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                            GpuEventSynchronizer *xReadyOnDevice,
 179                            gmx_wallcycle        *wcycle)
 180 {
 181     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 182
 183     PmeGpu            *pmeGpu = pme->gpu;
 184
 185     const unsigned int gridIndex  = 0;
 186     real              *fftgrid    = pme->fftgrid[gridIndex];
 187     if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
 188     {
 189         /* Spread the coefficients on a grid */
 190         const bool computeSplines = true;
 191         const bool spreadCharges  = true;
 192         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 193         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 194         pme_gpu_spread(pmeGpu, xReadyOnDevice, gridIndex, fftgrid, computeSplines, spreadCharges);
 195         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 196         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 197     }
 198 }
 199
 200 void pme_gpu_launch_complex_transforms(gmx_pme_t      *pme,
 201                                        gmx_wallcycle  *wcycle)
 202 {
 203     PmeGpu            *pmeGpu                 = pme->gpu;
 204     const bool         computeEnergyAndVirial = (pmeGpu->settings.currentFlags & GMX_PME_CALC_ENER_VIR) != 0;
 205     const bool         performBackFFT         = (pmeGpu->settings.currentFlags & (GMX_PME_CALC_F | GMX_PME_CALC_POT)) != 0;
 206     const unsigned int gridIndex              = 0;
 207     t_complex         *cfftgrid               = pme->cfftgrid[gridIndex];
 208
 209     if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
 210     {
 211         if (!pme_gpu_performs_FFT(pmeGpu))
 212         {
 213             wallcycle_start(wcycle, ewcWAIT_GPU_PME_SPREAD);
 214             pme_gpu_sync_spread_grid(pme->gpu);
 215             wallcycle_stop(wcycle, ewcWAIT_GPU_PME_SPREAD);
 216         }
 217     }
 218
 219     try
 220     {
 221         if (pmeGpu->settings.currentFlags & GMX_PME_SOLVE)
 222         {
 223             /* do R2C 3D-FFT */
 224             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_REAL_TO_COMPLEX, wcycle);
 225
 226             /* solve in k-space for our local cells */
 227             if (pme_gpu_performs_solve(pmeGpu))
 228             {
 229                 const auto gridOrdering = pme_gpu_uses_dd(pmeGpu) ? GridOrdering::YZX : GridOrdering::XYZ;
 230                 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 231                 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 232                 pme_gpu_solve(pmeGpu, cfftgrid, gridOrdering, computeEnergyAndVirial);
 233                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 234                 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 235             }
 236             else
 237             {
 238                 wallcycle_start(wcycle, ewcPME_SOLVE_MIXED_MODE);
 239 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
 240                 for (int thread = 0; thread < pme->nthread; thread++)
 241                 {
 242                     solve_pme_yzx(pme, cfftgrid, pme->boxVolume,
 243                                   computeEnergyAndVirial, pme->nthread, thread);
 244                 }
 245                 wallcycle_stop(wcycle, ewcPME_SOLVE_MIXED_MODE);
 246             }
 247         }
 248
 249         if (performBackFFT)
 250         {
 251             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_COMPLEX_TO_REAL, wcycle);
 252         }
 253     } GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 254 }
 255
 256 void pme_gpu_launch_gather(const gmx_pme_t                 *pme,
 257                            gmx_wallcycle gmx_unused        *wcycle,
 258                            PmeForceOutputHandling           forceTreatment)
 259 {
 260     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 261
 262     if (!pme_gpu_performs_gather(pme->gpu))
 263     {
 264         return;
 265     }
 266
 267     wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 268     wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 269     const unsigned int gridIndex  = 0;
 270     real              *fftgrid    = pme->fftgrid[gridIndex];
 271     pme_gpu_gather(pme->gpu, forceTreatment, reinterpret_cast<float *>(fftgrid));
 272     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 273     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 274 }
 275
 276 //! Accumulate the \c forcesToAdd to \c f, using the available threads.
 277 static void sum_forces(gmx::ArrayRef<gmx::RVec>       f,
 278                        gmx::ArrayRef<const gmx::RVec> forceToAdd)
 279 {
 280     const int      end = forceToAdd.size();
 281
 282     int gmx_unused nt = gmx_omp_nthreads_get(emntPME);
 283 #pragma omp parallel for num_threads(nt) schedule(static)
 284     for (int i = 0; i < end; i++)
 285     {
 286         f[i] += forceToAdd[i];
 287     }
 288 }
 289
 290 //! Reduce quantities from \c output to \c forceWithVirial and \c enerd.
 291 static void pme_gpu_reduce_outputs(const int             flags,
 292                                    const PmeOutput      &output,
 293                                    gmx_wallcycle        *wcycle,
 294                                    gmx::ForceWithVirial *forceWithVirial,
 295                                    gmx_enerdata_t       *enerd)
 296 {
 297     wallcycle_start(wcycle, ewcPME_GPU_F_REDUCTION);
 298     GMX_ASSERT(forceWithVirial, "Invalid force pointer");
 299
 300     const bool haveComputedEnergyAndVirial = (flags & GMX_PME_CALC_ENER_VIR) != 0;
 301     if (haveComputedEnergyAndVirial)
 302     {
 303         GMX_ASSERT(enerd, "Invalid energy output manager");
 304         forceWithVirial->addVirialContribution(output.coulombVirial_);
 305         enerd->term[F_COUL_RECIP] += output.coulombEnergy_;
 306     }
 307     if (output.haveForceOutput_)
 308     {
 309         sum_forces(forceWithVirial->force_, output.forces_);
 310     }
 311     wallcycle_stop(wcycle, ewcPME_GPU_F_REDUCTION);
 312 }
 313
 314 bool pme_gpu_try_finish_task(gmx_pme_t            *pme,
 315                              const int             flags,
 316                              gmx_wallcycle        *wcycle,
 317                              gmx::ForceWithVirial *forceWithVirial,
 318                              gmx_enerdata_t       *enerd,
 319                              GpuTaskCompletion     completionKind)
 320 {
 321     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 322     GMX_ASSERT(!pme->gpu->settings.useGpuForceReduction, "GPU force reduction should not be active on the pme_gpu_try_finish_task() path");
 323
 324     // First, if possible, check whether all tasks on the stream have
 325     // completed, and return fast if not. Accumulate to wcycle the
 326     // time needed for that checking, but do not yet record that the
 327     // gather has occured.
 328     bool           needToSynchronize      = true;
 329     constexpr bool c_streamQuerySupported = (GMX_GPU == GMX_GPU_CUDA);
 330     // TODO: implement c_streamQuerySupported with an additional GpuEventSynchronizer per stream (#2521)
 331     if ((completionKind == GpuTaskCompletion::Check) && c_streamQuerySupported)
 332     {
 333         wallcycle_start_nocount(wcycle, ewcWAIT_GPU_PME_GATHER);
 334         // Query the PME stream for completion of all tasks enqueued and
 335         // if we're not done, stop the timer before early return.
 336         const bool pmeGpuDone = pme_gpu_stream_query(pme->gpu);
 337         wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 338
 339         if (!pmeGpuDone)
 340         {
 341             return false;
 342         }
 343         needToSynchronize = false;
 344     }
 345
 346     wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
 347     // If the above check passed, then there is no need to make an
 348     // explicit synchronization call.
 349     if (needToSynchronize)
 350     {
 351         // Synchronize the whole PME stream at once, including D2H result transfers.
 352         pme_gpu_synchronize(pme->gpu);
 353     }
 354     pme_gpu_update_timings(pme->gpu);
 355     PmeOutput output = pme_gpu_getOutput(*pme, flags);
 356     wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 357
 358     GMX_ASSERT(pme->gpu->settings.useGpuForceReduction == !output.haveForceOutput_, "When forces are reduced on the CPU, there needs to be force output");
 359     pme_gpu_reduce_outputs(flags, output, wcycle, forceWithVirial, enerd);
 360
 361     return true;
 362 }
 363
 364 // This is used by PME-only ranks
 365 PmeOutput pme_gpu_wait_finish_task(gmx_pme_t     *pme,
 366                                    const int      flags,
 367                                    gmx_wallcycle *wcycle)
 368 {
 369     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 370
 371     wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
 372
 373     // Synchronize the whole PME stream at once, including D2H result transfers
 374     // if there are outputs we need to wait for at this step; we still call getOutputs
 375     // for uniformity and because it sets the PmeOutput.haveForceOutput_.
 376     const bool haveComputedEnergyAndVirial = (flags & GMX_PME_CALC_ENER_VIR) != 0;
 377     if (!pme->gpu->settings.useGpuForceReduction || haveComputedEnergyAndVirial)
 378     {
 379         pme_gpu_synchronize(pme->gpu);
 380     }
 381
 382     PmeOutput output = pme_gpu_getOutput(*pme, flags);
 383     wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 384     return output;
 385 }
 386
 387 // This is used when not using the alternate-waiting reduction
 388 void pme_gpu_wait_and_reduce(gmx_pme_t            *pme,
 389                              const int             flags,
 390                              gmx_wallcycle        *wcycle,
 391                              gmx::ForceWithVirial *forceWithVirial,
 392                              gmx_enerdata_t       *enerd)
 393 {
 394     PmeOutput output = pme_gpu_wait_finish_task(pme, flags, wcycle);
 395     GMX_ASSERT(pme->gpu->settings.useGpuForceReduction == !output.haveForceOutput_, "When forces are reduced on the CPU, there needs to be force output");
 396     pme_gpu_reduce_outputs(flags, output, wcycle, forceWithVirial, enerd);
 397 }
 398
 399 void pme_gpu_reinit_computation(const gmx_pme_t *pme,
 400                                 gmx_wallcycle   *wcycle)
 401 {
 402     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 403
 404     wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 405     wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 406
 407     pme_gpu_update_timings(pme->gpu);
 408
 409     pme_gpu_clear_grids(pme->gpu);
 410     pme_gpu_clear_energy_virial(pme->gpu);
 411
 412     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 413     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 414 }
 415
 416 DeviceBuffer<float> pme_gpu_get_device_x(const gmx_pme_t *pme)
 417 {
 418     GMX_ASSERT((pme && pme_gpu_active(pme)), "PME GPU coordinates buffer was requested from uninitialized PME module");
 419     return pme_gpu_get_kernelparam_coordinates(pme->gpu);
 420 }
 421
 422 void *pme_gpu_get_device_f(const gmx_pme_t *pme)
 423 {
 424     if (!pme || !pme_gpu_active(pme))
 425     {
 426         return nullptr;
 427     }
 428     return pme_gpu_get_kernelparam_forces(pme->gpu);
 429 }
 430
 431 void pme_gpu_set_device_x(const gmx_pme_t     *pme,
 432                           DeviceBuffer<float>  d_x)
 433 {
 434     GMX_ASSERT(pme != nullptr, "Null pointer is passed as a PME to the set coordinates function.");
 435     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 436
 437     pme_gpu_set_kernelparam_coordinates(pme->gpu, d_x);
 438 }
 439
 440 void *pme_gpu_get_device_stream(const gmx_pme_t *pme)
 441 {
 442     if (!pme || !pme_gpu_active(pme))
 443     {
 444         return nullptr;
 445     }
 446     return pme_gpu_get_stream(pme->gpu);
 447 }
 448
 449 void *pme_gpu_get_device_context(const gmx_pme_t *pme)
 450 {
 451     if (!pme || !pme_gpu_active(pme))
 452     {
 453         return nullptr;
 454     }
 455     return pme_gpu_get_context(pme->gpu);
 456 }
 457
 458 GpuEventSynchronizer * pme_gpu_get_f_ready_synchronizer(const gmx_pme_t *pme)
 459 {
 460     if (!pme || !pme_gpu_active(pme))
 461     {
 462         return nullptr;
 463     }
 464
 465     return pme_gpu_get_forces_ready_synchronizer(pme->gpu);
 466 }