src/gromacs/nbnxm/sycl/nbnxm_sycl_data_mgmt.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2020,2021, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35
  36 /*! \internal \file
  37  *  \brief
  38  *  Stubs of functions that must be defined by nbnxm sycl implementation.
  39  *
  40  *  \ingroup module_nbnxm
  41  */
  42 #include "gmxpre.h"
  43
  44 #include "gromacs/gpu_utils/device_stream_manager.h"
  45 #include "gromacs/hardware/device_information.h"
  46 #include "gromacs/mdtypes/interaction_const.h"
  47 #include "gromacs/nbnxm/atomdata.h"
  48 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  49 #include "gromacs/nbnxm/nbnxm_gpu.h"
  50 #include "gromacs/nbnxm/nbnxm_gpu_data_mgmt.h"
  51 #include "gromacs/pbcutil/ishift.h"
  52 #include "gromacs/utility/exceptions.h"
  53
  54 #include "nbnxm_sycl_types.h"
  55
  56 namespace Nbnxm
  57 {
  58
  59 //! This function is documented in the header file
  60 void gpu_clear_outputs(NbnxmGpu* nb, bool computeVirial)
  61 {
  62     NBAtomData*         adat        = nb->atdat;
  63     const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
  64     // Clear forces
  65     clearDeviceBufferAsync(&adat->f, 0, nb->atdat->numAtoms, localStream);
  66     // Clear shift force array and energies if the outputs were used in the current step
  67     if (computeVirial)
  68     {
  69         clearDeviceBufferAsync(&adat->fShift, 0, SHIFTS, localStream);
  70         clearDeviceBufferAsync(&adat->eLJ, 0, 1, localStream);
  71         clearDeviceBufferAsync(&adat->eElec, 0, 1, localStream);
  72     }
  73 }
  74
  75 /*! \brief Initialize \p atomdata first time; it only gets filled at pair-search. */
  76 static void initAtomdataFirst(NbnxmGpu* nb, int numTypes, const DeviceContext& deviceContext)
  77 {
  78     const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
  79     NBAtomData*         atomdata    = nb->atdat;
  80     atomdata->numTypes              = numTypes;
  81     allocateDeviceBuffer(&atomdata->shiftVec, SHIFTS, deviceContext);
  82     atomdata->shiftVecUploaded = false;
  83
  84     allocateDeviceBuffer(&atomdata->fShift, SHIFTS, deviceContext);
  85     allocateDeviceBuffer(&atomdata->eLJ, 1, deviceContext);
  86     allocateDeviceBuffer(&atomdata->eElec, 1, deviceContext);
  87
  88     clearDeviceBufferAsync(&atomdata->fShift, 0, SHIFTS, localStream);
  89     clearDeviceBufferAsync(&atomdata->eElec, 0, 1, localStream);
  90     clearDeviceBufferAsync(&atomdata->eLJ, 0, 1, localStream);
  91
  92     /* initialize to nullptr pointers to data that is not allocated here and will
  93        need reallocation in later */
  94     atomdata->xq = nullptr;
  95     atomdata->f  = nullptr;
  96
  97     /* size -1 indicates that the respective array hasn't been initialized yet */
  98     atomdata->numAtoms      = -1;
  99     atomdata->numAtomsAlloc = -1;
 100 }
 101
 102 /*! \brief Initialize the nonbonded parameter data structure. */
 103 static void initNbparam(NBParamGpu*                     nbp,
 104                         const interaction_const_t&      ic,
 105                         const PairlistParams&           listParams,
 106                         const nbnxn_atomdata_t::Params& nbatParams,
 107                         const DeviceContext&            deviceContext)
 108 {
 109     const int numTypes = nbatParams.numTypes;
 110
 111     set_cutoff_parameters(nbp, &ic, listParams);
 112
 113     nbp->vdwType  = nbnxmGpuPickVdwKernelType(&ic, nbatParams.ljCombinationRule);
 114     nbp->elecType = nbnxmGpuPickElectrostaticsKernelType(&ic, deviceContext.deviceInfo());
 115
 116     /* generate table for PME */
 117     nbp->coulomb_tab = nullptr;
 118     if (nbp->elecType == ElecType::EwaldTab || nbp->elecType == ElecType::EwaldTabTwin)
 119     {
 120         GMX_RELEASE_ASSERT(ic.coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
 121         init_ewald_coulomb_force_table(*ic.coulombEwaldTables, nbp, deviceContext);
 122     }
 123
 124     /* set up LJ parameter lookup table */
 125     if (!useLjCombRule(nbp->vdwType))
 126     {
 127         static_assert(sizeof(decltype(nbp->nbfp)) == 2 * sizeof(decltype(*nbatParams.nbfp.data())),
 128                       "Mismatch in the size of host / device data types");
 129         initParamLookupTable(&nbp->nbfp,
 130                              &nbp->nbfp_texobj,
 131                              reinterpret_cast<const Float2*>(nbatParams.nbfp.data()),
 132                              numTypes * numTypes,
 133                              deviceContext);
 134     }
 135
 136     /* set up LJ-PME parameter lookup table */
 137     if (ic.vdwtype == VanDerWaalsType::Pme)
 138     {
 139         static_assert(sizeof(decltype(nbp->nbfp_comb))
 140                               == 2 * sizeof(decltype(*nbatParams.nbfp_comb.data())),
 141                       "Mismatch in the size of host / device data types");
 142         initParamLookupTable(&nbp->nbfp_comb,
 143                              &nbp->nbfp_comb_texobj,
 144                              reinterpret_cast<const Float2*>(nbatParams.nbfp_comb.data()),
 145                              numTypes,
 146                              deviceContext);
 147     }
 148 }
 149
 150 NbnxmGpu* gpu_init(const gmx::DeviceStreamManager& deviceStreamManager,
 151                    const interaction_const_t*      ic,
 152                    const PairlistParams&           listParams,
 153                    const nbnxn_atomdata_t*         nbat,
 154                    const bool                      bLocalAndNonlocal)
 155 {
 156     auto* nb                              = new NbnxmGpu();
 157     nb->deviceContext_                    = &deviceStreamManager.context();
 158     nb->atdat                             = new NBAtomData;
 159     nb->nbparam                           = new NBParamGpu;
 160     nb->plist[InteractionLocality::Local] = new Nbnxm::gpu_plist;
 161     if (bLocalAndNonlocal)
 162     {
 163         nb->plist[InteractionLocality::NonLocal] = new Nbnxm::gpu_plist;
 164     }
 165
 166     nb->bUseTwoStreams = bLocalAndNonlocal;
 167
 168     nb->timers  = nullptr;
 169     nb->timings = nullptr;
 170
 171     /* init nbst */
 172     pmalloc(reinterpret_cast<void**>(&nb->nbst.eLJ), sizeof(*nb->nbst.eLJ));
 173     pmalloc(reinterpret_cast<void**>(&nb->nbst.eElec), sizeof(*nb->nbst.eElec));
 174     pmalloc(reinterpret_cast<void**>(&nb->nbst.fShift), SHIFTS * sizeof(*nb->nbst.fShift));
 175
 176     init_plist(nb->plist[InteractionLocality::Local]);
 177
 178     /* local/non-local GPU streams */
 179     GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedLocal),
 180                        "Local non-bonded stream should be initialized to use GPU for non-bonded.");
 181     nb->deviceStreams[InteractionLocality::Local] =
 182             &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedLocal);
 183     // In general, it's not strictly necessary to use 2 streams for SYCL, since they are
 184     // out-of-order. But for the time being, it will be less disruptive to keep them.
 185     if (nb->bUseTwoStreams)
 186     {
 187         init_plist(nb->plist[InteractionLocality::NonLocal]);
 188
 189         GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedNonLocal),
 190                            "Non-local non-bonded stream should be initialized to use GPU for "
 191                            "non-bonded with domain decomposition.");
 192         nb->deviceStreams[InteractionLocality::NonLocal] =
 193                 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedNonLocal);
 194     }
 195
 196     nb->bDoTime = false;
 197
 198     const nbnxn_atomdata_t::Params& nbatParams    = nbat->params();
 199     const DeviceContext&            deviceContext = *nb->deviceContext_;
 200
 201     initNbparam(nb->nbparam, *ic, listParams, nbatParams, deviceContext);
 202     initAtomdataFirst(nb, nbatParams.numTypes, deviceContext);
 203
 204     return nb;
 205 }
 206
 207 void gpu_upload_shiftvec(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom)
 208 {
 209     NBAtomData*         adat        = nb->atdat;
 210     const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
 211
 212     /* only if we have a dynamic box */
 213     if (nbatom->bDynamicBox || !adat->shiftVecUploaded)
 214     {
 215         GMX_ASSERT(adat->shiftVec.elementSize() == sizeof(nbatom->shift_vec[0]),
 216                    "Sizes of host- and device-side shift vectors should be the same.");
 217         copyToDeviceBuffer(&adat->shiftVec,
 218                            reinterpret_cast<const Float3*>(nbatom->shift_vec.data()),
 219                            0,
 220                            SHIFTS,
 221                            localStream,
 222                            GpuApiCallBehavior::Async,
 223                            nullptr);
 224         adat->shiftVecUploaded = true;
 225     }
 226 }
 227
 228 void gpu_init_atomdata(NbnxmGpu* nb, const nbnxn_atomdata_t* nbat)
 229 {
 230     GMX_ASSERT(!nb->bDoTime, "Timing on SYCL not supported yet");
 231     NBAtomData*          atdat         = nb->atdat;
 232     const DeviceContext& deviceContext = *nb->deviceContext_;
 233     const DeviceStream&  localStream   = *nb->deviceStreams[InteractionLocality::Local];
 234
 235     int  numAtoms    = nbat->numAtoms();
 236     bool reallocated = false;
 237     if (numAtoms > atdat->numAtomsAlloc)
 238     {
 239         int numAlloc = over_alloc_small(numAtoms);
 240
 241         /* free up first if the arrays have already been initialized */
 242         if (atdat->numAtomsAlloc != -1)
 243         {
 244             freeDeviceBuffer(&atdat->f);
 245             freeDeviceBuffer(&atdat->xq);
 246             freeDeviceBuffer(&atdat->atomTypes);
 247             freeDeviceBuffer(&atdat->ljComb);
 248         }
 249
 250         allocateDeviceBuffer(&atdat->f, numAlloc, deviceContext);
 251         allocateDeviceBuffer(&atdat->xq, numAlloc, deviceContext);
 252         if (useLjCombRule(nb->nbparam->vdwType))
 253         {
 254             allocateDeviceBuffer(&atdat->ljComb, numAlloc, deviceContext);
 255         }
 256         else
 257         {
 258             allocateDeviceBuffer(&atdat->atomTypes, numAlloc, deviceContext);
 259         }
 260
 261         atdat->numAtomsAlloc = numAlloc;
 262         reallocated          = true;
 263     }
 264
 265     atdat->numAtoms      = numAtoms;
 266     atdat->numAtomsLocal = nbat->natoms_local;
 267
 268     /* need to clear GPU f output if realloc happened */
 269     if (reallocated)
 270     {
 271         clearDeviceBufferAsync(&atdat->f, 0, atdat->numAtomsAlloc, localStream);
 272     }
 273
 274     if (useLjCombRule(nb->nbparam->vdwType))
 275     {
 276         GMX_ASSERT(atdat->ljComb.elementSize() == sizeof(Float2),
 277                    "Size of the LJ parameters element should be equal to the size of float2.");
 278         copyToDeviceBuffer(&atdat->ljComb,
 279                            reinterpret_cast<const Float2*>(nbat->params().lj_comb.data()),
 280                            0,
 281                            numAtoms,
 282                            localStream,
 283                            GpuApiCallBehavior::Async,
 284                            nullptr);
 285     }
 286     else
 287     {
 288         GMX_ASSERT(atdat->atomTypes.elementSize() == sizeof(nbat->params().type[0]),
 289                    "Sizes of host- and device-side atom types should be the same.");
 290         copyToDeviceBuffer(&atdat->atomTypes,
 291                            nbat->params().type.data(),
 292                            0,
 293                            numAtoms,
 294                            localStream,
 295                            GpuApiCallBehavior::Async,
 296                            nullptr);
 297     }
 298 }
 299
 300 void gpu_free(NbnxmGpu* nb)
 301 {
 302     if (nb == nullptr)
 303     {
 304         return;
 305     }
 306
 307     NBAtomData* atdat   = nb->atdat;
 308     NBParamGpu* nbparam = nb->nbparam;
 309
 310     if ((!nbparam->coulomb_tab)
 311         && (nbparam->elecType == ElecType::EwaldTab || nbparam->elecType == ElecType::EwaldTabTwin))
 312     {
 313         destroyParamLookupTable(&nbparam->coulomb_tab, nbparam->coulomb_tab_texobj);
 314     }
 315
 316     if (!useLjCombRule(nb->nbparam->vdwType))
 317     {
 318         destroyParamLookupTable(&nbparam->nbfp, nbparam->nbfp_texobj);
 319     }
 320
 321     if (nbparam->vdwType == VdwType::EwaldGeom || nbparam->vdwType == VdwType::EwaldLB)
 322     {
 323         destroyParamLookupTable(&nbparam->nbfp_comb, nbparam->nbfp_comb_texobj);
 324     }
 325
 326     /* Free plist */
 327     auto* plist = nb->plist[InteractionLocality::Local];
 328     delete plist;
 329     if (nb->bUseTwoStreams)
 330     {
 331         auto* plist_nl = nb->plist[InteractionLocality::NonLocal];
 332         delete plist_nl;
 333     }
 334
 335     /* Free nbst */
 336     pfree(nb->nbst.eLJ);
 337     nb->nbst.eLJ = nullptr;
 338
 339     pfree(nb->nbst.eElec);
 340     nb->nbst.eElec = nullptr;
 341
 342     pfree(nb->nbst.fShift);
 343     nb->nbst.fShift = nullptr;
 344
 345     delete atdat;
 346     delete nbparam;
 347     delete nb;
 348 }
 349
 350 int gpu_min_ci_balanced(NbnxmGpu* nb)
 351 {
 352     // SYCL-TODO: Logic and magic values taken from OpenCL
 353     static constexpr unsigned int balancedFactor = 50;
 354     if (nb == nullptr)
 355     {
 356         return 0;
 357     }
 358     const cl::sycl::device device = nb->deviceContext_->deviceInfo().syclDevice;
 359     const int numComputeUnits     = device.get_info<cl::sycl::info::device::max_compute_units>();
 360     return balancedFactor * numComputeUnits;
 361 }
 362
 363 } // namespace Nbnxm