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/mdlib/ppforceworkload.h"
  55 #include "gromacs/mdtypes/enerdata.h"
  56 #include "gromacs/mdtypes/forceoutput.h"
  57 #include "gromacs/mdtypes/inputrec.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
  63 #include "pme_gpu_internal.h"
  64 #include "pme_grid.h"
  65 #include "pme_internal.h"
  66 #include "pme_solve.h"
  67
  68 void pme_gpu_reset_timings(const gmx_pme_t *pme)
  69 {
  70     if (pme_gpu_active(pme))
  71     {
  72         pme_gpu_reset_timings(pme->gpu);
  73     }
  74 }
  75
  76 void pme_gpu_get_timings(const gmx_pme_t *pme, gmx_wallclock_gpu_pme_t *timings)
  77 {
  78     if (pme_gpu_active(pme))
  79     {
  80         pme_gpu_get_timings(pme->gpu, timings);
  81     }
  82 }
  83
  84 int pme_gpu_get_padding_size(const gmx_pme_t *pme)
  85 {
  86
  87     if (!pme || !pme_gpu_active(pme))
  88     {
  89         return 0;
  90     }
  91     else
  92     {
  93         return pme_gpu_get_atom_data_alignment(pme->gpu);
  94     }
  95 }
  96
  97 /*! \brief
  98  * A convenience wrapper for launching either the GPU or CPU FFT.
  99  *
 100  * \param[in] pme            The PME structure.
 101  * \param[in] gridIndex      The grid index - should currently always be 0.
 102  * \param[in] dir            The FFT direction enum.
 103  * \param[in] wcycle         The wallclock counter.
 104  */
 105 void inline parallel_3dfft_execute_gpu_wrapper(gmx_pme_t              *pme,
 106                                                const int               gridIndex,
 107                                                enum gmx_fft_direction  dir,
 108                                                gmx_wallcycle_t         wcycle)
 109 {
 110     GMX_ASSERT(gridIndex == 0, "Only single grid supported");
 111     if (pme_gpu_performs_FFT(pme->gpu))
 112     {
 113         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 114         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 115         pme_gpu_3dfft(pme->gpu, dir, gridIndex);
 116         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 117         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 118     }
 119     else
 120     {
 121         wallcycle_start(wcycle, ewcPME_FFT_MIXED_MODE);
 122 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
 123         for (int thread = 0; thread < pme->nthread; thread++)
 124         {
 125             gmx_parallel_3dfft_execute(pme->pfft_setup[gridIndex], dir, thread, wcycle);
 126         }
 127         wallcycle_stop(wcycle, ewcPME_FFT_MIXED_MODE);
 128     }
 129 }
 130
 131 /* The PME computation code split into a few separate functions. */
 132
 133 void pme_gpu_prepare_computation(gmx_pme_t            *pme,
 134                                  bool                  needToUpdateBox,
 135                                  const matrix          box,
 136                                  gmx_wallcycle        *wcycle,
 137                                  int                   flags,
 138                                  bool                  useGpuForceReduction)
 139 {
 140     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 141     GMX_ASSERT(pme->nnodes > 0, "");
 142     GMX_ASSERT(pme->nnodes == 1 || pme->ndecompdim > 0, "");
 143
 144     PmeGpu *pmeGpu = pme->gpu;
 145     pmeGpu->settings.currentFlags         = flags;
 146     // TODO these flags are only here to honor the CPU PME code, and probably should be removed
 147     pmeGpu->settings.useGpuForceReduction = useGpuForceReduction;
 148
 149     bool shouldUpdateBox = false;
 150     for (int i = 0; i < DIM; ++i)
 151     {
 152         for (int j = 0; j <= i; ++j)
 153         {
 154             shouldUpdateBox                  |= (pmeGpu->common->previousBox[i][j] != box[i][j]);
 155             pmeGpu->common->previousBox[i][j] = box[i][j];
 156         }
 157     }
 158
 159     if (needToUpdateBox || shouldUpdateBox) // || is to make the first computation always update
 160     {
 161         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 162         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 163         pme_gpu_update_input_box(pmeGpu, box);
 164         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 165         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 166
 167         if (!pme_gpu_performs_solve(pmeGpu))
 168         {
 169             // TODO remove code duplication and add test coverage
 170             matrix scaledBox;
 171             pmeGpu->common->boxScaler->scaleBox(box, scaledBox);
 172             gmx::invertBoxMatrix(scaledBox, pme->recipbox);
 173             pme->boxVolume = scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ];
 174         }
 175     }
 176 }
 177
 178 void pme_gpu_copy_coordinates_to_gpu(gmx_pme_t            *pme,
 179                                      const rvec           *coordinatesHost,
 180                                      gmx_wallcycle        *wcycle)
 181 {
 182     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 183
 184     PmeGpu *pmeGpu = pme->gpu;
 185
 186     // The only spot of PME GPU where LAUNCH_GPU counter increases call-count
 187     wallcycle_start(wcycle, ewcLAUNCH_GPU);
 188     // The only spot of PME GPU where ewcsLAUNCH_GPU_PME subcounter increases call-count
 189     wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_PME);
 190     pme_gpu_copy_input_coordinates(pmeGpu, coordinatesHost);
 191     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 192     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 193 }
 194
 195 void pme_gpu_launch_spread(gmx_pme_t            *pme,
 196                            gmx_wallcycle        *wcycle)
 197 {
 198     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 199
 200     PmeGpu            *pmeGpu = pme->gpu;
 201
 202     const unsigned int gridIndex  = 0;
 203     real              *fftgrid    = pme->fftgrid[gridIndex];
 204     if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
 205     {
 206         /* Spread the coefficients on a grid */
 207         const bool computeSplines = true;
 208         const bool spreadCharges  = true;
 209         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 210         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 211         pme_gpu_spread(pmeGpu, gridIndex, fftgrid, computeSplines, spreadCharges);
 212         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 213         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 214     }
 215 }
 216
 217 void pme_gpu_launch_complex_transforms(gmx_pme_t      *pme,
 218                                        gmx_wallcycle  *wcycle)
 219 {
 220     PmeGpu            *pmeGpu                 = pme->gpu;
 221     const bool         computeEnergyAndVirial = (pmeGpu->settings.currentFlags & GMX_PME_CALC_ENER_VIR) != 0;
 222     const bool         performBackFFT         = (pmeGpu->settings.currentFlags & (GMX_PME_CALC_F | GMX_PME_CALC_POT)) != 0;
 223     const unsigned int gridIndex              = 0;
 224     t_complex         *cfftgrid               = pme->cfftgrid[gridIndex];
 225
 226     if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
 227     {
 228         if (!pme_gpu_performs_FFT(pmeGpu))
 229         {
 230             wallcycle_start(wcycle, ewcWAIT_GPU_PME_SPREAD);
 231             pme_gpu_sync_spread_grid(pme->gpu);
 232             wallcycle_stop(wcycle, ewcWAIT_GPU_PME_SPREAD);
 233         }
 234     }
 235
 236     try
 237     {
 238         if (pmeGpu->settings.currentFlags & GMX_PME_SOLVE)
 239         {
 240             /* do R2C 3D-FFT */
 241             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_REAL_TO_COMPLEX, wcycle);
 242
 243             /* solve in k-space for our local cells */
 244             if (pme_gpu_performs_solve(pmeGpu))
 245             {
 246                 const auto gridOrdering = pme_gpu_uses_dd(pmeGpu) ? GridOrdering::YZX : GridOrdering::XYZ;
 247                 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 248                 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 249                 pme_gpu_solve(pmeGpu, cfftgrid, gridOrdering, computeEnergyAndVirial);
 250                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 251                 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 252             }
 253             else
 254             {
 255                 wallcycle_start(wcycle, ewcPME_SOLVE_MIXED_MODE);
 256 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
 257                 for (int thread = 0; thread < pme->nthread; thread++)
 258                 {
 259                     solve_pme_yzx(pme, cfftgrid, pme->boxVolume,
 260                                   computeEnergyAndVirial, pme->nthread, thread);
 261                 }
 262                 wallcycle_stop(wcycle, ewcPME_SOLVE_MIXED_MODE);
 263             }
 264         }
 265
 266         if (performBackFFT)
 267         {
 268             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_COMPLEX_TO_REAL, wcycle);
 269         }
 270     } GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 271 }
 272
 273 void pme_gpu_launch_gather(const gmx_pme_t                 *pme,
 274                            gmx_wallcycle gmx_unused        *wcycle,
 275                            PmeForceOutputHandling           forceTreatment)
 276 {
 277     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 278
 279     if (!pme_gpu_performs_gather(pme->gpu))
 280     {
 281         return;
 282     }
 283
 284     wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 285     wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 286     const unsigned int gridIndex  = 0;
 287     real              *fftgrid    = pme->fftgrid[gridIndex];
 288     pme_gpu_gather(pme->gpu, forceTreatment, reinterpret_cast<float *>(fftgrid));
 289     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 290     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 291 }
 292
 293 //! Accumulate the \c forcesToAdd to \c f, using the available threads.
 294 static void sum_forces(gmx::ArrayRef<gmx::RVec>       f,
 295                        gmx::ArrayRef<const gmx::RVec> forceToAdd)
 296 {
 297     const int      end = forceToAdd.size();
 298
 299     int gmx_unused nt = gmx_omp_nthreads_get(emntPME);
 300 #pragma omp parallel for num_threads(nt) schedule(static)
 301     for (int i = 0; i < end; i++)
 302     {
 303         f[i] += forceToAdd[i];
 304     }
 305 }
 306
 307 //! Reduce quantities from \c output to \c forceWithVirial and \c enerd.
 308 static void pme_gpu_reduce_outputs(const int             flags,
 309                                    const PmeOutput      &output,
 310                                    gmx_wallcycle        *wcycle,
 311                                    gmx::ForceWithVirial *forceWithVirial,
 312                                    gmx_enerdata_t       *enerd)
 313 {
 314     wallcycle_start(wcycle, ewcPME_GPU_F_REDUCTION);
 315     GMX_ASSERT(forceWithVirial, "Invalid force pointer");
 316
 317     const bool haveComputedEnergyAndVirial = (flags & GMX_PME_CALC_ENER_VIR) != 0;
 318     if (haveComputedEnergyAndVirial)
 319     {
 320         GMX_ASSERT(enerd, "Invalid energy output manager");
 321         forceWithVirial->addVirialContribution(output.coulombVirial_);
 322         enerd->term[F_COUL_RECIP] += output.coulombEnergy_;
 323     }
 324     if (output.haveForceOutput_)
 325     {
 326         sum_forces(forceWithVirial->force_, output.forces_);
 327     }
 328     wallcycle_stop(wcycle, ewcPME_GPU_F_REDUCTION);
 329 }
 330
 331 bool pme_gpu_try_finish_task(gmx_pme_t            *pme,
 332                              const int             flags,
 333                              gmx_wallcycle        *wcycle,
 334                              gmx::ForceWithVirial *forceWithVirial,
 335                              gmx_enerdata_t       *enerd,
 336                              GpuTaskCompletion     completionKind)
 337 {
 338     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 339     GMX_ASSERT(!pme->gpu->settings.useGpuForceReduction, "GPU force reduction should not be active on the pme_gpu_try_finish_task() path");
 340
 341     // First, if possible, check whether all tasks on the stream have
 342     // completed, and return fast if not. Accumulate to wcycle the
 343     // time needed for that checking, but do not yet record that the
 344     // gather has occured.
 345     bool           needToSynchronize      = true;
 346     constexpr bool c_streamQuerySupported = (GMX_GPU == GMX_GPU_CUDA);
 347     // TODO: implement c_streamQuerySupported with an additional GpuEventSynchronizer per stream (#2521)
 348     if ((completionKind == GpuTaskCompletion::Check) && c_streamQuerySupported)
 349     {
 350         wallcycle_start_nocount(wcycle, ewcWAIT_GPU_PME_GATHER);
 351         // Query the PME stream for completion of all tasks enqueued and
 352         // if we're not done, stop the timer before early return.
 353         const bool pmeGpuDone = pme_gpu_stream_query(pme->gpu);
 354         wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 355
 356         if (!pmeGpuDone)
 357         {
 358             return false;
 359         }
 360         needToSynchronize = false;
 361     }
 362
 363     wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
 364     // If the above check passed, then there is no need to make an
 365     // explicit synchronization call.
 366     if (needToSynchronize)
 367     {
 368         // Synchronize the whole PME stream at once, including D2H result transfers.
 369         pme_gpu_synchronize(pme->gpu);
 370     }
 371     pme_gpu_update_timings(pme->gpu);
 372     PmeOutput output = pme_gpu_getOutput(*pme, flags);
 373     wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 374
 375     GMX_ASSERT(pme->gpu->settings.useGpuForceReduction == !output.haveForceOutput_, "When forces are reduced on the CPU, there needs to be force output");
 376     pme_gpu_reduce_outputs(flags, output, wcycle, forceWithVirial, enerd);
 377
 378     return true;
 379 }
 380
 381 // This is used by PME-only ranks
 382 PmeOutput pme_gpu_wait_finish_task(gmx_pme_t     *pme,
 383                                    const int      flags,
 384                                    gmx_wallcycle *wcycle)
 385 {
 386     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 387
 388     wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
 389
 390     // Synchronize the whole PME stream at once, including D2H result transfers.
 391     //
 392     // TODO: make this sync conditional to wait only for virial/energies
 393     pme_gpu_synchronize(pme->gpu);
 394
 395     pme_gpu_update_timings(pme->gpu);
 396     PmeOutput output = pme_gpu_getOutput(*pme, flags);
 397     wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 398     return output;
 399 }
 400
 401 // This is used when not using the alternate-waiting reduction
 402 void pme_gpu_wait_and_reduce(gmx_pme_t            *pme,
 403                              const int             flags,
 404                              gmx_wallcycle        *wcycle,
 405                              gmx::ForceWithVirial *forceWithVirial,
 406                              gmx_enerdata_t       *enerd)
 407 {
 408     PmeOutput output = pme_gpu_wait_finish_task(pme, flags, wcycle);
 409     GMX_ASSERT(pme->gpu->settings.useGpuForceReduction == !output.haveForceOutput_, "When forces are reduced on the CPU, there needs to be force output");
 410     pme_gpu_reduce_outputs(flags, output, wcycle, forceWithVirial, enerd);
 411 }
 412
 413 void pme_gpu_reinit_computation(const gmx_pme_t *pme,
 414                                 gmx_wallcycle   *wcycle)
 415 {
 416     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 417
 418     wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 419     wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 420
 421     pme_gpu_clear_grids(pme->gpu);
 422     pme_gpu_clear_energy_virial(pme->gpu);
 423
 424     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 425     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 426 }
 427
 428 DeviceBuffer<float> pme_gpu_get_device_x(const gmx_pme_t *pme)
 429 {
 430     GMX_ASSERT((pme && pme_gpu_active(pme)), "PME GPU coordinates buffer was requested from uninitialized PME module");
 431     return pme_gpu_get_kernelparam_coordinates(pme->gpu);
 432 }
 433
 434 void *pme_gpu_get_device_f(const gmx_pme_t *pme)
 435 {
 436     if (!pme || !pme_gpu_active(pme))
 437     {
 438         return nullptr;
 439     }
 440     return pme_gpu_get_kernelparam_forces(pme->gpu);
 441 }
 442
 443 void *pme_gpu_get_device_stream(const gmx_pme_t *pme)
 444 {
 445     if (!pme || !pme_gpu_active(pme))
 446     {
 447         return nullptr;
 448     }
 449     return pme_gpu_get_stream(pme->gpu);
 450 }
 451
 452 GpuEventSynchronizer * pme_gpu_get_f_ready_synchronizer(const gmx_pme_t *pme)
 453 {
 454     if (!pme || !pme_gpu_active(pme))
 455     {
 456         return nullptr;
 457     }
 458
 459     return pme_gpu_get_forces_ready_synchronizer(pme->gpu);
 460 }