src/gromacs/ewald/pme_gpu.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2016,2017,2018,2019,2020,2021, 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 Implements high-level PME GPU functions which do not require GPU framework-specific code.
  38  *
  39  * \author Aleksei Iupinov <a.yupinov@gmail.com>
  40  * \ingroup module_ewald
  41  */
  42
  43 #include "gmxpre.h"
  44
  45 #include "config.h"
  46
  47 #include <list>
  48
  49 #include "gromacs/ewald/ewald_utils.h"
  50 #include "gromacs/ewald/pme.h"
  51 #include "gromacs/fft/parallel_3dfft.h"
  52 #include "gromacs/math/invertmatrix.h"
  53 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  54 #include "gromacs/mdtypes/enerdata.h"
  55 #include "gromacs/mdtypes/forceoutput.h"
  56 #include "gromacs/mdtypes/inputrec.h"
  57 #include "gromacs/mdtypes/simulation_workload.h"
  58 #include "gromacs/utility/exceptions.h"
  59 #include "gromacs/utility/fatalerror.h"
  60 #include "gromacs/utility/gmxassert.h"
  61 #include "gromacs/utility/stringutil.h"
  62 #include "gromacs/ewald/pme_coordinate_receiver_gpu.h"
  63
  64 #include "pme_gpu_internal.h"
  65 #include "pme_gpu_settings.h"
  66 #include "pme_gpu_timings.h"
  67 #include "pme_gpu_types_host.h"
  68 #include "pme_grid.h"
  69 #include "pme_internal.h"
  70 #include "pme_solve.h"
  71
  72 /*! \brief
  73  * Finds out if PME is currently running on GPU.
  74  *
  75  * \todo The GPU module should not be constructed (or at least called)
  76  * when it is not active, so there should be no need to check whether
  77  * it is active. An assertion that this is true makes sense.
  78  *
  79  * \param[in] pme  The PME structure.
  80  * \returns        True if PME runs on GPU currently, false otherwise.
  81  */
  82 static inline bool pme_gpu_active(const gmx_pme_t* pme)
  83 {
  84     return (pme != nullptr) && (pme->runMode != PmeRunMode::CPU);
  85 }
  86
  87 void pme_gpu_reset_timings(const gmx_pme_t* pme)
  88 {
  89     if (pme_gpu_active(pme))
  90     {
  91         pme_gpu_reset_timings(pme->gpu);
  92     }
  93 }
  94
  95 void pme_gpu_get_timings(const gmx_pme_t* pme, gmx_wallclock_gpu_pme_t* timings)
  96 {
  97     if (pme_gpu_active(pme))
  98     {
  99         pme_gpu_get_timings(pme->gpu, timings);
 100     }
 101 }
 102
 103 int pme_gpu_get_block_size(const gmx_pme_t* pme)
 104 {
 105
 106     if (!pme || !pme_gpu_active(pme))
 107     {
 108         return 0;
 109     }
 110     else
 111     {
 112         return pme_gpu_get_atom_data_block_size();
 113     }
 114 }
 115
 116 /*! \brief
 117  * A convenience wrapper for launching either the GPU or CPU FFT.
 118  *
 119  * \param[in] pme            The PME structure.
 120  * \param[in] gridIndex      The grid index - should currently always be 0.
 121  * \param[in] dir            The FFT direction enum.
 122  * \param[in] wcycle         The wallclock counter.
 123  */
 124 void inline parallel_3dfft_execute_gpu_wrapper(gmx_pme_t*             pme,
 125                                                const int              gridIndex,
 126                                                enum gmx_fft_direction dir,
 127                                                gmx_wallcycle*         wcycle)
 128 {
 129     if (pme_gpu_settings(pme->gpu).performGPUFFT)
 130     {
 131         wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
 132         wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuPme);
 133         pme_gpu_3dfft(pme->gpu, dir, gridIndex);
 134         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuPme);
 135         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
 136     }
 137     else
 138     {
 139         wallcycle_start(wcycle, WallCycleCounter::PmeFftMixedMode);
 140 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
 141         for (int thread = 0; thread < pme->nthread; thread++)
 142         {
 143             gmx_parallel_3dfft_execute(pme->pfft_setup[gridIndex], dir, thread, wcycle);
 144         }
 145         wallcycle_stop(wcycle, WallCycleCounter::PmeFftMixedMode);
 146     }
 147 }
 148
 149 /* The PME computation code split into a few separate functions. */
 150
 151 void pme_gpu_prepare_computation(gmx_pme_t*               pme,
 152                                  const matrix             box,
 153                                  gmx_wallcycle*           wcycle,
 154                                  const gmx::StepWorkload& stepWork)
 155 {
 156     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 157     GMX_ASSERT(pme->nnodes > 0, "");
 158     GMX_ASSERT(pme->nnodes == 1 || pme->ndecompdim > 0, "");
 159
 160     PmeGpu* pmeGpu = pme->gpu;
 161     // TODO these flags are only here to honor the CPU PME code, and probably should be removed
 162     pmeGpu->settings.useGpuForceReduction = stepWork.useGpuPmeFReduction;
 163
 164     bool shouldUpdateBox = false;
 165     for (int i = 0; i < DIM; ++i)
 166     {
 167         for (int j = 0; j <= i; ++j)
 168         {
 169             shouldUpdateBox |= (pmeGpu->common->previousBox[i][j] != box[i][j]);
 170             pmeGpu->common->previousBox[i][j] = box[i][j];
 171         }
 172     }
 173
 174     if (stepWork.haveDynamicBox || shouldUpdateBox) // || is to make the first computation always update
 175     {
 176         wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
 177         wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuPme);
 178         pme_gpu_update_input_box(pmeGpu, box);
 179         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuPme);
 180         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
 181
 182         if (!pme_gpu_settings(pmeGpu).performGPUSolve)
 183         {
 184             // TODO remove code duplication and add test coverage
 185             matrix scaledBox;
 186             pmeGpu->common->boxScaler->scaleBox(box, scaledBox);
 187             gmx::invertBoxMatrix(scaledBox, pme->recipbox);
 188             pme->boxVolume = scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ];
 189         }
 190     }
 191 }
 192
 193 void pme_gpu_launch_spread(gmx_pme_t*                     pme,
 194                            GpuEventSynchronizer*          xReadyOnDevice,
 195                            gmx_wallcycle*                 wcycle,
 196                            const real                     lambdaQ,
 197                            const bool                     useGpuDirectComm,
 198                            gmx::PmeCoordinateReceiverGpu* pmeCoordinateReceiverGpu)
 199 {
 200     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 201     GMX_ASSERT(!GMX_GPU_CUDA || xReadyOnDevice || !pme->bPPnode,
 202                "Need a valid xReadyOnDevice on PP+PME ranks with CUDA.");
 203     GMX_ASSERT(pme->doCoulomb, "Only Coulomb PME can be run on GPU.");
 204
 205     PmeGpu* pmeGpu = pme->gpu;
 206
 207     GMX_ASSERT(pmeGpu->common->ngrids == 1 || (pmeGpu->common->ngrids == 2 && pme->bFEP_q),
 208                "If not decoupling Coulomb interactions there should only be one FEP grid. If "
 209                "decoupling Coulomb interactions there should be two grids.");
 210
 211     /* PME on GPU can currently manage two grids:
 212      * grid_index=0: Coulomb PME with charges in the normal state or from FEP state A.
 213      * grid_index=1: Coulomb PME with charges from FEP state B.
 214      */
 215     real** fftgrids = pme->fftgrid;
 216     /* Spread the coefficients on a grid */
 217     const bool computeSplines = true;
 218     const bool spreadCharges  = true;
 219     wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
 220     wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuPme);
 221     pme_gpu_spread(
 222             pmeGpu, xReadyOnDevice, fftgrids, computeSplines, spreadCharges, lambdaQ, useGpuDirectComm, pmeCoordinateReceiverGpu);
 223     wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuPme);
 224     wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
 225 }
 226
 227 void pme_gpu_launch_complex_transforms(gmx_pme_t* pme, gmx_wallcycle* wcycle, const gmx::StepWorkload& stepWork)
 228 {
 229     PmeGpu*     pmeGpu   = pme->gpu;
 230     const auto& settings = pmeGpu->settings;
 231     // There's no support for computing energy without virial, or vice versa
 232     const bool computeEnergyAndVirial = stepWork.computeEnergy || stepWork.computeVirial;
 233     if (!settings.performGPUFFT)
 234     {
 235         wallcycle_start(wcycle, WallCycleCounter::WaitGpuPmeSpread);
 236         pme_gpu_sync_spread_grid(pme->gpu);
 237         wallcycle_stop(wcycle, WallCycleCounter::WaitGpuPmeSpread);
 238     }
 239
 240     try
 241     {
 242         /* The 3dffts and the solve are done in a loop to simplify things, even if this means that
 243          * there will be two kernel launches for solve. */
 244         for (int gridIndex = 0; gridIndex < pmeGpu->common->ngrids; gridIndex++)
 245         {
 246             /* do R2C 3D-FFT */
 247             t_complex* cfftgrid = pme->cfftgrid[gridIndex];
 248             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_REAL_TO_COMPLEX, wcycle);
 249
 250             /* solve in k-space for our local cells */
 251             if (settings.performGPUSolve)
 252             {
 253                 const auto gridOrdering =
 254                         settings.useDecomposition ? GridOrdering::YZX : GridOrdering::XYZ;
 255                 wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
 256                 wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuPme);
 257                 pme_gpu_solve(pmeGpu, gridIndex, cfftgrid, gridOrdering, computeEnergyAndVirial);
 258                 wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuPme);
 259                 wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
 260             }
 261             else
 262             {
 263                 wallcycle_start(wcycle, WallCycleCounter::PmeSolveMixedMode);
 264 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
 265                 for (int thread = 0; thread < pme->nthread; thread++)
 266                 {
 267                     solve_pme_yzx(pme, cfftgrid, pme->boxVolume, computeEnergyAndVirial, pme->nthread, thread);
 268                 }
 269                 wallcycle_stop(wcycle, WallCycleCounter::PmeSolveMixedMode);
 270             }
 271
 272             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_COMPLEX_TO_REAL, wcycle);
 273         }
 274     }
 275     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 276 }
 277
 278 void pme_gpu_launch_gather(const gmx_pme_t* pme, gmx_wallcycle gmx_unused* wcycle, const real lambdaQ)
 279 {
 280     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 281
 282     if (!pme_gpu_settings(pme->gpu).performGPUGather)
 283     {
 284         return;
 285     }
 286
 287     wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
 288     wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuPme);
 289
 290     float** fftgrids = pme->fftgrid;
 291     pme_gpu_gather(pme->gpu, fftgrids, lambdaQ);
 292     wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuPme);
 293     wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
 294 }
 295
 296 //! Accumulate the \c forcesToAdd to \c f, using the available threads.
 297 static void sum_forces(gmx::ArrayRef<gmx::RVec> f, gmx::ArrayRef<const gmx::RVec> forceToAdd)
 298 {
 299     const int end = forceToAdd.size();
 300
 301     int gmx_unused nt = gmx_omp_nthreads_get(ModuleMultiThread::Pme);
 302 #pragma omp parallel for num_threads(nt) schedule(static)
 303     for (int i = 0; i < end; i++)
 304     {
 305         f[i] += forceToAdd[i];
 306     }
 307 }
 308
 309 //! Reduce quantities from \c output to \c forceWithVirial and \c enerd.
 310 static void pme_gpu_reduce_outputs(const bool            computeEnergyAndVirial,
 311                                    const PmeOutput&      output,
 312                                    gmx_wallcycle*        wcycle,
 313                                    gmx::ForceWithVirial* forceWithVirial,
 314                                    gmx_enerdata_t*       enerd)
 315 {
 316     wallcycle_start(wcycle, WallCycleCounter::PmeGpuFReduction);
 317     GMX_ASSERT(forceWithVirial, "Invalid force pointer");
 318
 319     if (computeEnergyAndVirial)
 320     {
 321         GMX_ASSERT(enerd, "Invalid energy output manager");
 322         forceWithVirial->addVirialContribution(output.coulombVirial_);
 323         enerd->term[F_COUL_RECIP] += output.coulombEnergy_;
 324         enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Coul] += output.coulombDvdl_;
 325     }
 326     if (output.haveForceOutput_)
 327     {
 328         sum_forces(forceWithVirial->force_, output.forces_);
 329     }
 330     wallcycle_stop(wcycle, WallCycleCounter::PmeGpuFReduction);
 331 }
 332
 333 bool pme_gpu_try_finish_task(gmx_pme_t*               pme,
 334                              const gmx::StepWorkload& stepWork,
 335                              gmx_wallcycle*           wcycle,
 336                              gmx::ForceWithVirial*    forceWithVirial,
 337                              gmx_enerdata_t*          enerd,
 338                              const real               lambdaQ,
 339                              GpuTaskCompletion        completionKind)
 340 {
 341     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 342     GMX_ASSERT(!pme->gpu->settings.useGpuForceReduction,
 343                "GPU force reduction should not be active on the pme_gpu_try_finish_task() path");
 344
 345     // First, if possible, check whether all tasks on the stream have
 346     // completed, and return fast if not. Accumulate to wcycle the
 347     // time needed for that checking, but do not yet record that the
 348     // gather has occured.
 349     bool           needToSynchronize      = true;
 350     constexpr bool c_streamQuerySupported = GMX_GPU_CUDA;
 351
 352     // TODO: implement c_streamQuerySupported with an additional GpuEventSynchronizer per stream (#2521)
 353     if ((completionKind == GpuTaskCompletion::Check) && c_streamQuerySupported)
 354     {
 355         wallcycle_start_nocount(wcycle, WallCycleCounter::WaitGpuPmeGather);
 356         // Query the PME stream for completion of all tasks enqueued and
 357         // if we're not done, stop the timer before early return.
 358         const bool pmeGpuDone = pme_gpu_stream_query(pme->gpu);
 359         wallcycle_stop(wcycle, WallCycleCounter::WaitGpuPmeGather);
 360
 361         if (!pmeGpuDone)
 362         {
 363             return false;
 364         }
 365         needToSynchronize = false;
 366     }
 367
 368     wallcycle_start(wcycle, WallCycleCounter::WaitGpuPmeGather);
 369     // If the above check passed, then there is no need to make an
 370     // explicit synchronization call.
 371     if (needToSynchronize)
 372     {
 373         // Synchronize the whole PME stream at once, including D2H result transfers.
 374         pme_gpu_synchronize(pme->gpu);
 375     }
 376     pme_gpu_update_timings(pme->gpu);
 377     // There's no support for computing energy without virial, or vice versa
 378     const bool computeEnergyAndVirial = stepWork.computeEnergy || stepWork.computeVirial;
 379     PmeOutput  output                 = pme_gpu_getOutput(
 380             *pme, computeEnergyAndVirial, pme->gpu->common->ngrids > 1 ? lambdaQ : 1.0);
 381     wallcycle_stop(wcycle, WallCycleCounter::WaitGpuPmeGather);
 382
 383     GMX_ASSERT(pme->gpu->settings.useGpuForceReduction == !output.haveForceOutput_,
 384                "When forces are reduced on the CPU, there needs to be force output");
 385     pme_gpu_reduce_outputs(computeEnergyAndVirial, output, wcycle, forceWithVirial, enerd);
 386
 387     return true;
 388 }
 389
 390 // This is used by PME-only ranks
 391 PmeOutput pme_gpu_wait_finish_task(gmx_pme_t*     pme,
 392                                    const bool     computeEnergyAndVirial,
 393                                    const real     lambdaQ,
 394                                    gmx_wallcycle* wcycle)
 395 {
 396     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 397
 398     wallcycle_start(wcycle, WallCycleCounter::WaitGpuPmeGather);
 399
 400     // Synchronize the whole PME stream at once, including D2H result transfers
 401     // if there are outputs we need to wait for at this step; we still call getOutputs
 402     // for uniformity and because it sets the PmeOutput.haveForceOutput_.
 403     if (!pme->gpu->settings.useGpuForceReduction || computeEnergyAndVirial)
 404     {
 405         pme_gpu_synchronize(pme->gpu);
 406     }
 407
 408     PmeOutput output = pme_gpu_getOutput(
 409             *pme, computeEnergyAndVirial, pme->gpu->common->ngrids > 1 ? lambdaQ : 1.0);
 410     wallcycle_stop(wcycle, WallCycleCounter::WaitGpuPmeGather);
 411     return output;
 412 }
 413
 414 // This is used when not using the alternate-waiting reduction
 415 void pme_gpu_wait_and_reduce(gmx_pme_t*               pme,
 416                              const gmx::StepWorkload& stepWork,
 417                              gmx_wallcycle*           wcycle,
 418                              gmx::ForceWithVirial*    forceWithVirial,
 419                              gmx_enerdata_t*          enerd,
 420                              const real               lambdaQ)
 421 {
 422     // There's no support for computing energy without virial, or vice versa
 423     const bool computeEnergyAndVirial = stepWork.computeEnergy || stepWork.computeVirial;
 424     PmeOutput  output                 = pme_gpu_wait_finish_task(
 425             pme, computeEnergyAndVirial, pme->gpu->common->ngrids > 1 ? lambdaQ : 1.0, wcycle);
 426     GMX_ASSERT(pme->gpu->settings.useGpuForceReduction == !output.haveForceOutput_,
 427                "When forces are reduced on the CPU, there needs to be force output");
 428     pme_gpu_reduce_outputs(computeEnergyAndVirial, output, wcycle, forceWithVirial, enerd);
 429 }
 430
 431 void pme_gpu_reinit_computation(const gmx_pme_t* pme, gmx_wallcycle* wcycle)
 432 {
 433     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 434
 435     wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
 436     wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuPme);
 437
 438     pme_gpu_update_timings(pme->gpu);
 439
 440     pme_gpu_clear_grids(pme->gpu);
 441     pme_gpu_clear_energy_virial(pme->gpu);
 442
 443     wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuPme);
 444     wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
 445 }
 446
 447 DeviceBuffer<gmx::RVec> pme_gpu_get_device_f(const gmx_pme_t* pme)
 448 {
 449     if (!pme || !pme_gpu_active(pme))
 450     {
 451         return DeviceBuffer<gmx::RVec>{};
 452     }
 453     return pme_gpu_get_kernelparam_forces(pme->gpu);
 454 }
 455
 456 void pme_gpu_set_device_x(const gmx_pme_t* pme, DeviceBuffer<gmx::RVec> d_x)
 457 {
 458     GMX_ASSERT(pme != nullptr, "Null pointer is passed as a PME to the set coordinates function.");
 459     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 460
 461     pme_gpu_set_kernelparam_coordinates(pme->gpu, d_x);
 462 }
 463
 464 GpuEventSynchronizer* pme_gpu_get_f_ready_synchronizer(const gmx_pme_t* pme)
 465 {
 466     if (!pme || !pme_gpu_active(pme))
 467     {
 468         return nullptr;
 469     }
 470
 471     return pme_gpu_get_forces_ready_synchronizer(pme->gpu);
 472 }