {
            fprintf(fplog, ", each using %d OpenMP threads", nth_pp);
        }
        /* Don't report doing PP+PME, because we can't tell here if
         * this is RF, etc. */
    }
    else
    {
        fprintf(fplog, "On %d MPI rank%s doing PP", nrank_pp, nrank_pp == 1 ? "" : "s");
        if (nth_pp > 1)
        {
            fprintf(fplog, ",%s using %d OpenMP threads", nrank_pp > 1 ? " each" : "", nth_pp);
        }
        fprintf(fplog, ", and\non %d MPI rank%s doing PME", nrank_pme, nrank_pme == 1 ? "" : "s");
        if (nth_pme > 1)
        {
            fprintf(fplog, ",%s using %d OpenMP threads", nrank_pme > 1 ? " each" : "", nth_pme);
        }
    }

    fprintf(fplog, "\n\n");
    fprintf(fplog, " Computing:          Num   Num      Call    Wall time         Giga-Cycles\n");
    fprintf(fplog, "                     Ranks Threads  Count      (s)         total sum    %%\n");
}


void wallcycle_print(FILE *fplog, const gmx::MDLogger &mdlog, int nnodes, int npme,
                     int nth_pp, int nth_pme, double realtime,
                     gmx_wallcycle_t wc, const WallcycleCounts &cyc_sum,
                     const gmx_wallclock_gpu_nbnxn_t *gpu_nbnxn_t,
                     const gmx_wallclock_gpu_pme_t *gpu_pme_t)
{
    double      tot, tot_for_pp, tot_for_rest, tot_cpu_overlap, gpu_cpu_ratio;
    double      c2t, c2t_pp, c2t_pme = 0;
    int         i, j, npp, nth_tot;
    char        buf[STRLEN];
    const char *hline = "-----------------------------------------------------------------------------";

    if (wc == nullptr)
    {
        return;
    }

    GMX_ASSERT(nth_pp > 0, "Number of particle-particle threads must be >0");
    GMX_ASSERT(nth_pme > 0, "Number of PME threads must be >0");
    GMX_ASSERT(nnodes > 0, "Number of nodes must be >0");
    GMX_ASSERT(npme >= 0, "Number of PME nodes cannot be negative");
    npp     = nnodes - npme;
    /* npme is the number of PME-only ranks used, and we always do PP work */
    GMX_ASSERT(npp > 0, "Number of particle-particle nodes must be >0");

    nth_tot = npp*nth_pp + npme*nth_pme;

    /* When using PME-only nodes, the next line is valid for both
       PP-only and PME-only nodes because they started ewcRUN at the
       same time. */
    tot        = cyc_sum[ewcRUN];
    tot_for_pp = 0;

    if (tot <= 0.0)
    {
        /* TODO This is heavy handed, but until someone reworks the
           code so that it is provably robust with respect to
           non-positive values for all possible timer and cycle
           counters, there is less value gained from printing whatever
           timing data might still be sensible for some non-Jenkins
           run, than is lost from diagnosing Jenkins FP exceptions on
           runs about whose execution time we don't care. */
        GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
                "WARNING: A total of %f CPU cycles was recorded, so mdrun cannot print a time accounting",
                tot);
        return;
    }

    if (wc->haveInvalidCount)
    {
        GMX_LOG(mdlog.warning).asParagraph().appendText("NOTE: Detected invalid cycle counts, probably because threads moved between CPU cores that do not have synchronized cycle counters. Will not print the cycle accounting.");
        return;
    }


    /* Conversion factor from cycles to seconds */
    c2t     = realtime/tot;
    c2t_pp  = c2t * nth_tot / static_cast<double>(npp*nth_pp);
    if (npme > 0)
    {
        c2t_pme = c2t * nth_tot / static_cast<double>(npme*nth_pme);
    }
    else
    {
        c2t_pme = 0;
    }

    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");

    print_header(fplog, npp, nth_pp, npme, nth_pme);

    fprintf(fplog, "%s\n", hline);
    for (i = ewcPPDURINGPME+1; i < ewcNR; i++)
    {
        if (is_pme_subcounter(i))
        {
            /* Do not count these at all */
        }
        else if (npme > 0 && is_pme_counter(i))
        {
            /* Print timing information for PME-only nodes, but add an
             * asterisk so the reader of the table can know that the
             * walltimes are not meant to add up. The asterisk still
             * fits in the required maximum of 19 characters. */
            char buffer[STRLEN];
            snprintf(buffer, STRLEN, "%s *", wcn[i]);
            print_cycles(fplog, c2t_pme, buffer,
                         npme, nth_pme,
                         wc->wcc[i].n, cyc_sum[i], tot);
        }
        else
        {
            /* Print timing information when it is for a PP or PP+PME
               node */
            print_cycles(fplog, c2t_pp, wcn[i],
                         npp, nth_pp,
                         wc->wcc[i].n, cyc_sum[i], tot);
            tot_for_pp += cyc_sum[i];
        }
    }
    if (wc->wcc_all != nullptr)
    {
        for (i = 0; i < ewcNR; i++)
        {
            for (j = 0; j < ewcNR; j++)
            {
                snprintf(buf, 20, "%-9.9s %-9.9s", wcn[i], wcn[j]);
                print_cycles(fplog, c2t_pp, buf,
                             npp, nth_pp,
                             wc->wcc_all[i*ewcNR+j].n,
                             wc->wcc_all[i*ewcNR+j].c,
                             tot);
            }
        }
    }
    tot_for_rest = tot * npp * nth_pp / static_cast<double>(nth_tot);
    print_cycles(fplog, c2t_pp, "Rest",
                 npp, nth_pp,
                 -1, tot_for_rest - tot_for_pp, tot);
    fprintf(fplog, "%s\n", hline);
    print_cycles(fplog, c2t, "Total",
                 npp, nth_pp,
                 -1, tot, tot);
    fprintf(fplog, "%s\n", hline);

    if (npme > 0)
    {
        fprintf(fplog,
                "(*) Note that with separate PME ranks, the walltime column actually sums to\n"
                "    twice the total reported, but the cycle count total and %% are correct.\n"
                "%s\n", hline);
    }

    if (wc->wcc[ewcPMEMESH].n > 0)
    {
        // A workaround to not print breakdown when no subcounters were recorded.
        // TODO: figure out and record PME GPU counters (what to do with the waiting ones?)
        std::vector<int> validPmeSubcounterIndices;
        for (i = ewcPPDURINGPME+1; i < ewcNR; i++)
        {
            if (is_pme_subcounter(i) && wc->wcc[i].n > 0)
            {
                validPmeSubcounterIndices.push_back(i);
            }
        }

        if (!validPmeSubcounterIndices.empty())
        {
            fprintf(fplog, " Breakdown of PME mesh computation\n");
            fprintf(fplog, "%s\n", hline);
            for (auto i : validPmeSubcounterIndices)
            {
                print_cycles(fplog, npme > 0 ? c2t_pme : c2t_pp, wcn[i],
                             npme > 0 ? npme : npp, nth_pme,
                             wc->wcc[i].n, cyc_sum[i], tot);
            }
            fprintf(fplog, "%s\n", hline);
        }
    }

    if (useCycleSubcounters && wc->wcsc)
    {
        fprintf(fplog, " Breakdown of PP computation\n");
        fprintf(fplog, "%s\n", hline);
        for (i = 0; i < ewcsNR; i++)
        {
            print_cycles(fplog, c2t_pp, wcsn[i],
                         npp, nth_pp,
                         wc->wcsc[i].n, cyc_sum[ewcNR+i], tot);
        }
        fprintf(fplog, "%s\n", hline);
    }

    /* print GPU timing summary */
    double tot_gpu = 0.0;
    if (gpu_pme_t)
    {
        for (size_t k = 0; k < gtPME_EVENT_COUNT; k++)
        {
            tot_gpu += gpu_pme_t->timing[k].t;
        }
    }
    if (gpu_nbnxn_t)
    {
        const char *k_log_str[2][2] = {
            {"Nonbonded F kernel", "Nonbonded F+ene k."},
            {"Nonbonded F+prune k.", "Nonbonded F+ene+prune k."}
        };
        tot_gpu += gpu_nbnxn_t->pl_h2d_t + gpu_nbnxn_t->nb_h2d_t + gpu_nbnxn_t->nb_d2h_t;

        /* add up the kernel timings */
        for (i = 0; i < 2; i++)
        {
            for (j = 0; j < 2; j++)
            {
                tot_gpu += gpu_nbnxn_t->ktime[i][j].t;
            }
        }
        tot_gpu += gpu_nbnxn_t->pruneTime.t;

        tot_cpu_overlap = wc->wcc[ewcFORCE].c;
        if (wc->wcc[ewcPMEMESH].n > 0)
        {
            tot_cpu_overlap += wc->wcc[ewcPMEMESH].c;
        }
        tot_cpu_overlap *= realtime*1000/tot; /* convert s to ms */

        fprintf(fplog, "\n GPU timings\n%s\n", hline);
        fprintf(fplog, " Computing:                         Count  Wall t (s)      ms/step       %c\n", '%');
        fprintf(fplog, "%s\n", hline);
        print_gputimes(fplog, "Pair list H2D",
                       gpu_nbnxn_t->pl_h2d_c, gpu_nbnxn_t->pl_h2d_t, tot_gpu);
        print_gputimes(fplog, "X / q H2D",
                       gpu_nbnxn_t->nb_c, gpu_nbnxn_t->nb_h2d_t, tot_gpu);

        for (i = 0; i < 2; i++)
        {
            for (j = 0; j < 2; j++)
            {
                if (gpu_nbnxn_t->ktime[i][j].c)
                {
                    print_gputimes(fplog, k_log_str[i][j],
                                   gpu_nbnxn_t->ktime[i][j].c, gpu_nbnxn_t->ktime[i][j].t, tot_gpu);
                }
            }
        }
        if (gpu_pme_t)
        {
            for (size_t k = 0; k < gtPME_EVENT_COUNT; k++)
            {
                if (gpu_pme_t->timing[k].c)
                {
                    print_gputimes(fplog, PMEStageNames[k],
                                   gpu_pme_t->timing[k].c,
                                   gpu_pme_t->timing[k].t,
                                   tot_gpu);
                }
            }
        }
        if (gpu_nbnxn_t->pruneTime.c)
        {
            print_gputimes(fplog, "Pruning kernel", gpu_nbnxn_t->pruneTime.c, gpu_nbnxn_t->pruneTime.t, tot_gpu);
        }
        print_gputimes(fplog, "F D2H",  gpu_nbnxn_t->nb_c, gpu_nbnxn_t->nb_d2h_t, tot_gpu);
        fprintf(fplog, "%s\n", hline);
        print_gputimes(fplog, "Total ", gpu_nbnxn_t->nb_c, tot_gpu, tot_gpu);
        fprintf(fplog, "%s\n", hline);
        if (gpu_nbnxn_t->dynamicPruneTime.c)
        {
            /* We print the dynamic pruning kernel timings after a separator
             * and avoid adding it to tot_gpu as this is not in the force
             * overlap. We print the fraction as relative to the rest.
             */
            print_gputimes(fplog, "*Dynamic pruning", gpu_nbnxn_t->dynamicPruneTime.c, gpu_nbnxn_t->dynamicPruneTime.t, tot_gpu);
            fprintf(fplog, "%s\n", hline);
        }
        gpu_cpu_ratio = tot_gpu/tot_cpu_overlap;
        if (gpu_nbnxn_t->nb_c > 0 && wc->wcc[ewcFORCE].n > 0)
        {
            fprintf(fplog, "\nAverage per-step force GPU/CPU evaluation time ratio: %.3f ms/%.3f ms = %.3f\n",
                    tot_gpu/gpu_nbnxn_t->nb_c, tot_cpu_overlap/wc->wcc[ewcFORCE].n,
                    gpu_cpu_ratio);
        }

        /* only print notes related to CPU-GPU load balance with PME */
        if (wc->wcc[ewcPMEMESH].n > 0)
        {
            fprintf(fplog, "For optimal resource utilization this ratio should be close to 1\n");

            /* print note if the imbalance is high with PME case in which
             * CPU-GPU load balancing is possible */
            if (gpu_cpu_ratio < 0.8 || gpu_cpu_ratio > 1.25)
            {
                /* Only the sim master calls this function, so always print to stderr */
                if (gpu_cpu_ratio < 0.8)
                {
                    if (npp > 1)
                    {
                        /* The user could have used -notunepme,
                         * but we currently can't check that here.
                         */
                        GMX_LOG(mdlog.warning).asParagraph().appendText(
                                "NOTE: The CPU has >25% more load than the GPU. This imbalance wastes\n"
                                "      GPU resources. Maybe the domain decomposition limits the PME tuning.\n"
                                "      In that case, try setting the DD grid manually (-dd) or lowering -dds.");
                    }
                    else
                    {
                        /* We should not end up here, unless the box is
                         * too small for increasing the cut-off for PME tuning.
                         */
                        GMX_LOG(mdlog.warning).asParagraph().appendText(
                                "NOTE: The CPU has >25% more load than the GPU. This imbalance wastes\n"
                                "      GPU resources.");
                    }
                }
                if (gpu_cpu_ratio > 1.25)
                {
                    GMX_LOG(mdlog.warning).asParagraph().appendText(
                            "NOTE: The GPU has >25% more load than the CPU. This imbalance wastes\n"
                            "      CPU resources.");
                }
            }
        }
    }

    if (wc->wc_barrier)
    {
        GMX_LOG(mdlog.warning).asParagraph().appendText(
                "MPI_Barrier was called before each cycle start/stop\n"
                "call, so timings are not those of real runs.");
    }

    if (wc->wcc[ewcNB_XF_BUF_OPS].n > 0 &&
        (cyc_sum[ewcDOMDEC] > tot*0.1 ||
         cyc_sum[ewcNS] > tot*0.1))
    {
        /* Only the sim master calls this function, so always print to stderr */
        if (wc->wcc[ewcDOMDEC].n == 0)
        {
            GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
                    "NOTE: %d %% of the run time was spent in pair search,\n"
                    "      you might want to increase nstlist (this has no effect on accuracy)\n",
                    gmx::roundToInt(100*cyc_sum[ewcNS]/tot));
        }
        else
        {
            GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
                    "NOTE: %d %% of the run time was spent in domain decomposition,\n"
                    "      %d %% of the run time was spent in pair search,\n"
                    "      you might want to increase nstlist (this has no effect on accuracy)\n",
                    gmx::roundToInt(100*cyc_sum[ewcDOMDEC]/tot),
                    gmx::roundToInt(100*cyc_sum[ewcNS]/tot));
        }
    }

    if (cyc_sum[ewcMoveE] > tot*0.05)
    {
        GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
                "NOTE: %d %% of the run time was spent communicating energies,\n"
                "      you might want to increase some nst* mdp options\n",
                gmx::roundToInt(100*cyc_sum[ewcMoveE]/tot));
    }
}

extern int64_t wcycle_get_reset_counters(gmx_wallcycle_t wc)
{
    if (wc == nullptr)
    {
        return -1;
    }

    return wc->reset_counters;
}

extern void wcycle_set_reset_counters(gmx_wallcycle_t wc, int64_t reset_counters)
{
    if (wc == nullptr)
    {
        return;
    }

    wc->reset_counters = reset_counters;
}

void wallcycle_sub_start(gmx_wallcycle_t wc, int ewcs)
{
    if (useCycleSubcounters && wc != nullptr)
    {
        wc->wcsc[ewcs].start = gmx_cycles_read();
    }
}

void wallcycle_sub_start_nocount(gmx_wallcycle_t wc, int ewcs)
{
    if (useCycleSubcounters && wc != nullptr)
    {
        wallcycle_sub_start(wc, ewcs);
        wc->wcsc[ewcs].n--;
    }
}

void wallcycle_sub_stop(gmx_wallcycle_t wc, int ewcs)
{
    if (useCycleSubcounters && wc != nullptr)
    {
        wc->wcsc[ewcs].c += gmx_cycles_read() - wc->wcsc[ewcs].start;
        wc->wcsc[ewcs].n++;
    }
}