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     sycl_atomdata_t*    adat        = nb->atdat;
  63     const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
  64     // Clear forces
  65     clearDeviceBufferAsync(&adat->f, 0, nb->atdat->natoms, 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     sycl_atomdata_t*    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->natoms   = -1;
  99     atomdata->numAlloc = -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         initParamLookupTable(
 128                 &nbp->nbfp, &nbp->nbfp_texobj, nbatParams.nbfp.data(), 2 * numTypes * numTypes, deviceContext);
 129     }
 130
 131     /* set up LJ-PME parameter lookup table */
 132     if (ic.vdwtype == evdwPME)
 133     {
 134         initParamLookupTable(
 135                 &nbp->nbfp_comb, &nbp->nbfp_comb_texobj, nbatParams.nbfp_comb.data(), 2 * numTypes, deviceContext);
 136     }
 137 }
 138
 139 NbnxmGpu* gpu_init(const gmx::DeviceStreamManager& deviceStreamManager,
 140                    const interaction_const_t*      ic,
 141                    const PairlistParams&           listParams,
 142                    const nbnxn_atomdata_t*         nbat,
 143                    const bool                      bLocalAndNonlocal)
 144 {
 145     auto* nb                              = new NbnxmGpu();
 146     nb->deviceContext_                    = &deviceStreamManager.context();
 147     nb->atdat                             = new sycl_atomdata_t;
 148     nb->nbparam                           = new NBParamGpu;
 149     nb->plist[InteractionLocality::Local] = new Nbnxm::gpu_plist;
 150     if (bLocalAndNonlocal)
 151     {
 152         nb->plist[InteractionLocality::NonLocal] = new Nbnxm::gpu_plist;
 153     }
 154
 155     nb->bUseTwoStreams = bLocalAndNonlocal;
 156
 157     nb->timers  = nullptr;
 158     nb->timings = nullptr;
 159
 160     /* init nbst */
 161     pmalloc(reinterpret_cast<void**>(&nb->nbst.e_lj), sizeof(*nb->nbst.e_lj));
 162     pmalloc(reinterpret_cast<void**>(&nb->nbst.e_el), sizeof(*nb->nbst.e_el));
 163     pmalloc(reinterpret_cast<void**>(&nb->nbst.fshift), SHIFTS * sizeof(*nb->nbst.fshift));
 164
 165     init_plist(nb->plist[InteractionLocality::Local]);
 166
 167     /* local/non-local GPU streams */
 168     GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedLocal),
 169                        "Local non-bonded stream should be initialized to use GPU for non-bonded.");
 170     nb->deviceStreams[InteractionLocality::Local] =
 171             &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedLocal);
 172     // In general, it's not strictly necessary to use 2 streams for SYCL, since they are
 173     // out-of-order. But for the time being, it will be less disruptive to keep them.
 174     if (nb->bUseTwoStreams)
 175     {
 176         init_plist(nb->plist[InteractionLocality::NonLocal]);
 177
 178         GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedNonLocal),
 179                            "Non-local non-bonded stream should be initialized to use GPU for "
 180                            "non-bonded with domain decomposition.");
 181         nb->deviceStreams[InteractionLocality::NonLocal] =
 182                 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedNonLocal);
 183     }
 184
 185     nb->xNonLocalCopyD2HDone = new GpuEventSynchronizer();
 186
 187     nb->bDoTime = false;
 188
 189     const nbnxn_atomdata_t::Params& nbatParams    = nbat->params();
 190     const DeviceContext&            deviceContext = *nb->deviceContext_;
 191
 192     initNbparam(nb->nbparam, *ic, listParams, nbatParams, deviceContext);
 193     initAtomdataFirst(nb, nbatParams.numTypes, deviceContext);
 194
 195     return nb;
 196 }
 197
 198 void gpu_upload_shiftvec(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom)
 199 {
 200     sycl_atomdata_t*    adat        = nb->atdat;
 201     const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
 202
 203     /* only if we have a dynamic box */
 204     if (nbatom->bDynamicBox || !adat->shiftVecUploaded)
 205     {
 206         GMX_ASSERT(adat->shiftVec.elementSize() == sizeof(nbatom->shift_vec[0]),
 207                    "Sizes of host- and device-side shift vectors should be the same.");
 208         copyToDeviceBuffer(&adat->shiftVec,
 209                            reinterpret_cast<const float3*>(nbatom->shift_vec.data()),
 210                            0,
 211                            SHIFTS,
 212                            localStream,
 213                            GpuApiCallBehavior::Async,
 214                            nullptr);
 215         adat->shiftVecUploaded = true;
 216     }
 217 }
 218
 219 void gpu_init_atomdata(NbnxmGpu* nb, const nbnxn_atomdata_t* nbat)
 220 {
 221     GMX_ASSERT(!nb->bDoTime, "Timing on SYCL not supported yet");
 222     sycl_atomdata_t*     atdat         = nb->atdat;
 223     const DeviceContext& deviceContext = *nb->deviceContext_;
 224     const DeviceStream&  localStream   = *nb->deviceStreams[InteractionLocality::Local];
 225
 226     int  numAtoms    = nbat->numAtoms();
 227     bool reallocated = false;
 228     if (numAtoms > atdat->numAlloc)
 229     {
 230         int numAlloc = over_alloc_small(numAtoms);
 231
 232         /* free up first if the arrays have already been initialized */
 233         if (atdat->numAlloc != -1)
 234         {
 235             freeDeviceBuffer(&atdat->f);
 236             freeDeviceBuffer(&atdat->xq);
 237             freeDeviceBuffer(&atdat->atomTypes);
 238             freeDeviceBuffer(&atdat->ljComb);
 239         }
 240
 241         allocateDeviceBuffer(&atdat->f, numAlloc, deviceContext);
 242         allocateDeviceBuffer(&atdat->xq, numAlloc, deviceContext);
 243         if (useLjCombRule(nb->nbparam->vdwType))
 244         {
 245             allocateDeviceBuffer(&atdat->ljComb, numAlloc, deviceContext);
 246         }
 247         else
 248         {
 249             allocateDeviceBuffer(&atdat->atomTypes, numAlloc, deviceContext);
 250         }
 251
 252         atdat->numAlloc = numAlloc;
 253         reallocated     = true;
 254     }
 255
 256     atdat->natoms       = numAtoms;
 257     atdat->natoms_local = nbat->natoms_local;
 258
 259     /* need to clear GPU f output if realloc happened */
 260     if (reallocated)
 261     {
 262         clearDeviceBufferAsync(&atdat->f, 0, atdat->numAlloc, localStream);
 263     }
 264
 265     if (useLjCombRule(nb->nbparam->vdwType))
 266     {
 267         GMX_ASSERT(atdat->ljComb.elementSize() == sizeof(float2),
 268                    "Size of the LJ parameters element should be equal to the size of float2.");
 269         copyToDeviceBuffer(&atdat->ljComb,
 270                            reinterpret_cast<const float2*>(nbat->params().lj_comb.data()),
 271                            0,
 272                            numAtoms,
 273                            localStream,
 274                            GpuApiCallBehavior::Async,
 275                            nullptr);
 276     }
 277     else
 278     {
 279         GMX_ASSERT(atdat->atomTypes.elementSize() == sizeof(nbat->params().type[0]),
 280                    "Sizes of host- and device-side atom types should be the same.");
 281         copyToDeviceBuffer(&atdat->atomTypes,
 282                            nbat->params().type.data(),
 283                            0,
 284                            numAtoms,
 285                            localStream,
 286                            GpuApiCallBehavior::Async,
 287                            nullptr);
 288     }
 289 }
 290
 291 void gpu_free(NbnxmGpu* nb)
 292 {
 293     if (nb == nullptr)
 294     {
 295         return;
 296     }
 297
 298     sycl_atomdata_t* atdat   = nb->atdat;
 299     NBParamGpu*      nbparam = nb->nbparam;
 300
 301     if ((!nbparam->coulomb_tab)
 302         && (nbparam->elecType == ElecType::EwaldTab || nbparam->elecType == ElecType::EwaldTabTwin))
 303     {
 304         destroyParamLookupTable(&nbparam->coulomb_tab, nbparam->coulomb_tab_texobj);
 305     }
 306
 307     if (!useLjCombRule(nb->nbparam->vdwType))
 308     {
 309         destroyParamLookupTable(&nbparam->nbfp, nbparam->nbfp_texobj);
 310     }
 311
 312     if (nbparam->vdwType == VdwType::EwaldGeom || nbparam->vdwType == VdwType::EwaldLB)
 313     {
 314         destroyParamLookupTable(&nbparam->nbfp_comb, nbparam->nbfp_comb_texobj);
 315     }
 316
 317     /* Free plist */
 318     auto* plist = nb->plist[InteractionLocality::Local];
 319     delete plist;
 320     if (nb->bUseTwoStreams)
 321     {
 322         auto* plist_nl = nb->plist[InteractionLocality::NonLocal];
 323         delete plist_nl;
 324     }
 325
 326     /* Free nbst */
 327     pfree(nb->nbst.e_lj);
 328     nb->nbst.e_lj = nullptr;
 329
 330     pfree(nb->nbst.e_el);
 331     nb->nbst.e_el = nullptr;
 332
 333     pfree(nb->nbst.fshift);
 334     nb->nbst.fshift = nullptr;
 335
 336     delete atdat;
 337     delete nbparam;
 338     delete nb;
 339 }
 340
 341 int gpu_min_ci_balanced(NbnxmGpu* nb)
 342 {
 343     // SYCL-TODO: Logic and magic values taken from OpenCL
 344     static constexpr unsigned int balancedFactor = 50;
 345     if (nb == nullptr)
 346     {
 347         return 0;
 348     }
 349     const cl::sycl::device device = nb->deviceContext_->deviceInfo().syclDevice;
 350     const int numComputeUnits     = device.get_info<cl::sycl::info::device::max_compute_units>();
 351     return balancedFactor * numComputeUnits;
 352 }
 353
 354 } // namespace Nbnxm