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                                  bool                  useGpuForceReduction)
 138 {
 139     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 140     GMX_ASSERT(pme->nnodes > 0, "");
 141     GMX_ASSERT(pme->nnodes == 1 || pme->ndecompdim > 0, "");
 142
 143     PmeGpu *pmeGpu = pme->gpu;
 144     pmeGpu->settings.currentFlags         = flags;
 145     // TODO these flags are only here to honor the CPU PME code, and probably should be removed
 146     pmeGpu->settings.useGpuForceReduction = useGpuForceReduction;
 147
 148     bool shouldUpdateBox = false;
 149     for (int i = 0; i < DIM; ++i)
 150     {
 151         for (int j = 0; j <= i; ++j)
 152         {
 153             shouldUpdateBox                  |= (pmeGpu->common->previousBox[i][j] != box[i][j]);
 154             pmeGpu->common->previousBox[i][j] = box[i][j];
 155         }
 156     }
 157
 158     if (needToUpdateBox || shouldUpdateBox) // || is to make the first computation always update
 159     {
 160         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 161         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 162         pme_gpu_update_input_box(pmeGpu, box);
 163         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 164         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 165
 166         if (!pme_gpu_performs_solve(pmeGpu))
 167         {
 168             // TODO remove code duplication and add test coverage
 169             matrix scaledBox;
 170             pmeGpu->common->boxScaler->scaleBox(box, scaledBox);
 171             gmx::invertBoxMatrix(scaledBox, pme->recipbox);
 172             pme->boxVolume = scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ];
 173         }
 174     }
 175 }
 176
 177 void pme_gpu_launch_spread(gmx_pme_t            *pme,
 178                            gmx_wallcycle        *wcycle)
 179 {
 180     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 181
 182     PmeGpu            *pmeGpu = pme->gpu;
 183
 184     const unsigned int gridIndex  = 0;
 185     real              *fftgrid    = pme->fftgrid[gridIndex];
 186     if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
 187     {
 188         /* Spread the coefficients on a grid */
 189         const bool computeSplines = true;
 190         const bool spreadCharges  = true;
 191         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 192         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 193         pme_gpu_spread(pmeGpu, gridIndex, fftgrid, computeSplines, spreadCharges);
 194         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 195         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 196     }
 197 }
 198
 199 void pme_gpu_launch_complex_transforms(gmx_pme_t      *pme,
 200                                        gmx_wallcycle  *wcycle)
 201 {
 202     PmeGpu            *pmeGpu                 = pme->gpu;
 203     const bool         computeEnergyAndVirial = (pmeGpu->settings.currentFlags & GMX_PME_CALC_ENER_VIR) != 0;
 204     const bool         performBackFFT         = (pmeGpu->settings.currentFlags & (GMX_PME_CALC_F | GMX_PME_CALC_POT)) != 0;
 205     const unsigned int gridIndex              = 0;
 206     t_complex         *cfftgrid               = pme->cfftgrid[gridIndex];
 207
 208     if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
 209     {
 210         if (!pme_gpu_performs_FFT(pmeGpu))
 211         {
 212             wallcycle_start(wcycle, ewcWAIT_GPU_PME_SPREAD);
 213             pme_gpu_sync_spread_grid(pme->gpu);
 214             wallcycle_stop(wcycle, ewcWAIT_GPU_PME_SPREAD);
 215         }
 216     }
 217
 218     try
 219     {
 220         if (pmeGpu->settings.currentFlags & GMX_PME_SOLVE)
 221         {
 222             /* do R2C 3D-FFT */
 223             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_REAL_TO_COMPLEX, wcycle);
 224
 225             /* solve in k-space for our local cells */
 226             if (pme_gpu_performs_solve(pmeGpu))
 227             {
 228                 const auto gridOrdering = pme_gpu_uses_dd(pmeGpu) ? GridOrdering::YZX : GridOrdering::XYZ;
 229                 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 230                 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 231                 pme_gpu_solve(pmeGpu, cfftgrid, gridOrdering, computeEnergyAndVirial);
 232                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 233                 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 234             }
 235             else
 236             {
 237                 wallcycle_start(wcycle, ewcPME_SOLVE_MIXED_MODE);
 238 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
 239                 for (int thread = 0; thread < pme->nthread; thread++)
 240                 {
 241                     solve_pme_yzx(pme, cfftgrid, pme->boxVolume,
 242                                   computeEnergyAndVirial, pme->nthread, thread);
 243                 }
 244                 wallcycle_stop(wcycle, ewcPME_SOLVE_MIXED_MODE);
 245             }
 246         }
 247
 248         if (performBackFFT)
 249         {
 250             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_COMPLEX_TO_REAL, wcycle);
 251         }
 252     } GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 253 }
 254
 255 void pme_gpu_launch_gather(const gmx_pme_t                 *pme,
 256                            gmx_wallcycle gmx_unused        *wcycle,
 257                            PmeForceOutputHandling           forceTreatment)
 258 {
 259     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 260
 261     if (!pme_gpu_performs_gather(pme->gpu))
 262     {
 263         return;
 264     }
 265
 266     wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 267     wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 268     const unsigned int gridIndex  = 0;
 269     real              *fftgrid    = pme->fftgrid[gridIndex];
 270     pme_gpu_gather(pme->gpu, forceTreatment, reinterpret_cast<float *>(fftgrid));
 271     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 272     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 273 }
 274
 275 //! Accumulate the \c forcesToAdd to \c f, using the available threads.
 276 static void sum_forces(gmx::ArrayRef<gmx::RVec>       f,
 277                        gmx::ArrayRef<const gmx::RVec> forceToAdd)
 278 {
 279     const int      end = forceToAdd.size();
 280
 281     int gmx_unused nt = gmx_omp_nthreads_get(emntPME);
 282 #pragma omp parallel for num_threads(nt) schedule(static)
 283     for (int i = 0; i < end; i++)
 284     {
 285         f[i] += forceToAdd[i];
 286     }
 287 }
 288
 289 //! Reduce quantities from \c output to \c forceWithVirial and \c enerd.
 290 static void pme_gpu_reduce_outputs(const int             flags,
 291                                    const PmeOutput      &output,
 292                                    gmx_wallcycle        *wcycle,
 293                                    gmx::ForceWithVirial *forceWithVirial,
 294                                    gmx_enerdata_t       *enerd)
 295 {
 296     wallcycle_start(wcycle, ewcPME_GPU_F_REDUCTION);
 297     GMX_ASSERT(forceWithVirial, "Invalid force pointer");
 298
 299     const bool haveComputedEnergyAndVirial = (flags & GMX_PME_CALC_ENER_VIR) != 0;
 300     if (haveComputedEnergyAndVirial)
 301     {
 302         GMX_ASSERT(enerd, "Invalid energy output manager");
 303         forceWithVirial->addVirialContribution(output.coulombVirial_);
 304         enerd->term[F_COUL_RECIP] += output.coulombEnergy_;
 305     }
 306     if (output.haveForceOutput_)
 307     {
 308         sum_forces(forceWithVirial->force_, output.forces_);
 309     }
 310     wallcycle_stop(wcycle, ewcPME_GPU_F_REDUCTION);
 311 }
 312
 313 bool pme_gpu_try_finish_task(gmx_pme_t            *pme,
 314                              const int             flags,
 315                              gmx_wallcycle        *wcycle,
 316                              gmx::ForceWithVirial *forceWithVirial,
 317                              gmx_enerdata_t       *enerd,
 318                              GpuTaskCompletion     completionKind)
 319 {
 320     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 321     GMX_ASSERT(!pme->gpu->settings.useGpuForceReduction, "GPU force reduction should not be active on the pme_gpu_try_finish_task() path");
 322
 323     // First, if possible, check whether all tasks on the stream have
 324     // completed, and return fast if not. Accumulate to wcycle the
 325     // time needed for that checking, but do not yet record that the
 326     // gather has occured.
 327     bool           needToSynchronize      = true;
 328     constexpr bool c_streamQuerySupported = (GMX_GPU == GMX_GPU_CUDA);
 329     // TODO: implement c_streamQuerySupported with an additional GpuEventSynchronizer per stream (#2521)
 330     if ((completionKind == GpuTaskCompletion::Check) && c_streamQuerySupported)
 331     {
 332         wallcycle_start_nocount(wcycle, ewcWAIT_GPU_PME_GATHER);
 333         // Query the PME stream for completion of all tasks enqueued and
 334         // if we're not done, stop the timer before early return.
 335         const bool pmeGpuDone = pme_gpu_stream_query(pme->gpu);
 336         wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 337
 338         if (!pmeGpuDone)
 339         {
 340             return false;
 341         }
 342         needToSynchronize = false;
 343     }
 344
 345     wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
 346     // If the above check passed, then there is no need to make an
 347     // explicit synchronization call.
 348     if (needToSynchronize)
 349     {
 350         // Synchronize the whole PME stream at once, including D2H result transfers.
 351         pme_gpu_synchronize(pme->gpu);
 352     }
 353     pme_gpu_update_timings(pme->gpu);
 354     PmeOutput output = pme_gpu_getOutput(*pme, flags);
 355     wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 356
 357     GMX_ASSERT(pme->gpu->settings.useGpuForceReduction == !output.haveForceOutput_, "When forces are reduced on the CPU, there needs to be force output");
 358     pme_gpu_reduce_outputs(flags, output, wcycle, forceWithVirial, enerd);
 359
 360     return true;
 361 }
 362
 363 // This is used by PME-only ranks
 364 PmeOutput pme_gpu_wait_finish_task(gmx_pme_t     *pme,
 365                                    const int      flags,
 366                                    gmx_wallcycle *wcycle)
 367 {
 368     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 369
 370     wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
 371
 372     // Synchronize the whole PME stream at once, including D2H result transfers
 373     // if there are outputs we need to wait for at this step; we still call getOutputs
 374     // for uniformity and because it sets the PmeOutput.haveForceOutput_.
 375     const bool haveComputedEnergyAndVirial = (flags & GMX_PME_CALC_ENER_VIR) != 0;
 376     if (!pme->gpu->settings.useGpuForceReduction || haveComputedEnergyAndVirial)
 377     {
 378         pme_gpu_synchronize(pme->gpu);
 379     }
 380
 381     PmeOutput output = pme_gpu_getOutput(*pme, flags);
 382     wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 383     return output;
 384 }
 385
 386 // This is used when not using the alternate-waiting reduction
 387 void pme_gpu_wait_and_reduce(gmx_pme_t            *pme,
 388                              const int             flags,
 389                              gmx_wallcycle        *wcycle,
 390                              gmx::ForceWithVirial *forceWithVirial,
 391                              gmx_enerdata_t       *enerd)
 392 {
 393     PmeOutput output = pme_gpu_wait_finish_task(pme, flags, wcycle);
 394     GMX_ASSERT(pme->gpu->settings.useGpuForceReduction == !output.haveForceOutput_, "When forces are reduced on the CPU, there needs to be force output");
 395     pme_gpu_reduce_outputs(flags, output, wcycle, forceWithVirial, enerd);
 396 }
 397
 398 void pme_gpu_reinit_computation(const gmx_pme_t *pme,
 399                                 gmx_wallcycle   *wcycle)
 400 {
 401     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 402
 403     wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 404     wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 405
 406     pme_gpu_update_timings(pme->gpu);
 407
 408     pme_gpu_clear_grids(pme->gpu);
 409     pme_gpu_clear_energy_virial(pme->gpu);
 410
 411     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 412     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 413 }
 414
 415 DeviceBuffer<float> pme_gpu_get_device_x(const gmx_pme_t *pme)
 416 {
 417     GMX_ASSERT((pme && pme_gpu_active(pme)), "PME GPU coordinates buffer was requested from uninitialized PME module");
 418     return pme_gpu_get_kernelparam_coordinates(pme->gpu);
 419 }
 420
 421 void *pme_gpu_get_device_f(const gmx_pme_t *pme)
 422 {
 423     if (!pme || !pme_gpu_active(pme))
 424     {
 425         return nullptr;
 426     }
 427     return pme_gpu_get_kernelparam_forces(pme->gpu);
 428 }
 429
 430 void pme_gpu_set_device_x(const gmx_pme_t     *pme,
 431                           DeviceBuffer<float>  d_x)
 432 {
 433     GMX_ASSERT(pme != nullptr, "Null pointer is passed as a PME to the set coordinates function.");
 434     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 435
 436     pme_gpu_set_kernelparam_coordinates(pme->gpu, d_x);
 437 }
 438
 439 void *pme_gpu_get_device_stream(const gmx_pme_t *pme)
 440 {
 441     if (!pme || !pme_gpu_active(pme))
 442     {
 443         return nullptr;
 444     }
 445     return pme_gpu_get_stream(pme->gpu);
 446 }
 447
 448 void *pme_gpu_get_device_context(const gmx_pme_t *pme)
 449 {
 450     if (!pme || !pme_gpu_active(pme))
 451     {
 452         return nullptr;
 453     }
 454     return pme_gpu_get_context(pme->gpu);
 455 }
 456
 457 GpuEventSynchronizer * pme_gpu_get_f_ready_synchronizer(const gmx_pme_t *pme)
 458 {
 459     if (!pme || !pme_gpu_active(pme))
 460     {
 461         return nullptr;
 462     }
 463
 464     return pme_gpu_get_forces_ready_synchronizer(pme->gpu);
 465 }