Simplify PME GPU constants

[gromacs.git] / src / gromacs / ewald / pme_pp.cpp

/*
 * This file is part of the GROMACS molecular simulation package.
 *
 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
 * Copyright (c) 2001-2004, The GROMACS development team.
 * Copyright (c) 2013,2014,2015,2016,2017 by the GROMACS development team.
 * Copyright (c) 2018,2019,2020, 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.
 */
/*! \internal \file
 *
 * \brief This file contains function definitions necessary for
 * managing the offload of long-ranged PME work to separate MPI rank,
 * for computing energies and forces (Coulomb and LJ).
 *
 * \author Berk Hess <hess@kth.se>
 * \ingroup module_ewald
 */

#include "gmxpre.h"

#include "pme_pp.h"

#include "config.h"

#include <cstdio>
#include <cstring>

#include "gromacs/domdec/domdec.h"
#include "gromacs/domdec/domdec_struct.h"
#include "gromacs/ewald/pme.h"
#include "gromacs/ewald/pme_pp_comm_gpu.h"
#include "gromacs/gmxlib/network.h"
#include "gromacs/math/vec.h"
#include "gromacs/mdlib/gmx_omp_nthreads.h"
#include "gromacs/mdtypes/commrec.h"
#include "gromacs/mdtypes/forceoutput.h"
#include "gromacs/mdtypes/forcerec.h"
#include "gromacs/mdtypes/interaction_const.h"
#include "gromacs/mdtypes/md_enums.h"
#include "gromacs/mdtypes/state_propagator_data_gpu.h"
#include "gromacs/nbnxm/nbnxm.h"
#include "gromacs/timing/wallcycle.h"
#include "gromacs/utility/fatalerror.h"
#include "gromacs/utility/gmxmpi.h"
#include "gromacs/utility/smalloc.h"

#include "pme_pp_communication.h"

/*! \brief Block to wait for communication to PME ranks to complete
 *
 * This should be faster with a real non-blocking MPI implementation
 */
static constexpr bool c_useDelayedWait = false;

/*! \brief Wait for the pending data send requests to PME ranks to complete */
static void gmx_pme_send_coeffs_coords_wait(gmx_domdec_t* dd)
{
    if (dd->nreq_pme)
    {
#if GMX_MPI
        MPI_Waitall(dd->nreq_pme, dd->req_pme, MPI_STATUSES_IGNORE);
#endif
        dd->nreq_pme = 0;
    }
}

/*! \brief Send data to PME ranks */
static void gmx_pme_send_coeffs_coords(t_forcerec*      fr,
                                       const t_commrec* cr,
                                       unsigned int     flags,
                                       real gmx_unused* chargeA,
                                       real gmx_unused* chargeB,
                                       real gmx_unused* c6A,
                                       real gmx_unused* c6B,
                                       real gmx_unused* sigmaA,
                                       real gmx_unused* sigmaB,
                                       const matrix     box,
                                       const rvec gmx_unused* x,
                                       real                   lambda_q,
                                       real                   lambda_lj,
                                       int                    maxshift_x,
                                       int                    maxshift_y,
                                       int64_t                step,
                                       bool                   useGpuPmePpComms,
                                       bool                   reinitGpuPmePpComms,
                                       bool                   sendCoordinatesFromGpu,
                                       GpuEventSynchronizer*  coordinatesReadyOnDeviceEvent)
{
    gmx_domdec_t*         dd;
    gmx_pme_comm_n_box_t* cnb;
    int                   n;

    dd = cr->dd;
    n  = dd_numHomeAtoms(*dd);

    if (debug)
    {
        fprintf(debug, "PP rank %d sending to PME rank %d: %d%s%s%s%s\n", cr->sim_nodeid, dd->pme_nodeid,
                n, (flags & PP_PME_CHARGE) ? " charges" : "", (flags & PP_PME_SQRTC6) ? " sqrtC6" : "",
                (flags & PP_PME_SIGMA) ? " sigma" : "", (flags & PP_PME_COORD) ? " coordinates" : "");
    }

    if (useGpuPmePpComms)
    {
        flags |= PP_PME_GPUCOMMS;
    }

    if (c_useDelayedWait)
    {
        /* We can not use cnb until pending communication has finished */
        gmx_pme_send_coeffs_coords_wait(dd);
    }

    if (dd->pme_receive_vir_ener)
    {
        /* Peer PP node: communicate all data */
        if (dd->cnb == nullptr)
        {
            snew(dd->cnb, 1);
        }
        cnb = dd->cnb;

        cnb->flags      = flags;
        cnb->natoms     = n;
        cnb->maxshift_x = maxshift_x;
        cnb->maxshift_y = maxshift_y;
        cnb->lambda_q   = lambda_q;
        cnb->lambda_lj  = lambda_lj;
        cnb->step       = step;
        if (flags & PP_PME_COORD)
        {
            copy_mat(box, cnb->box);
        }
#if GMX_MPI
        MPI_Isend(cnb, sizeof(*cnb), MPI_BYTE, dd->pme_nodeid, eCommType_CNB, cr->mpi_comm_mysim,
                  &dd->req_pme[dd->nreq_pme++]);
#endif
    }
    else if (flags & (PP_PME_CHARGE | PP_PME_SQRTC6 | PP_PME_SIGMA))
    {
#if GMX_MPI
        /* Communicate only the number of atoms */
        MPI_Isend(&n, sizeof(n), MPI_BYTE, dd->pme_nodeid, eCommType_CNB, cr->mpi_comm_mysim,
                  &dd->req_pme[dd->nreq_pme++]);
#endif
    }

#if GMX_MPI
    if (n > 0)
    {
        if (flags & PP_PME_CHARGE)
        {
            MPI_Isend(chargeA, n * sizeof(real), MPI_BYTE, dd->pme_nodeid, eCommType_ChargeA,
                      cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
        }
        if (flags & PP_PME_CHARGEB)
        {
            MPI_Isend(chargeB, n * sizeof(real), MPI_BYTE, dd->pme_nodeid, eCommType_ChargeB,
                      cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
        }
        if (flags & PP_PME_SQRTC6)
        {
            MPI_Isend(c6A, n * sizeof(real), MPI_BYTE, dd->pme_nodeid, eCommType_SQRTC6A,
                      cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
        }
        if (flags & PP_PME_SQRTC6B)
        {
            MPI_Isend(c6B, n * sizeof(real), MPI_BYTE, dd->pme_nodeid, eCommType_SQRTC6B,
                      cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
        }
        if (flags & PP_PME_SIGMA)
        {
            MPI_Isend(sigmaA, n * sizeof(real), MPI_BYTE, dd->pme_nodeid, eCommType_SigmaA,
                      cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
        }
        if (flags & PP_PME_SIGMAB)
        {
            MPI_Isend(sigmaB, n * sizeof(real), MPI_BYTE, dd->pme_nodeid, eCommType_SigmaB,
                      cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
        }
        if (flags & PP_PME_COORD)
        {
            if (reinitGpuPmePpComms)
            {
                fr->pmePpCommGpu->reinit(n);
            }


            /* MPI_Isend does not accept a const buffer pointer */
            real* xRealPtr = const_cast<real*>(x[0]);
            if (useGpuPmePpComms && (fr != nullptr))
            {
                void* sendPtr = sendCoordinatesFromGpu
                                        ? static_cast<void*>(fr->stateGpu->getCoordinates())
                                        : static_cast<void*>(xRealPtr);
                fr->pmePpCommGpu->sendCoordinatesToPmeCudaDirect(sendPtr, n, sendCoordinatesFromGpu,
                                                                 coordinatesReadyOnDeviceEvent);
            }
            else
            {
                MPI_Isend(xRealPtr, n * sizeof(rvec), MPI_BYTE, dd->pme_nodeid, eCommType_COORD,
                          cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
            }
        }
    }
#else
    GMX_UNUSED_VALUE(fr);
    GMX_UNUSED_VALUE(reinitGpuPmePpComms);
    GMX_UNUSED_VALUE(sendCoordinatesFromGpu);
    GMX_UNUSED_VALUE(coordinatesReadyOnDeviceEvent);
#endif
    if (!c_useDelayedWait)
    {
        /* Wait for the data to arrive */
        /* We can skip this wait as we are sure x and q will not be modified
         * before the next call to gmx_pme_send_x_q or gmx_pme_receive_f.
         */
        gmx_pme_send_coeffs_coords_wait(dd);
    }
}

void gmx_pme_send_parameters(const t_commrec*           cr,
                             const interaction_const_t* ic,
                             gmx_bool                   bFreeEnergy_q,
                             gmx_bool                   bFreeEnergy_lj,
                             real*                      chargeA,
                             real*                      chargeB,
                             real*                      sqrt_c6A,
                             real*                      sqrt_c6B,
                             real*                      sigmaA,
                             real*                      sigmaB,
                             int                        maxshift_x,
                             int                        maxshift_y)
{
    unsigned int flags = 0;

    if (EEL_PME(ic->eeltype))
    {
        flags |= PP_PME_CHARGE;
    }
    if (EVDW_PME(ic->vdwtype))
    {
        flags |= (PP_PME_SQRTC6 | PP_PME_SIGMA);
    }
    if (bFreeEnergy_q || bFreeEnergy_lj)
    {
        /* Assumes that the B state flags are in the bits just above
         * the ones for the A state. */
        flags |= (flags << 1);
    }

    gmx_pme_send_coeffs_coords(nullptr, cr, flags, chargeA, chargeB, sqrt_c6A, sqrt_c6B, sigmaA,
                               sigmaB, nullptr, nullptr, 0, 0, maxshift_x, maxshift_y, -1, false,
                               false, false, nullptr);
}

void gmx_pme_send_coordinates(t_forcerec*           fr,
                              const t_commrec*      cr,
                              const matrix          box,
                              const rvec*           x,
                              real                  lambda_q,
                              real                  lambda_lj,
                              bool                  computeEnergyAndVirial,
                              int64_t               step,
                              bool                  useGpuPmePpComms,
                              bool                  receiveCoordinateAddressFromPme,
                              bool                  sendCoordinatesFromGpu,
                              GpuEventSynchronizer* coordinatesReadyOnDeviceEvent,
                              gmx_wallcycle*        wcycle)
{
    wallcycle_start(wcycle, ewcPP_PMESENDX);

    unsigned int flags = PP_PME_COORD;
    if (computeEnergyAndVirial)
    {
        flags |= PP_PME_ENER_VIR;
    }
    gmx_pme_send_coeffs_coords(fr, cr, flags, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr,
                               box, x, lambda_q, lambda_lj, 0, 0, step, useGpuPmePpComms,
                               receiveCoordinateAddressFromPme, sendCoordinatesFromGpu,
                               coordinatesReadyOnDeviceEvent);

    wallcycle_stop(wcycle, ewcPP_PMESENDX);
}

void gmx_pme_send_finish(const t_commrec* cr)
{
    unsigned int flags = PP_PME_FINISH;

    gmx_pme_send_coeffs_coords(nullptr, cr, flags, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr,
                               nullptr, nullptr, 0, 0, 0, 0, -1, false, false, false, nullptr);
}

void gmx_pme_send_switchgrid(const t_commrec* cr, ivec grid_size, real ewaldcoeff_q, real ewaldcoeff_lj)
{
#if GMX_MPI
    gmx_pme_comm_n_box_t cnb;

    /* Only let one PP node signal each PME node */
    if (cr->dd->pme_receive_vir_ener)
    {
        cnb.flags = PP_PME_SWITCHGRID;
        copy_ivec(grid_size, cnb.grid_size);
        cnb.ewaldcoeff_q  = ewaldcoeff_q;
        cnb.ewaldcoeff_lj = ewaldcoeff_lj;

        /* We send this, uncommon, message blocking to simplify the code */
        MPI_Send(&cnb, sizeof(cnb), MPI_BYTE, cr->dd->pme_nodeid, eCommType_CNB, cr->mpi_comm_mysim);
    }
#else
    GMX_UNUSED_VALUE(cr);
    GMX_UNUSED_VALUE(grid_size);
    GMX_UNUSED_VALUE(ewaldcoeff_q);
    GMX_UNUSED_VALUE(ewaldcoeff_lj);
#endif
}

void gmx_pme_send_resetcounters(const t_commrec gmx_unused* cr, int64_t gmx_unused step)
{
#if GMX_MPI
    gmx_pme_comm_n_box_t cnb;

    /* Only let one PP node signal each PME node */
    if (cr->dd->pme_receive_vir_ener)
    {
        cnb.flags = PP_PME_RESETCOUNTERS;
        cnb.step  = step;

        /* We send this, uncommon, message blocking to simplify the code */
        MPI_Send(&cnb, sizeof(cnb), MPI_BYTE, cr->dd->pme_nodeid, eCommType_CNB, cr->mpi_comm_mysim);
    }
#endif
}

/*! \brief Receive virial and energy from PME rank */
static void receive_virial_energy(const t_commrec*      cr,
                                  gmx::ForceWithVirial* forceWithVirial,
                                  real*                 energy_q,
                                  real*                 energy_lj,
                                  real*                 dvdlambda_q,
                                  real*                 dvdlambda_lj,
                                  float*                pme_cycles)
{
    gmx_pme_comm_vir_ene_t cve;

    if (cr->dd->pme_receive_vir_ener)
    {
        if (debug)
        {
            fprintf(debug, "PP rank %d receiving from PME rank %d: virial and energy\n",
                    cr->sim_nodeid, cr->dd->pme_nodeid);
        }
#if GMX_MPI
        MPI_Recv(&cve, sizeof(cve), MPI_BYTE, cr->dd->pme_nodeid, 1, cr->mpi_comm_mysim, MPI_STATUS_IGNORE);
#else
        memset(&cve, 0, sizeof(cve));
#endif

        forceWithVirial->addVirialContribution(cve.vir_q);
        forceWithVirial->addVirialContribution(cve.vir_lj);
        *energy_q  = cve.energy_q;
        *energy_lj = cve.energy_lj;
        *dvdlambda_q += cve.dvdlambda_q;
        *dvdlambda_lj += cve.dvdlambda_lj;
        *pme_cycles = cve.cycles;

        if (cve.stop_cond != gmx_stop_cond_none)
        {
            gmx_set_stop_condition(cve.stop_cond);
        }
    }
    else
    {
        *energy_q   = 0;
        *energy_lj  = 0;
        *pme_cycles = 0;
    }
}

/*! \brief Recieve force data from PME ranks */
static void recvFFromPme(gmx::PmePpCommGpu* pmePpCommGpu,
                         void*              recvptr,
                         int                n,
                         const t_commrec*   cr,
                         bool               useGpuPmePpComms,
                         bool               receivePmeForceToGpu)
{
    if (useGpuPmePpComms)
    {
        GMX_ASSERT(pmePpCommGpu != nullptr, "Need valid pmePpCommGpu");
        // Receive directly using CUDA memory copy
        pmePpCommGpu->receiveForceFromPmeCudaDirect(recvptr, n, receivePmeForceToGpu);
    }
    else
    {
        // Receive data using MPI
#if GMX_MPI
        MPI_Recv(recvptr, n * sizeof(rvec), MPI_BYTE, cr->dd->pme_nodeid, 0, cr->mpi_comm_mysim,
                 MPI_STATUS_IGNORE);
#else
        GMX_UNUSED_VALUE(cr);
#endif
    }
}


void gmx_pme_receive_f(gmx::PmePpCommGpu*    pmePpCommGpu,
                       const t_commrec*      cr,
                       gmx::ForceWithVirial* forceWithVirial,
                       real*                 energy_q,
                       real*                 energy_lj,
                       real*                 dvdlambda_q,
                       real*                 dvdlambda_lj,
                       bool                  useGpuPmePpComms,
                       bool                  receivePmeForceToGpu,
                       float*                pme_cycles)
{
    if (c_useDelayedWait)
    {
        /* Wait for the x request to finish */
        gmx_pme_send_coeffs_coords_wait(cr->dd);
    }

    const int               natoms = dd_numHomeAtoms(*cr->dd);
    std::vector<gmx::RVec>& buffer = cr->dd->pmeForceReceiveBuffer;
    buffer.resize(natoms);

    void* recvptr = reinterpret_cast<void*>(buffer.data());
    recvFFromPme(pmePpCommGpu, recvptr, natoms, cr, useGpuPmePpComms, receivePmeForceToGpu);

    int nt = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, natoms);

    gmx::ArrayRef<gmx::RVec> f = forceWithVirial->force_;

    if (!receivePmeForceToGpu)
    {
        /* Note that we would like to avoid this conditional by putting it
         * into the omp pragma instead, but then we still take the full
         * omp parallel for overhead (at least with gcc5).
         */
        if (nt == 1)
        {
            for (int i = 0; i < natoms; i++)
            {
                f[i] += buffer[i];
            }
        }
        else
        {
#pragma omp parallel for num_threads(nt) schedule(static)
            for (int i = 0; i < natoms; i++)
            {
                f[i] += buffer[i];
            }
        }
    }

    receive_virial_energy(cr, forceWithVirial, energy_q, energy_lj, dvdlambda_q, dvdlambda_lj, pme_cycles);
}