Prepared t_mdatoms for using vector

[gromacs.git] / src / gromacs / mdlib / gmx_omp_nthreads.cpp

/*
 * This file is part of the GROMACS molecular simulation package.
 *
 * Copyright (c) 2012,2013,2014,2015,2016,2017, 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 "gmx_omp_nthreads.h"

#include "config.h"

#include <cstdio>
#include <cstdlib>
#include <cstring>

#include "gromacs/gmxlib/network.h"
#include "gromacs/mdtypes/commrec.h"
#include "gromacs/utility/cstringutil.h"
#include "gromacs/utility/fatalerror.h"
#include "gromacs/utility/gmxassert.h"
#include "gromacs/utility/gmxomp.h"
#include "gromacs/utility/logger.h"
#include "gromacs/utility/programcontext.h"

/** Structure with the number of threads for each OpenMP multi-threaded
 *  algorithmic module in mdrun. */
typedef struct
{
    int      gnth;          /**< Global num. of threads per PP or PP+PME process/tMPI thread. */
    int      gnth_pme;      /**< Global num. of threads per PME only process/tMPI thread. */

    int      nth[emntNR];   /**< Number of threads for each module, indexed with module_nth_t */
    gmx_bool initialized;   /**< TRUE if the module as been initialized. */
} omp_module_nthreads_t;

/** Names of environment variables to set the per module number of threads.
 *
 *  Indexed with the values of module_nth_t.
 * */
static const char *modth_env_var[emntNR] =
{
    "GMX_DEFAULT_NUM_THREADS should never be set",
    "GMX_DOMDEC_NUM_THREADS", "GMX_PAIRSEARCH_NUM_THREADS",
    "GMX_NONBONDED_NUM_THREADS", "GMX_LISTED_FORCES_NUM_THREADS",
    "GMX_PME_NUM_THREADS", "GMX_UPDATE_NUM_THREADS",
    "GMX_VSITE_NUM_THREADS",
    "GMX_LINCS_NUM_THREADS", "GMX_SETTLE_NUM_THREADS"
};

/** Names of the modules. */
static const char *mod_name[emntNR] =
{
    "default", "domain decomposition", "pair search", "non-bonded",
    "bonded", "PME", "update", "LINCS", "SETTLE"
};

/** Number of threads for each algorithmic module.
 *
 *  File-scope global variable that gets set once in pick_module_nthreads()
 *  and queried via gmx_omp_nthreads_get().
 *
 *  All fields are initialized to 0 which should result in errors if
 *  the init call is omitted.
 * */
static omp_module_nthreads_t modth = { 0, 0, {0, 0, 0, 0, 0, 0, 0, 0, 0}, FALSE};


/** Determine the number of threads for module \p mod.
 *
 *  \p m takes values form the module_nth_t enum and maps these to the
 *  corresponding value in modth_env_var.
 *
 *  Each number of threads per module takes the default value unless
 *  GMX_*_NUM_THERADS env var is set, case in which its value overrides
 *  the deafult.
 *
 *  The "group" scheme supports OpenMP only in PME and in thise case all but
 *  the PME nthread values default to 1.
 */
static void pick_module_nthreads(const gmx::MDLogger &mdlog, int m,
                                 gmx_bool bFullOmpSupport,
                                 gmx_bool bSepPME)
{
    char      *env;
    int        nth;

    const bool bOMP = GMX_OPENMP;

    /* The default should never be set through a GMX_*_NUM_THREADS env var
     * as it's always equal with gnth. */
    if (m == emntDefault)
    {
        return;
    }

    /* check the environment variable */
    if ((env = getenv(modth_env_var[m])) != nullptr)
    {
        sscanf(env, "%d", &nth);

        // cppcheck-suppress knownConditionTrueFalse
        if (!bOMP)
        {
            gmx_warning("%s=%d is set, but %s is compiled without OpenMP!",
                        modth_env_var[m], nth,
                        gmx::getProgramContext().displayName());
        }

        /* with the verlet codepath, when any GMX_*_NUM_THREADS env var is set,
         * OMP_NUM_THREADS also has to be set */
        if (bFullOmpSupport && getenv("OMP_NUM_THREADS") == nullptr)
        {
            gmx_warning("%s=%d is set, the default number of threads also "
                        "needs to be set with OMP_NUM_THREADS!",
                        modth_env_var[m], nth);
        }

        /* with the group scheme warn if any env var except PME is set */
        if (!bFullOmpSupport)
        {
            if (m != emntPME)
            {
                gmx_warning("%s=%d is set, but OpenMP multithreading is not "
                            "supported in %s!",
                            modth_env_var[m], nth, mod_name[m]);
                nth = 1;
            }
        }

        /* only babble if we are really overriding with a different value */
        if ((bSepPME && m == emntPME && nth != modth.gnth_pme) || (nth != modth.gnth))
        {
            GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
                    "%s=%d set, overriding the default number of %s threads",
                    modth_env_var[m], nth, mod_name[m]);
        }
    }
    else
    {
        /* pick the global PME node nthreads if we are setting the number
         * of threads in separate PME nodes  */
        nth = (bSepPME && m == emntPME) ? modth.gnth_pme : modth.gnth;
    }

    gmx_omp_nthreads_set(m, nth);
}

void gmx_omp_nthreads_read_env(int     *nthreads_omp,
                               gmx_bool bIsSimMaster)
{
    char    *env;
    gmx_bool bCommandLineSetNthreadsOMP = *nthreads_omp > 0;
    char     buffer[STRLEN];

    GMX_RELEASE_ASSERT(nthreads_omp, "nthreads_omp must be a non-NULL pointer");

    if ((env = getenv("OMP_NUM_THREADS")) != nullptr)
    {
        int nt_omp;

        sscanf(env, "%d", &nt_omp);
        if (nt_omp <= 0)
        {
            gmx_fatal(FARGS, "OMP_NUM_THREADS is invalid: '%s'", env);
        }

        if (bCommandLineSetNthreadsOMP && nt_omp != *nthreads_omp)
        {
            gmx_fatal(FARGS, "Environment variable OMP_NUM_THREADS (%d) and the number of threads requested on the command line (%d) have different values. Either omit one, or set them both to the same value.", nt_omp, *nthreads_omp);
        }

        /* Setting the number of OpenMP threads. */
        *nthreads_omp = nt_omp;

        /* Output the results */
        sprintf(buffer,
                "The number of OpenMP threads was set by environment variable OMP_NUM_THREADS to %d%s\n",
                nt_omp,
                bCommandLineSetNthreadsOMP ? " (and the command-line setting agreed with that)" : "");
        if (bIsSimMaster)
        {
            /* This prints once per simulation for multi-simulations,
             * which might help diagnose issues with inhomogenous
             * cluster setups. */
            fputs(buffer, stderr);
        }
        if (debug)
        {
            /* This prints once per process for real MPI (i.e. once
             * per debug file), and once per simulation for thread MPI
             * (because of logic in the calling function). */
            fputs(buffer, debug);
        }
    }
}

/*! \brief Helper function for parsing various input about the number
    of OpenMP threads to use in various modules and deciding what to
    do about it. */
static void manage_number_of_openmp_threads(const gmx::MDLogger &mdlog,
                                            const t_commrec     *cr,
                                            bool                 bOMP,
                                            int                  nthreads_hw_avail,
                                            int                  omp_nthreads_req,
                                            int                  omp_nthreads_pme_req,
                                            gmx_bool gmx_unused  bThisNodePMEOnly,
                                            gmx_bool             bFullOmpSupport,
                                            int                  nppn,
                                            gmx_bool             bSepPME)
{
    int      nth;
    char    *env;

#if GMX_THREAD_MPI
    /* modth is shared among tMPI threads, so for thread safety, the
     * detection is done on the master only. It is not thread-safe
     * with multiple simulations, but that's anyway not supported by
     * tMPI. */
    if (!SIMMASTER(cr))
    {
        return;
    }
#else
    GMX_UNUSED_VALUE(cr);
#endif

    if (modth.initialized)
    {
        /* Just return if the initialization has already been
           done. This could only happen if gmx_omp_nthreads_init() has
           already been called. */
        return;
    }

    /* With full OpenMP support (verlet scheme) set the number of threads
     * per process / default:
     * - 1 if not compiled with OpenMP or
     * - OMP_NUM_THREADS if the env. var is set, or
     * - omp_nthreads_req = #of threads requested by the user on the mdrun
     *   command line, otherwise
     * - take the max number of available threads and distribute them
     *   on the processes/tMPI threads.
     * ~ The GMX_*_NUM_THREADS env var overrides the number of threads of
     *   the respective module and it has to be used in conjunction with
     *   OMP_NUM_THREADS.
     *
     * With the group scheme OpenMP multithreading is only supported in PME,
     * for all other modules nthreads is set to 1.
     * The number of PME threads is equal to:
     * - 1 if not compiled with OpenMP or
     * - GMX_PME_NUM_THREADS if defined, otherwise
     * - OMP_NUM_THREADS if defined, otherwise
     * - 1
     */
    nth = 1;
    if ((env = getenv("OMP_NUM_THREADS")) != nullptr)
    {
        if (!bOMP && (std::strncmp(env, "1", 1) != 0))
        {
            gmx_warning("OMP_NUM_THREADS is set, but %s was compiled without OpenMP support!",
                        gmx::getProgramContext().displayName());
        }
        else
        {
            nth = gmx_omp_get_max_threads();
        }
    }
    else if (omp_nthreads_req > 0)
    {
        nth = omp_nthreads_req;
    }
    else if (bFullOmpSupport && bOMP)
    {
        /* max available threads per node */
        nth = nthreads_hw_avail;

        /* divide the threads among the MPI processes/tMPI threads */
        if (nth >= nppn)
        {
            nth /= nppn;
        }
        else
        {
            nth = 1;
        }
    }

    /* now we have the global values, set them:
     * - 1 if not compiled with OpenMP and for the group scheme
     * - nth for the verlet scheme when compiled with OpenMP
     */
    if (bFullOmpSupport && bOMP)
    {
        modth.gnth = nth;
    }
    else
    {
        modth.gnth = 1;
    }

    if (bSepPME)
    {
        if (omp_nthreads_pme_req > 0)
        {
            modth.gnth_pme = omp_nthreads_pme_req;
        }
        else
        {
            modth.gnth_pme = nth;
        }
    }
    else
    {
        modth.gnth_pme = 0;
    }

    /* now set the per-module values */
    modth.nth[emntDefault] = modth.gnth;
    pick_module_nthreads(mdlog, emntDomdec, bFullOmpSupport, bSepPME);
    pick_module_nthreads(mdlog, emntPairsearch, bFullOmpSupport, bSepPME);
    pick_module_nthreads(mdlog, emntNonbonded, bFullOmpSupport, bSepPME);
    pick_module_nthreads(mdlog, emntBonded, bFullOmpSupport, bSepPME);
    pick_module_nthreads(mdlog, emntPME, bFullOmpSupport, bSepPME);
    pick_module_nthreads(mdlog, emntUpdate, bFullOmpSupport, bSepPME);
    pick_module_nthreads(mdlog, emntVSITE, bFullOmpSupport, bSepPME);
    pick_module_nthreads(mdlog, emntLINCS, bFullOmpSupport, bSepPME);
    pick_module_nthreads(mdlog, emntSETTLE, bFullOmpSupport, bSepPME);

    /* set the number of threads globally */
    if (bOMP)
    {
#if !GMX_THREAD_MPI
        if (bThisNodePMEOnly)
        {
            gmx_omp_set_num_threads(modth.gnth_pme);
        }
        else
#endif      /* GMX_THREAD_MPI */
        {
            if (bFullOmpSupport)
            {
                gmx_omp_set_num_threads(nth);
            }
            else
            {
                gmx_omp_set_num_threads(1);
            }
        }
    }

    modth.initialized = TRUE;
}

/*! \brief Report on the OpenMP settings that will be used */
static void
reportOpenmpSettings(const gmx::MDLogger &mdlog,
                     const t_commrec     *cr,
                     gmx_bool             bOMP,
                     gmx_bool             bFullOmpSupport,
                     gmx_bool             bSepPME)
{
#if GMX_THREAD_MPI
    const char *mpi_str = "per tMPI thread";
#else
    const char *mpi_str = "per MPI process";
#endif
    int         nth_min, nth_max, nth_pme_min, nth_pme_max;

    /* inform the user about the settings */
    if (!bOMP)
    {
        return;
    }

#if GMX_MPI
    if (cr->nnodes + cr->npmenodes > 1)
    {
        /* Get the min and max thread counts over the MPI ranks */
        int buf_in[4], buf_out[4];

        buf_in[0] = -modth.gnth;
        buf_in[1] =  modth.gnth;
        buf_in[2] = -modth.gnth_pme;
        buf_in[3] =  modth.gnth_pme;

        MPI_Allreduce(buf_in, buf_out, 4, MPI_INT, MPI_MAX, cr->mpi_comm_mysim);

        nth_min     = -buf_out[0];
        nth_max     =  buf_out[1];
        nth_pme_min = -buf_out[2];
        nth_pme_max =  buf_out[3];
    }
    else
#endif
    {
        nth_min     = modth.gnth;
        nth_max     = modth.gnth;
        nth_pme_min = modth.gnth_pme;
        nth_pme_max = modth.gnth_pme;
    }

    /* for group scheme we print PME threads info only */
    if (bFullOmpSupport)
    {
        if (nth_max == nth_min)
        {
            GMX_LOG(mdlog.warning).appendTextFormatted(
                    "Using %d OpenMP thread%s %s",
                    nth_min, nth_min > 1 ? "s" : "",
                    cr->nnodes > 1 ? mpi_str : "");
        }
        else
        {
            GMX_LOG(mdlog.warning).appendTextFormatted(
                    "Using %d - %d OpenMP threads %s",
                    nth_min, nth_max, mpi_str);
        }
    }
    if (bSepPME && (nth_pme_min != nth_min || nth_pme_max != nth_max))
    {
        if (nth_pme_max == nth_pme_min)
        {
            GMX_LOG(mdlog.warning).appendTextFormatted(
                    "Using %d OpenMP thread%s %s for PME",
                    nth_pme_min, nth_pme_min > 1 ? "s" : "",
                    cr->nnodes > 1 ? mpi_str : "");
        }
        else
        {
            GMX_LOG(mdlog.warning).appendTextFormatted(
                    "Using %d - %d OpenMP threads %s for PME",
                    nth_pme_min, nth_pme_max, mpi_str);
        }
    }
    GMX_LOG(mdlog.warning);
}

/*! \brief Detect and warn about oversubscription of cores.
 *
 * \todo This could probably live elsewhere, since it is not specifc
 * to OpenMP, and only needs modth.gnth.
 *
 * \todo Enable this for separate PME nodes as well! */
static void
issueOversubscriptionWarning(const gmx::MDLogger &mdlog,
                             const t_commrec     *cr,
                             int                  nthreads_hw_avail,
                             int                  nppn,
                             gmx_bool             bSepPME)
{
    char sbuf[STRLEN], sbuf1[STRLEN], sbuf2[STRLEN];

    if (bSepPME || 0 != cr->rank_pp_intranode)
    {
        return;
    }

    if (modth.gnth*nppn > nthreads_hw_avail)
    {
        sprintf(sbuf, "threads");
        sbuf1[0] = '\0';
        sprintf(sbuf2, "O");
#if GMX_MPI
        if (modth.gnth == 1)
        {
#if GMX_THREAD_MPI
            sprintf(sbuf, "thread-MPI threads");
#else
            sprintf(sbuf, "MPI processes");
            sprintf(sbuf1, " per rank");
            sprintf(sbuf2, "On rank %d: o", cr->sim_nodeid);
#endif
        }
#endif
        GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
                "WARNING: %sversubscribing the available %d logical CPU cores%s with %d %s.\n"
                "         This will cause considerable performance loss!",
                sbuf2, nthreads_hw_avail, sbuf1, nppn*modth.gnth, sbuf);
    }
}

void gmx_omp_nthreads_init(const gmx::MDLogger &mdlog, t_commrec *cr,
                           int nthreads_hw_avail,
                           int omp_nthreads_req,
                           int omp_nthreads_pme_req,
                           gmx_bool bThisNodePMEOnly,
                           gmx_bool bFullOmpSupport)
{
    int        nppn;
    gmx_bool   bSepPME;

    const bool bOMP = GMX_OPENMP;

    /* number of MPI processes/threads per physical node */
    nppn = cr->nrank_intranode;

    bSepPME = (thisRankHasDuty(cr, DUTY_PP) != thisRankHasDuty(cr, DUTY_PME));

    manage_number_of_openmp_threads(mdlog, cr, bOMP,
                                    nthreads_hw_avail,
                                    omp_nthreads_req, omp_nthreads_pme_req,
                                    bThisNodePMEOnly, bFullOmpSupport,
                                    nppn, bSepPME);
#if GMX_THREAD_MPI
    /* Non-master threads have to wait for the OpenMP management to be
     * done, so that code elsewhere that uses OpenMP can be certain
     * the setup is complete. */
    if (PAR(cr))
    {
        MPI_Barrier(cr->mpi_comm_mysim);
    }
#endif

    reportOpenmpSettings(mdlog, cr, bOMP, bFullOmpSupport, bSepPME);
    issueOversubscriptionWarning(mdlog, cr, nthreads_hw_avail, nppn, bSepPME);
}

int gmx_omp_nthreads_get(int mod)
{
    if (mod < 0 || mod >= emntNR)
    {
        /* invalid module queried */
        return -1;
    }
    else
    {
        return modth.nth[mod];
    }
}

void
gmx_omp_nthreads_set(int mod, int nthreads)
{
    /* Catch an attempt to set the number of threads on an invalid
     * OpenMP module. */
    GMX_RELEASE_ASSERT(mod >= 0 && mod < emntNR, "Trying to set nthreads on invalid OpenMP module");

    modth.nth[mod] = nthreads;
}