src/gromacs/ewald/pme_gpu.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2016,2017,2018,2019,2020, 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_gpu_settings.h"
  64 #include "pme_gpu_timings.h"
  65 #include "pme_gpu_types_host.h"
  66 #include "pme_grid.h"
  67 #include "pme_internal.h"
  68 #include "pme_solve.h"
  69
  70 /*! \brief
  71  * Finds out if PME is currently running on GPU.
  72  *
  73  * \todo The GPU module should not be constructed (or at least called)
  74  * when it is not active, so there should be no need to check whether
  75  * it is active. An assertion that this is true makes sense.
  76  *
  77  * \param[in] pme  The PME structure.
  78  * \returns        True if PME runs on GPU currently, false otherwise.
  79  */
  80 static inline bool pme_gpu_active(const gmx_pme_t* pme)
  81 {
  82     return (pme != nullptr) && (pme->runMode != PmeRunMode::CPU);
  83 }
  84
  85 void pme_gpu_reset_timings(const gmx_pme_t* pme)
  86 {
  87     if (pme_gpu_active(pme))
  88     {
  89         pme_gpu_reset_timings(pme->gpu);
  90     }
  91 }
  92
  93 void pme_gpu_get_timings(const gmx_pme_t* pme, gmx_wallclock_gpu_pme_t* timings)
  94 {
  95     if (pme_gpu_active(pme))
  96     {
  97         pme_gpu_get_timings(pme->gpu, timings);
  98     }
  99 }
 100
 101 int pme_gpu_get_padding_size(const gmx_pme_t* pme)
 102 {
 103
 104     if (!pme || !pme_gpu_active(pme))
 105     {
 106         return 0;
 107     }
 108     else
 109     {
 110         return pme_gpu_get_atom_data_alignment(pme->gpu);
 111     }
 112 }
 113
 114 /*! \brief
 115  * A convenience wrapper for launching either the GPU or CPU FFT.
 116  *
 117  * \param[in] pme            The PME structure.
 118  * \param[in] gridIndex      The grid index - should currently always be 0.
 119  * \param[in] dir            The FFT direction enum.
 120  * \param[in] wcycle         The wallclock counter.
 121  */
 122 void inline parallel_3dfft_execute_gpu_wrapper(gmx_pme_t*             pme,
 123                                                const int              gridIndex,
 124                                                enum gmx_fft_direction dir,
 125                                                gmx_wallcycle_t        wcycle)
 126 {
 127     GMX_ASSERT(gridIndex == 0, "Only single grid supported");
 128     if (pme_gpu_settings(pme->gpu).performGPUFFT)
 129     {
 130         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 131         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 132         pme_gpu_3dfft(pme->gpu, dir, gridIndex);
 133         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 134         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 135     }
 136     else
 137     {
 138         wallcycle_start(wcycle, ewcPME_FFT_MIXED_MODE);
 139 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
 140         for (int thread = 0; thread < pme->nthread; thread++)
 141         {
 142             gmx_parallel_3dfft_execute(pme->pfft_setup[gridIndex], dir, thread, wcycle);
 143         }
 144         wallcycle_stop(wcycle, ewcPME_FFT_MIXED_MODE);
 145     }
 146 }
 147
 148 /* The PME computation code split into a few separate functions. */
 149
 150 void pme_gpu_prepare_computation(gmx_pme_t*     pme,
 151                                  bool           needToUpdateBox,
 152                                  const matrix   box,
 153                                  gmx_wallcycle* wcycle,
 154                                  int            flags,
 155                                  bool           useGpuForceReduction)
 156 {
 157     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 158     GMX_ASSERT(pme->nnodes > 0, "");
 159     GMX_ASSERT(pme->nnodes == 1 || pme->ndecompdim > 0, "");
 160
 161     PmeGpu* pmeGpu                = pme->gpu;
 162     pmeGpu->settings.currentFlags = flags;
 163     // TODO these flags are only here to honor the CPU PME code, and probably should be removed
 164     pmeGpu->settings.useGpuForceReduction = useGpuForceReduction;
 165
 166     bool shouldUpdateBox = false;
 167     for (int i = 0; i < DIM; ++i)
 168     {
 169         for (int j = 0; j <= i; ++j)
 170         {
 171             shouldUpdateBox |= (pmeGpu->common->previousBox[i][j] != box[i][j]);
 172             pmeGpu->common->previousBox[i][j] = box[i][j];
 173         }
 174     }
 175
 176     if (needToUpdateBox || shouldUpdateBox) // || is to make the first computation always update
 177     {
 178         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 179         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 180         pme_gpu_update_input_box(pmeGpu, box);
 181         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 182         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 183
 184         if (!pme_gpu_settings(pmeGpu).performGPUSolve)
 185         {
 186             // TODO remove code duplication and add test coverage
 187             matrix scaledBox;
 188             pmeGpu->common->boxScaler->scaleBox(box, scaledBox);
 189             gmx::invertBoxMatrix(scaledBox, pme->recipbox);
 190             pme->boxVolume = scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ];
 191         }
 192     }
 193 }
 194
 195 void pme_gpu_launch_spread(gmx_pme_t* pme, GpuEventSynchronizer* xReadyOnDevice, gmx_wallcycle* wcycle)
 196 {
 197     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 198     GMX_ASSERT(xReadyOnDevice || !pme->bPPnode || (GMX_GPU != GMX_GPU_CUDA),
 199                "Need a valid xReadyOnDevice on PP+PME ranks with CUDA.");
 200
 201     PmeGpu* pmeGpu = pme->gpu;
 202
 203     const unsigned int gridIndex = 0;
 204     real*              fftgrid   = pme->fftgrid[gridIndex];
 205     if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
 206     {
 207         /* Spread the coefficients on a grid */
 208         const bool computeSplines = true;
 209         const bool spreadCharges  = true;
 210         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 211         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 212         pme_gpu_spread(pmeGpu, xReadyOnDevice, gridIndex, fftgrid, computeSplines, spreadCharges);
 213         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 214         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 215     }
 216 }
 217
 218 void pme_gpu_launch_complex_transforms(gmx_pme_t* pme, gmx_wallcycle* wcycle)
 219 {
 220     PmeGpu*     pmeGpu                 = pme->gpu;
 221     const auto& settings               = pmeGpu->settings;
 222     const bool  computeEnergyAndVirial = (settings.currentFlags & GMX_PME_CALC_ENER_VIR) != 0;
 223     const bool  performBackFFT = (settings.currentFlags & (GMX_PME_CALC_F | GMX_PME_CALC_POT)) != 0;
 224     const unsigned int gridIndex = 0;
 225     t_complex*         cfftgrid  = pme->cfftgrid[gridIndex];
 226
 227     if (settings.currentFlags & GMX_PME_SPREAD)
 228     {
 229         if (!settings.performGPUFFT)
 230         {
 231             wallcycle_start(wcycle, ewcWAIT_GPU_PME_SPREAD);
 232             pme_gpu_sync_spread_grid(pme->gpu);
 233             wallcycle_stop(wcycle, ewcWAIT_GPU_PME_SPREAD);
 234         }
 235     }
 236
 237     try
 238     {
 239         if (settings.currentFlags & GMX_PME_SOLVE)
 240         {
 241             /* do R2C 3D-FFT */
 242             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_REAL_TO_COMPLEX, wcycle);
 243
 244             /* solve in k-space for our local cells */
 245             if (settings.performGPUSolve)
 246             {
 247                 // TODO grid ordering should be set up at pme init time.
 248                 const auto gridOrdering =
 249                         settings.useDecomposition ? GridOrdering::YZX : GridOrdering::XYZ;
 250                 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 251                 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 252                 pme_gpu_solve(pmeGpu, cfftgrid, gridOrdering, computeEnergyAndVirial);
 253                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 254                 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 255             }
 256             else
 257             {
 258                 wallcycle_start(wcycle, ewcPME_SOLVE_MIXED_MODE);
 259 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
 260                 for (int thread = 0; thread < pme->nthread; thread++)
 261                 {
 262                     solve_pme_yzx(pme, cfftgrid, pme->boxVolume, computeEnergyAndVirial,
 263                                   pme->nthread, thread);
 264                 }
 265                 wallcycle_stop(wcycle, ewcPME_SOLVE_MIXED_MODE);
 266             }
 267         }
 268
 269         if (performBackFFT)
 270         {
 271             parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_COMPLEX_TO_REAL, wcycle);
 272         }
 273     }
 274     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 275 }
 276
 277 void pme_gpu_launch_gather(const gmx_pme_t* pme, gmx_wallcycle gmx_unused* wcycle, PmeForceOutputHandling forceTreatment)
 278 {
 279     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 280
 281     if (!pme_gpu_settings(pme->gpu).performGPUGather)
 282     {
 283         return;
 284     }
 285
 286     wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 287     wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 288     const unsigned int gridIndex = 0;
 289     real*              fftgrid   = pme->fftgrid[gridIndex];
 290     pme_gpu_gather(pme->gpu, forceTreatment, reinterpret_cast<float*>(fftgrid));
 291     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 292     wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 293 }
 294
 295 //! Accumulate the \c forcesToAdd to \c f, using the available threads.
 296 static void sum_forces(gmx::ArrayRef<gmx::RVec> f, gmx::ArrayRef<const gmx::RVec> forceToAdd)
 297 {
 298     const int end = forceToAdd.size();
 299
 300     int gmx_unused nt = gmx_omp_nthreads_get(emntPME);
 301 #pragma omp parallel for num_threads(nt) schedule(static)
 302     for (int i = 0; i < end; i++)
 303     {
 304         f[i] += forceToAdd[i];
 305     }
 306 }
 307
 308 //! Reduce quantities from \c output to \c forceWithVirial and \c enerd.
 309 static void pme_gpu_reduce_outputs(const int             flags,
 310                                    const PmeOutput&      output,
 311                                    gmx_wallcycle*        wcycle,
 312                                    gmx::ForceWithVirial* forceWithVirial,
 313                                    gmx_enerdata_t*       enerd)
 314 {
 315     wallcycle_start(wcycle, ewcPME_GPU_F_REDUCTION);
 316     GMX_ASSERT(forceWithVirial, "Invalid force pointer");
 317
 318     const bool haveComputedEnergyAndVirial = (flags & GMX_PME_CALC_ENER_VIR) != 0;
 319     if (haveComputedEnergyAndVirial)
 320     {
 321         GMX_ASSERT(enerd, "Invalid energy output manager");
 322         forceWithVirial->addVirialContribution(output.coulombVirial_);
 323         enerd->term[F_COUL_RECIP] += output.coulombEnergy_;
 324     }
 325     if (output.haveForceOutput_)
 326     {
 327         sum_forces(forceWithVirial->force_, output.forces_);
 328     }
 329     wallcycle_stop(wcycle, ewcPME_GPU_F_REDUCTION);
 330 }
 331
 332 bool pme_gpu_try_finish_task(gmx_pme_t*            pme,
 333                              const int             flags,
 334                              gmx_wallcycle*        wcycle,
 335                              gmx::ForceWithVirial* forceWithVirial,
 336                              gmx_enerdata_t*       enerd,
 337                              GpuTaskCompletion     completionKind)
 338 {
 339     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 340     GMX_ASSERT(!pme->gpu->settings.useGpuForceReduction,
 341                "GPU force reduction should not be active on the pme_gpu_try_finish_task() path");
 342
 343     // First, if possible, check whether all tasks on the stream have
 344     // completed, and return fast if not. Accumulate to wcycle the
 345     // time needed for that checking, but do not yet record that the
 346     // gather has occured.
 347     bool           needToSynchronize      = true;
 348     constexpr bool c_streamQuerySupported = (GMX_GPU == GMX_GPU_CUDA);
 349     // TODO: implement c_streamQuerySupported with an additional GpuEventSynchronizer per stream (#2521)
 350     if ((completionKind == GpuTaskCompletion::Check) && c_streamQuerySupported)
 351     {
 352         wallcycle_start_nocount(wcycle, ewcWAIT_GPU_PME_GATHER);
 353         // Query the PME stream for completion of all tasks enqueued and
 354         // if we're not done, stop the timer before early return.
 355         const bool pmeGpuDone = pme_gpu_stream_query(pme->gpu);
 356         wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 357
 358         if (!pmeGpuDone)
 359         {
 360             return false;
 361         }
 362         needToSynchronize = false;
 363     }
 364
 365     wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
 366     // If the above check passed, then there is no need to make an
 367     // explicit synchronization call.
 368     if (needToSynchronize)
 369     {
 370         // Synchronize the whole PME stream at once, including D2H result transfers.
 371         pme_gpu_synchronize(pme->gpu);
 372     }
 373     pme_gpu_update_timings(pme->gpu);
 374     PmeOutput output = pme_gpu_getOutput(*pme, flags);
 375     wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 376
 377     GMX_ASSERT(pme->gpu->settings.useGpuForceReduction == !output.haveForceOutput_,
 378                "When forces are reduced on the CPU, there needs to be force output");
 379     pme_gpu_reduce_outputs(flags, output, wcycle, forceWithVirial, enerd);
 380
 381     return true;
 382 }
 383
 384 // This is used by PME-only ranks
 385 PmeOutput pme_gpu_wait_finish_task(gmx_pme_t* pme, const int flags, gmx_wallcycle* wcycle)
 386 {
 387     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 388
 389     wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
 390
 391     // Synchronize the whole PME stream at once, including D2H result transfers
 392     // if there are outputs we need to wait for at this step; we still call getOutputs
 393     // for uniformity and because it sets the PmeOutput.haveForceOutput_.
 394     const bool haveComputedEnergyAndVirial = (flags & GMX_PME_CALC_ENER_VIR) != 0;
 395     if (!pme->gpu->settings.useGpuForceReduction || haveComputedEnergyAndVirial)
 396     {
 397         pme_gpu_synchronize(pme->gpu);
 398     }
 399
 400     PmeOutput output = pme_gpu_getOutput(*pme, flags);
 401     wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
 402     return output;
 403 }
 404
 405 // This is used when not using the alternate-waiting reduction
 406 void pme_gpu_wait_and_reduce(gmx_pme_t*            pme,
 407                              const int             flags,
 408                              gmx_wallcycle*        wcycle,
 409                              gmx::ForceWithVirial* forceWithVirial,
 410                              gmx_enerdata_t*       enerd)
 411 {
 412     PmeOutput output = pme_gpu_wait_finish_task(pme, flags, wcycle);
 413     GMX_ASSERT(pme->gpu->settings.useGpuForceReduction == !output.haveForceOutput_,
 414                "When forces are reduced on the CPU, there needs to be force output");
 415     pme_gpu_reduce_outputs(flags, output, wcycle, forceWithVirial, enerd);
 416 }
 417
 418 void pme_gpu_reinit_computation(const gmx_pme_t* pme, gmx_wallcycle* wcycle)
 419 {
 420     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 421
 422     wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 423     wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
 424
 425     pme_gpu_update_timings(pme->gpu);
 426
 427     pme_gpu_clear_grids(pme->gpu);
 428     pme_gpu_clear_energy_virial(pme->gpu);
 429
 430     wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
 431     wallcycle_stop(wcycle, ewcLAUNCH_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_set_device_x(const gmx_pme_t* pme, DeviceBuffer<float> d_x)
 444 {
 445     GMX_ASSERT(pme != nullptr, "Null pointer is passed as a PME to the set coordinates function.");
 446     GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
 447
 448     pme_gpu_set_kernelparam_coordinates(pme->gpu, d_x);
 449 }
 450
 451 void* pme_gpu_get_device_stream(const gmx_pme_t* pme)
 452 {
 453     if (!pme || !pme_gpu_active(pme))
 454     {
 455         return nullptr;
 456     }
 457     return pme_gpu_get_stream(pme->gpu);
 458 }
 459
 460 void* pme_gpu_get_device_context(const gmx_pme_t* pme)
 461 {
 462     if (!pme || !pme_gpu_active(pme))
 463     {
 464         return nullptr;
 465     }
 466     return pme_gpu_get_context(pme->gpu);
 467 }
 468
 469 GpuEventSynchronizer* pme_gpu_get_f_ready_synchronizer(const gmx_pme_t* pme)
 470 {
 471     if (!pme || !pme_gpu_active(pme))
 472     {
 473         return nullptr;
 474     }
 475
 476     return pme_gpu_get_forces_ready_synchronizer(pme->gpu);
 477 }