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, 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/utility/exceptions.h"
  58 #include "gromacs/utility/fatalerror.h"
  59 #include "gromacs/utility/gmxassert.h"
  60 #include "gromacs/utility/stringutil.h"
  61
  62 #include "pme_gpu_internal.h"
  63 #include "pme_grid.h"
  64 #include "pme_internal.h"
  65 #include "pme_solve.h"
  66
  67 void pme_gpu_reset_timings(const gmx_pme_t *pme)
  68 {
  69     if (pme_gpu_active(pme))
  70     {
  71         pme_gpu_reset_timings(pme->gpu);
  72     }
  73 }
  74
  75 void pme_gpu_get_timings(const gmx_pme_t *pme, gmx_wallclock_gpu_pme_t *timings)
  76 {
  77     if (pme_gpu_active(pme))
  78     {
  79         pme_gpu_get_timings(pme->gpu, timings);
  80     }
  81 }
  82
  83 /*! \brief
  84  * A convenience wrapper for launching either the GPU or CPU FFT.
  85  *
  86  * \param[in] pme            The PME structure.
  87  * \param[in] gridIndex      The grid index - should currently always be 0.
  88  * \param[in] dir            The FFT direction enum.
  89  * \param[in] wcycle         The wallclock counter.
  90  */
  91 void inline parallel_3dfft_execute_gpu_wrapper(gmx_pme_t              *pme,
  92                                                const int               gridIndex,
  93                                                enum gmx_fft_direction  dir,
  94                                                gmx_wallcycle_t         wcycle)
  95 {
  96     GMX_ASSERT(gridIndex == 0, "Only single grid supported");
  97     if (pme_gpu_performs_FFT(pme->gpu))
  98     {
  99         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 100         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 101         pme_gpu_3dfft(pme->gpu, dir, gridIndex);
 102         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 103         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 104     }
 105     else
 106     {
 107         wallcycle_start(wcycle, ewcPME_FFT_MIXED_MODE);
 108 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
 109         for (int thread = 0; thread < pme->nthread; thread++)
 110         {
 111             gmx_parallel_3dfft_execute(pme->pfft_setup[gridIndex], dir, thread, wcycle);
 112         }
 113         wallcycle_stop(wcycle, ewcPME_FFT_MIXED_MODE);
 114     }
 115 }
 116
 117 /* The PME computation code split into a few separate functions. */
 118
 119 void pme_gpu_prepare_computation(gmx_pme_t            *pme,
 120                                  bool                  needToUpdateBox,
 121                                  const matrix          box,
 122                                  gmx_wallcycle        *wcycle,
 123                                  int                   flags)
 124 {
 125     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 126     GMX_ASSERT(pme->nnodes > 0, "");
 127     GMX_ASSERT(pme->nnodes == 1 || pme->ndecompdim > 0, "");
 128
 129     PmeGpu *pmeGpu = pme->gpu;
 130     pmeGpu->settings.currentFlags = flags;
 131     // TODO these flags are only here to honor the CPU PME code, and probably should be removed
 132
 133     bool shouldUpdateBox = false;
 134     for (int i = 0; i < DIM; ++i)
 135     {
 136         for (int j = 0; j <= i; ++j)
 137         {
 138             shouldUpdateBox                  |= (pmeGpu->common->previousBox[i][j] != box[i][j]);
 139             pmeGpu->common->previousBox[i][j] = box[i][j];
 140         }
 141     }
 142
 143     if (needToUpdateBox || shouldUpdateBox) // || is to make the first computation always update
 144     {
 145         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 146         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 147         pme_gpu_update_input_box(pmeGpu, box);
 148         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 149         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 150
 151         if (!pme_gpu_performs_solve(pmeGpu))
 152         {
 153             // TODO remove code duplication and add test coverage
 154             matrix scaledBox;
 155             pmeGpu->common->boxScaler->scaleBox(box, scaledBox);
 156             gmx::invertBoxMatrix(scaledBox, pme->recipbox);
 157             pme->boxVolume = scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ];
 158         }
 159     }
 160 }
 161
 162
 163 void pme_gpu_launch_spread(gmx_pme_t            *pme,
 164                            const rvec           *x,
 165                            gmx_wallcycle        *wcycle)
 166 {
 167     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 168
 169     PmeGpu *pmeGpu = pme->gpu;
 170
 171     // The only spot of PME GPU where LAUNCH_GPU counter increases call-count
 172     wallcycle_start(wcycle, ewcLAUNCH_GPU);
 173     // The only spot of PME GPU where ewcsLAUNCH_GPU_PME subcounter increases call-count
 174     wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_PME);
 175     pme_gpu_copy_input_coordinates(pmeGpu, x);
 176     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 177     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 178
 179     const unsigned int gridIndex  = 0;
 180     real              *fftgrid    = pme->fftgrid[gridIndex];
 181     if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
 182     {
 183         /* Spread the coefficients on a grid */
 184         const bool computeSplines = true;
 185         const bool spreadCharges  = true;
 186         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 187         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 188         pme_gpu_spread(pmeGpu, gridIndex, fftgrid, computeSplines, spreadCharges);
 189         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 190         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 191     }
 192 }
 193
 194 void pme_gpu_launch_complex_transforms(gmx_pme_t      *pme,
 195                                        gmx_wallcycle  *wcycle)
 196 {
 197     PmeGpu            *pmeGpu                 = pme->gpu;
 198     const bool         computeEnergyAndVirial = (pmeGpu->settings.currentFlags & GMX_PME_CALC_ENER_VIR) != 0;
 199     const bool         performBackFFT         = (pmeGpu->settings.currentFlags & (GMX_PME_CALC_F | GMX_PME_CALC_POT)) != 0;
 200     const unsigned int gridIndex              = 0;
 201     t_complex         *cfftgrid               = pme->cfftgrid[gridIndex];
 202
 203     if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
 204     {
 205         if (!pme_gpu_performs_FFT(pmeGpu))
 206         {
 207             wallcycle_start(wcycle, ewcWAIT_GPU_PME_SPREAD);
 208             pme_gpu_sync_spread_grid(pme->gpu);
 209             wallcycle_stop(wcycle, ewcWAIT_GPU_PME_SPREAD);
 210         }
 211     }
 212
 213     try
 214     {
 215         if (pmeGpu->settings.currentFlags & GMX_PME_SOLVE)
 216         {
 217             /* do R2C 3D-FFT */
 218             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_REAL_TO_COMPLEX, wcycle);
 219
 220             /* solve in k-space for our local cells */
 221             if (pme_gpu_performs_solve(pmeGpu))
 222             {
 223                 const auto gridOrdering = pme_gpu_uses_dd(pmeGpu) ? GridOrdering::YZX : GridOrdering::XYZ;
 224                 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 225                 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 226                 pme_gpu_solve(pmeGpu, cfftgrid, gridOrdering, computeEnergyAndVirial);
 227                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 228                 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 229             }
 230             else
 231             {
 232                 wallcycle_start(wcycle, ewcPME_SOLVE_MIXED_MODE);
 233 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
 234                 for (int thread = 0; thread < pme->nthread; thread++)
 235                 {
 236                     solve_pme_yzx(pme, cfftgrid, pme->boxVolume,
 237                                   computeEnergyAndVirial, pme->nthread, thread);
 238                 }
 239                 wallcycle_stop(wcycle, ewcPME_SOLVE_MIXED_MODE);
 240             }
 241         }
 242
 243         if (performBackFFT)
 244         {
 245             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_COMPLEX_TO_REAL, wcycle);
 246         }
 247     } GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 248 }
 249
 250 void pme_gpu_launch_gather(const gmx_pme_t                 *pme,
 251                            gmx_wallcycle gmx_unused        *wcycle,
 252                            PmeForceOutputHandling           forceTreatment,
 253                            bool                             useGpuFPmeReduction)
 254 {
 255     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 256
 257     if (!pme_gpu_performs_gather(pme->gpu))
 258     {
 259         return;
 260     }
 261
 262     wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 263     wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 264     const unsigned int gridIndex  = 0;
 265     real              *fftgrid    = pme->fftgrid[gridIndex];
 266     pme_gpu_gather(pme->gpu, forceTreatment, reinterpret_cast<float *>(fftgrid), useGpuFPmeReduction);
 267     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 268     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 269 }
 270
 271 //! Accumulate the \c forcesToAdd to \c f, using the available threads.
 272 static void sum_forces(gmx::ArrayRef<gmx::RVec>       f,
 273                        gmx::ArrayRef<const gmx::RVec> forceToAdd)
 274 {
 275     const int      end = forceToAdd.size();
 276
 277     int gmx_unused nt = gmx_omp_nthreads_get(emntPME);
 278 #pragma omp parallel for num_threads(nt) schedule(static)
 279     for (int i = 0; i < end; i++)
 280     {
 281         f[i] += forceToAdd[i];
 282     }
 283 }
 284
 285 //! Reduce quantities from \c output to \c forceWithVirial and \c enerd.
 286 static void pme_gpu_reduce_outputs(const int             flags,
 287                                    const PmeOutput      &output,
 288                                    gmx_wallcycle        *wcycle,
 289                                    gmx::ForceWithVirial *forceWithVirial,
 290                                    gmx_enerdata_t       *enerd,
 291                                    bool                  useGpuFPmeReduction)
 292 {
 293     wallcycle_start(wcycle, ewcPME_GPU_F_REDUCTION);
 294     GMX_ASSERT(forceWithVirial, "Invalid force pointer");
 295     const bool haveComputedEnergyAndVirial = (flags & GMX_PME_CALC_ENER_VIR) != 0;
 296     if (haveComputedEnergyAndVirial)
 297     {
 298         GMX_ASSERT(enerd, "Invalid energy output manager");
 299         forceWithVirial->addVirialContribution(output.coulombVirial_);
 300         enerd->term[F_COUL_RECIP] += output.coulombEnergy_;
 301     }
 302     if (!useGpuFPmeReduction)
 303     {
 304         sum_forces(forceWithVirial->force_, output.forces_);
 305     }
 306     wallcycle_stop(wcycle, ewcPME_GPU_F_REDUCTION);
 307 }
 308
 309 bool pme_gpu_try_finish_task(gmx_pme_t            *pme,
 310                              const int             flags,
 311                              gmx_wallcycle        *wcycle,
 312                              gmx::ForceWithVirial *forceWithVirial,
 313                              gmx_enerdata_t       *enerd,
 314                              GpuTaskCompletion     completionKind)
 315 {
 316     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 317
 318     // First, if possible, check whether all tasks on the stream have
 319     // completed, and return fast if not. Accumulate to wcycle the
 320     // time needed for that checking, but do not yet record that the
 321     // gather has occured.
 322     bool           needToSynchronize      = true;
 323     constexpr bool c_streamQuerySupported = (GMX_GPU == GMX_GPU_CUDA);
 324     // TODO: implement c_streamQuerySupported with an additional GpuEventSynchronizer per stream (#2521)
 325     if ((completionKind == GpuTaskCompletion::Check) && c_streamQuerySupported)
 326     {
 327         wallcycle_start_nocount(wcycle, ewcWAIT_GPU_PME_GATHER);
 328         // Query the PME stream for completion of all tasks enqueued and
 329         // if we're not done, stop the timer before early return.
 330         const bool pmeGpuDone = pme_gpu_stream_query(pme->gpu);
 331         wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 332
 333         if (!pmeGpuDone)
 334         {
 335             return false;
 336         }
 337         needToSynchronize = false;
 338     }
 339
 340     wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
 341     // If the above check passed, then there is no need to make an
 342     // explicit synchronization call.
 343     if (needToSynchronize)
 344     {
 345         // Synchronize the whole PME stream at once, including D2H result transfers.
 346         pme_gpu_synchronize(pme->gpu);
 347     }
 348     pme_gpu_update_timings(pme->gpu);
 349     PmeOutput output = pme_gpu_getOutput(*pme, flags);
 350     wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 351
 352     pme_gpu_reduce_outputs(flags, output, wcycle, forceWithVirial, enerd, false);
 353
 354     return true;
 355 }
 356
 357 // This is used by PME-only ranks
 358 PmeOutput pme_gpu_wait_finish_task(gmx_pme_t     *pme,
 359                                    const int      flags,
 360                                    gmx_wallcycle *wcycle)
 361 {
 362     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 363
 364     wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
 365
 366     // Synchronize the whole PME stream at once, including D2H result transfers.
 367     // TODO: make this sync conditional with useGpuFPmeReduction to wait only for virial/energies
 368     pme_gpu_synchronize(pme->gpu);
 369
 370     pme_gpu_update_timings(pme->gpu);
 371     PmeOutput output = pme_gpu_getOutput(*pme, flags);
 372     wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 373     return output;
 374 }
 375
 376 // This is used when not using the alternate-waiting reduction
 377 void pme_gpu_wait_and_reduce(gmx_pme_t            *pme,
 378                              const int             flags,
 379                              gmx_wallcycle        *wcycle,
 380                              gmx::ForceWithVirial *forceWithVirial,
 381                              gmx_enerdata_t       *enerd,
 382                              bool                  useGpuFPmeReduction)
 383 {
 384     PmeOutput output = pme_gpu_wait_finish_task(pme, flags, wcycle);
 385     pme_gpu_reduce_outputs(flags, output, wcycle, forceWithVirial, enerd, useGpuFPmeReduction);
 386 }
 387
 388 void pme_gpu_reinit_computation(const gmx_pme_t *pme,
 389                                 gmx_wallcycle   *wcycle)
 390 {
 391     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 392
 393     wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 394     wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 395
 396     pme_gpu_clear_grids(pme->gpu);
 397     pme_gpu_clear_energy_virial(pme->gpu);
 398
 399     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 400     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 401 }
 402
 403 void *pme_gpu_get_device_x(const gmx_pme_t *pme)
 404 {
 405     if (!pme || !pme_gpu_active(pme))
 406     {
 407         return nullptr;
 408     }
 409     return pme_gpu_get_kernelparam_coordinates(pme->gpu);
 410 }
 411
 412 void *pme_gpu_get_device_f(const gmx_pme_t *pme)
 413 {
 414     if (!pme || !pme_gpu_active(pme))
 415     {
 416         return nullptr;
 417     }
 418     return pme_gpu_get_kernelparam_forces(pme->gpu);
 419 }
 420
 421 GpuEventSynchronizer * pme_gpu_get_f_ready_synchronizer(const gmx_pme_t *pme)
 422 {
 423     if (!pme || !pme_gpu_active(pme))
 424     {
 425         return nullptr;
 426     }
 427
 428     return pme_gpu_get_forces_ready_synchronizer(pme->gpu);
 429 }