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 int pme_gpu_get_padding_size(const gmx_pme_t *pme)
  84 {
  85
  86     if (!pme || !pme_gpu_active(pme))
  87     {
  88         return 0;
  89     }
  90     else
  91     {
  92         return pme_gpu_get_atom_data_alignment(pme->gpu);
  93     }
  94 }
  95
  96 /*! \brief
  97  * A convenience wrapper for launching either the GPU or CPU FFT.
  98  *
  99  * \param[in] pme            The PME structure.
 100  * \param[in] gridIndex      The grid index - should currently always be 0.
 101  * \param[in] dir            The FFT direction enum.
 102  * \param[in] wcycle         The wallclock counter.
 103  */
 104 void inline parallel_3dfft_execute_gpu_wrapper(gmx_pme_t              *pme,
 105                                                const int               gridIndex,
 106                                                enum gmx_fft_direction  dir,
 107                                                gmx_wallcycle_t         wcycle)
 108 {
 109     GMX_ASSERT(gridIndex == 0, "Only single grid supported");
 110     if (pme_gpu_performs_FFT(pme->gpu))
 111     {
 112         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 113         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 114         pme_gpu_3dfft(pme->gpu, dir, gridIndex);
 115         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 116         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 117     }
 118     else
 119     {
 120         wallcycle_start(wcycle, ewcPME_FFT_MIXED_MODE);
 121 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
 122         for (int thread = 0; thread < pme->nthread; thread++)
 123         {
 124             gmx_parallel_3dfft_execute(pme->pfft_setup[gridIndex], dir, thread, wcycle);
 125         }
 126         wallcycle_stop(wcycle, ewcPME_FFT_MIXED_MODE);
 127     }
 128 }
 129
 130 /* The PME computation code split into a few separate functions. */
 131
 132 void pme_gpu_prepare_computation(gmx_pme_t            *pme,
 133                                  bool                  needToUpdateBox,
 134                                  const matrix          box,
 135                                  gmx_wallcycle        *wcycle,
 136                                  int                   flags)
 137 {
 138     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 139     GMX_ASSERT(pme->nnodes > 0, "");
 140     GMX_ASSERT(pme->nnodes == 1 || pme->ndecompdim > 0, "");
 141
 142     PmeGpu *pmeGpu = pme->gpu;
 143     pmeGpu->settings.currentFlags = flags;
 144     // TODO these flags are only here to honor the CPU PME code, and probably should be removed
 145
 146     bool shouldUpdateBox = false;
 147     for (int i = 0; i < DIM; ++i)
 148     {
 149         for (int j = 0; j <= i; ++j)
 150         {
 151             shouldUpdateBox                  |= (pmeGpu->common->previousBox[i][j] != box[i][j]);
 152             pmeGpu->common->previousBox[i][j] = box[i][j];
 153         }
 154     }
 155
 156     if (needToUpdateBox || shouldUpdateBox) // || is to make the first computation always update
 157     {
 158         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 159         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 160         pme_gpu_update_input_box(pmeGpu, box);
 161         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 162         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 163
 164         if (!pme_gpu_performs_solve(pmeGpu))
 165         {
 166             // TODO remove code duplication and add test coverage
 167             matrix scaledBox;
 168             pmeGpu->common->boxScaler->scaleBox(box, scaledBox);
 169             gmx::invertBoxMatrix(scaledBox, pme->recipbox);
 170             pme->boxVolume = scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ];
 171         }
 172     }
 173 }
 174
 175 void pme_gpu_copy_coordinates_to_gpu(gmx_pme_t            *pme,
 176                                      const rvec           *coordinatesHost,
 177                                      gmx_wallcycle        *wcycle)
 178 {
 179     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 180
 181     PmeGpu *pmeGpu = pme->gpu;
 182
 183     // The only spot of PME GPU where LAUNCH_GPU counter increases call-count
 184     wallcycle_start(wcycle, ewcLAUNCH_GPU);
 185     // The only spot of PME GPU where ewcsLAUNCH_GPU_PME subcounter increases call-count
 186     wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_PME);
 187     pme_gpu_copy_input_coordinates(pmeGpu, coordinatesHost);
 188     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 189     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 190 }
 191
 192 void pme_gpu_launch_spread(gmx_pme_t            *pme,
 193                            gmx_wallcycle        *wcycle)
 194 {
 195     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 196
 197     PmeGpu            *pmeGpu = pme->gpu;
 198
 199     const unsigned int gridIndex  = 0;
 200     real              *fftgrid    = pme->fftgrid[gridIndex];
 201     if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
 202     {
 203         /* Spread the coefficients on a grid */
 204         const bool computeSplines = true;
 205         const bool spreadCharges  = true;
 206         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 207         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 208         pme_gpu_spread(pmeGpu, gridIndex, fftgrid, computeSplines, spreadCharges);
 209         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 210         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 211     }
 212 }
 213
 214 void pme_gpu_launch_complex_transforms(gmx_pme_t      *pme,
 215                                        gmx_wallcycle  *wcycle)
 216 {
 217     PmeGpu            *pmeGpu                 = pme->gpu;
 218     const bool         computeEnergyAndVirial = (pmeGpu->settings.currentFlags & GMX_PME_CALC_ENER_VIR) != 0;
 219     const bool         performBackFFT         = (pmeGpu->settings.currentFlags & (GMX_PME_CALC_F | GMX_PME_CALC_POT)) != 0;
 220     const unsigned int gridIndex              = 0;
 221     t_complex         *cfftgrid               = pme->cfftgrid[gridIndex];
 222
 223     if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
 224     {
 225         if (!pme_gpu_performs_FFT(pmeGpu))
 226         {
 227             wallcycle_start(wcycle, ewcWAIT_GPU_PME_SPREAD);
 228             pme_gpu_sync_spread_grid(pme->gpu);
 229             wallcycle_stop(wcycle, ewcWAIT_GPU_PME_SPREAD);
 230         }
 231     }
 232
 233     try
 234     {
 235         if (pmeGpu->settings.currentFlags & GMX_PME_SOLVE)
 236         {
 237             /* do R2C 3D-FFT */
 238             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_REAL_TO_COMPLEX, wcycle);
 239
 240             /* solve in k-space for our local cells */
 241             if (pme_gpu_performs_solve(pmeGpu))
 242             {
 243                 const auto gridOrdering = pme_gpu_uses_dd(pmeGpu) ? GridOrdering::YZX : GridOrdering::XYZ;
 244                 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 245                 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 246                 pme_gpu_solve(pmeGpu, cfftgrid, gridOrdering, computeEnergyAndVirial);
 247                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 248                 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 249             }
 250             else
 251             {
 252                 wallcycle_start(wcycle, ewcPME_SOLVE_MIXED_MODE);
 253 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
 254                 for (int thread = 0; thread < pme->nthread; thread++)
 255                 {
 256                     solve_pme_yzx(pme, cfftgrid, pme->boxVolume,
 257                                   computeEnergyAndVirial, pme->nthread, thread);
 258                 }
 259                 wallcycle_stop(wcycle, ewcPME_SOLVE_MIXED_MODE);
 260             }
 261         }
 262
 263         if (performBackFFT)
 264         {
 265             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_COMPLEX_TO_REAL, wcycle);
 266         }
 267     } GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 268 }
 269
 270 void pme_gpu_launch_gather(const gmx_pme_t                 *pme,
 271                            gmx_wallcycle gmx_unused        *wcycle,
 272                            PmeForceOutputHandling           forceTreatment,
 273                            bool                             useGpuFPmeReduction)
 274 {
 275     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 276
 277     if (!pme_gpu_performs_gather(pme->gpu))
 278     {
 279         return;
 280     }
 281
 282     wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 283     wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 284     const unsigned int gridIndex  = 0;
 285     real              *fftgrid    = pme->fftgrid[gridIndex];
 286     pme_gpu_gather(pme->gpu, forceTreatment, reinterpret_cast<float *>(fftgrid), useGpuFPmeReduction);
 287     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 288     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 289 }
 290
 291 //! Accumulate the \c forcesToAdd to \c f, using the available threads.
 292 static void sum_forces(gmx::ArrayRef<gmx::RVec>       f,
 293                        gmx::ArrayRef<const gmx::RVec> forceToAdd)
 294 {
 295     const int      end = forceToAdd.size();
 296
 297     int gmx_unused nt = gmx_omp_nthreads_get(emntPME);
 298 #pragma omp parallel for num_threads(nt) schedule(static)
 299     for (int i = 0; i < end; i++)
 300     {
 301         f[i] += forceToAdd[i];
 302     }
 303 }
 304
 305 //! Reduce quantities from \c output to \c forceWithVirial and \c enerd.
 306 static void pme_gpu_reduce_outputs(const int             flags,
 307                                    const PmeOutput      &output,
 308                                    gmx_wallcycle        *wcycle,
 309                                    gmx::ForceWithVirial *forceWithVirial,
 310                                    gmx_enerdata_t       *enerd,
 311                                    bool                  useGpuFPmeReduction)
 312 {
 313     wallcycle_start(wcycle, ewcPME_GPU_F_REDUCTION);
 314     GMX_ASSERT(forceWithVirial, "Invalid force pointer");
 315     const bool haveComputedEnergyAndVirial = (flags & GMX_PME_CALC_ENER_VIR) != 0;
 316     if (haveComputedEnergyAndVirial)
 317     {
 318         GMX_ASSERT(enerd, "Invalid energy output manager");
 319         forceWithVirial->addVirialContribution(output.coulombVirial_);
 320         enerd->term[F_COUL_RECIP] += output.coulombEnergy_;
 321     }
 322     if (!useGpuFPmeReduction)
 323     {
 324         sum_forces(forceWithVirial->force_, output.forces_);
 325     }
 326     wallcycle_stop(wcycle, ewcPME_GPU_F_REDUCTION);
 327 }
 328
 329 bool pme_gpu_try_finish_task(gmx_pme_t            *pme,
 330                              const int             flags,
 331                              gmx_wallcycle        *wcycle,
 332                              gmx::ForceWithVirial *forceWithVirial,
 333                              gmx_enerdata_t       *enerd,
 334                              GpuTaskCompletion     completionKind)
 335 {
 336     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 337
 338     // First, if possible, check whether all tasks on the stream have
 339     // completed, and return fast if not. Accumulate to wcycle the
 340     // time needed for that checking, but do not yet record that the
 341     // gather has occured.
 342     bool           needToSynchronize      = true;
 343     constexpr bool c_streamQuerySupported = (GMX_GPU == GMX_GPU_CUDA);
 344     // TODO: implement c_streamQuerySupported with an additional GpuEventSynchronizer per stream (#2521)
 345     if ((completionKind == GpuTaskCompletion::Check) && c_streamQuerySupported)
 346     {
 347         wallcycle_start_nocount(wcycle, ewcWAIT_GPU_PME_GATHER);
 348         // Query the PME stream for completion of all tasks enqueued and
 349         // if we're not done, stop the timer before early return.
 350         const bool pmeGpuDone = pme_gpu_stream_query(pme->gpu);
 351         wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 352
 353         if (!pmeGpuDone)
 354         {
 355             return false;
 356         }
 357         needToSynchronize = false;
 358     }
 359
 360     wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
 361     // If the above check passed, then there is no need to make an
 362     // explicit synchronization call.
 363     if (needToSynchronize)
 364     {
 365         // Synchronize the whole PME stream at once, including D2H result transfers.
 366         pme_gpu_synchronize(pme->gpu);
 367     }
 368     pme_gpu_update_timings(pme->gpu);
 369     PmeOutput output = pme_gpu_getOutput(*pme, flags);
 370     wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 371
 372     pme_gpu_reduce_outputs(flags, output, wcycle, forceWithVirial, enerd, false);
 373
 374     return true;
 375 }
 376
 377 // This is used by PME-only ranks
 378 PmeOutput pme_gpu_wait_finish_task(gmx_pme_t     *pme,
 379                                    const int      flags,
 380                                    gmx_wallcycle *wcycle)
 381 {
 382     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 383
 384     wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
 385
 386     // Synchronize the whole PME stream at once, including D2H result transfers.
 387     // TODO: make this sync conditional with useGpuFPmeReduction to wait only for virial/energies
 388     pme_gpu_synchronize(pme->gpu);
 389
 390     pme_gpu_update_timings(pme->gpu);
 391     PmeOutput output = pme_gpu_getOutput(*pme, flags);
 392     wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 393     return output;
 394 }
 395
 396 // This is used when not using the alternate-waiting reduction
 397 void pme_gpu_wait_and_reduce(gmx_pme_t            *pme,
 398                              const int             flags,
 399                              gmx_wallcycle        *wcycle,
 400                              gmx::ForceWithVirial *forceWithVirial,
 401                              gmx_enerdata_t       *enerd,
 402                              bool                  useGpuFPmeReduction)
 403 {
 404     PmeOutput output = pme_gpu_wait_finish_task(pme, flags, wcycle);
 405     pme_gpu_reduce_outputs(flags, output, wcycle, forceWithVirial, enerd, useGpuFPmeReduction);
 406 }
 407
 408 void pme_gpu_reinit_computation(const gmx_pme_t *pme,
 409                                 gmx_wallcycle   *wcycle)
 410 {
 411     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 412
 413     wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 414     wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 415
 416     pme_gpu_clear_grids(pme->gpu);
 417     pme_gpu_clear_energy_virial(pme->gpu);
 418
 419     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 420     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 421 }
 422
 423 DeviceBuffer<float> pme_gpu_get_device_x(const gmx_pme_t *pme)
 424 {
 425     GMX_ASSERT((pme && pme_gpu_active(pme)), "PME GPU coordinates buffer was requested from uninitialized PME module");
 426     return pme_gpu_get_kernelparam_coordinates(pme->gpu);
 427 }
 428
 429 void *pme_gpu_get_device_f(const gmx_pme_t *pme)
 430 {
 431     if (!pme || !pme_gpu_active(pme))
 432     {
 433         return nullptr;
 434     }
 435     return pme_gpu_get_kernelparam_forces(pme->gpu);
 436 }
 437
 438 void *pme_gpu_get_device_stream(const gmx_pme_t *pme)
 439 {
 440     if (!pme || !pme_gpu_active(pme))
 441     {
 442         return nullptr;
 443     }
 444     return pme_gpu_get_stream(pme->gpu);
 445 }
 446
 447 GpuEventSynchronizer * pme_gpu_get_f_ready_synchronizer(const gmx_pme_t *pme)
 448 {
 449     if (!pme || !pme_gpu_active(pme))
 450     {
 451         return nullptr;
 452     }
 453
 454     return pme_gpu_get_forces_ready_synchronizer(pme->gpu);
 455 }