2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2016,2017,2018,2019,2020,2021, 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.
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.
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.
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.
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.
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.
37 * \brief Implements high-level PME GPU functions which do not require GPU framework-specific code.
39 * \author Aleksei Iupinov <a.yupinov@gmail.com>
40 * \ingroup module_ewald
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/mdtypes/simulation_workload.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 #include "gromacs/ewald/pme_coordinate_receiver_gpu.h"
64 #include "pme_gpu_internal.h"
65 #include "pme_gpu_settings.h"
66 #include "pme_gpu_timings.h"
67 #include "pme_gpu_types_host.h"
69 #include "pme_internal.h"
70 #include "pme_solve.h"
73 * Finds out if PME is currently running on GPU.
75 * \todo The GPU module should not be constructed (or at least called)
76 * when it is not active, so there should be no need to check whether
77 * it is active. An assertion that this is true makes sense.
79 * \param[in] pme The PME structure.
80 * \returns True if PME runs on GPU currently, false otherwise.
82 static inline bool pme_gpu_active(const gmx_pme_t* pme)
84 return (pme != nullptr) && (pme->runMode != PmeRunMode::CPU);
87 void pme_gpu_reset_timings(const gmx_pme_t* pme)
89 if (pme_gpu_active(pme))
91 pme_gpu_reset_timings(pme->gpu);
95 void pme_gpu_get_timings(const gmx_pme_t* pme, gmx_wallclock_gpu_pme_t* timings)
97 if (pme_gpu_active(pme))
99 pme_gpu_get_timings(pme->gpu, timings);
103 int pme_gpu_get_block_size(const gmx_pme_t* pme)
106 if (!pme || !pme_gpu_active(pme))
112 return pme_gpu_get_atom_data_block_size();
117 * A convenience wrapper for launching either the GPU or CPU FFT.
119 * \param[in] pme The PME structure.
120 * \param[in] gridIndex The grid index - should currently always be 0.
121 * \param[in] dir The FFT direction enum.
122 * \param[in] wcycle The wallclock counter.
124 void inline parallel_3dfft_execute_gpu_wrapper(gmx_pme_t* pme,
126 enum gmx_fft_direction dir,
127 gmx_wallcycle* wcycle)
129 if (pme_gpu_settings(pme->gpu).performGPUFFT)
131 wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
132 wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuPme);
133 pme_gpu_3dfft(pme->gpu, dir, gridIndex);
134 wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuPme);
135 wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
139 wallcycle_start(wcycle, WallCycleCounter::PmeFftMixedMode);
140 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
141 for (int thread = 0; thread < pme->nthread; thread++)
143 gmx_parallel_3dfft_execute(pme->pfft_setup[gridIndex], dir, thread, wcycle);
145 wallcycle_stop(wcycle, WallCycleCounter::PmeFftMixedMode);
149 /* The PME computation code split into a few separate functions. */
151 void pme_gpu_prepare_computation(gmx_pme_t* pme,
153 gmx_wallcycle* wcycle,
154 const gmx::StepWorkload& stepWork)
156 GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
157 GMX_ASSERT(pme->nnodes > 0, "");
158 GMX_ASSERT(pme->nnodes == 1 || pme->ndecompdim > 0, "");
160 PmeGpu* pmeGpu = pme->gpu;
161 // TODO these flags are only here to honor the CPU PME code, and probably should be removed
162 pmeGpu->settings.useGpuForceReduction = stepWork.useGpuPmeFReduction;
164 bool shouldUpdateBox = false;
165 for (int i = 0; i < DIM; ++i)
167 for (int j = 0; j <= i; ++j)
169 shouldUpdateBox |= (pmeGpu->common->previousBox[i][j] != box[i][j]);
170 pmeGpu->common->previousBox[i][j] = box[i][j];
174 if (stepWork.haveDynamicBox || shouldUpdateBox) // || is to make the first computation always update
176 wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
177 wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuPme);
178 pme_gpu_update_input_box(pmeGpu, box);
179 wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuPme);
180 wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
182 if (!pme_gpu_settings(pmeGpu).performGPUSolve)
184 // TODO remove code duplication and add test coverage
186 pmeGpu->common->boxScaler->scaleBox(box, scaledBox);
187 gmx::invertBoxMatrix(scaledBox, pme->recipbox);
188 pme->boxVolume = scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ];
193 void pme_gpu_launch_spread(gmx_pme_t* pme,
194 GpuEventSynchronizer* xReadyOnDevice,
195 gmx_wallcycle* wcycle,
197 const bool useGpuDirectComm,
198 gmx::PmeCoordinateReceiverGpu* pmeCoordinateReceiverGpu)
200 GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
201 GMX_ASSERT(!GMX_GPU_CUDA || xReadyOnDevice || !pme->bPPnode,
202 "Need a valid xReadyOnDevice on PP+PME ranks with CUDA.");
203 GMX_ASSERT(pme->doCoulomb, "Only Coulomb PME can be run on GPU.");
205 PmeGpu* pmeGpu = pme->gpu;
207 GMX_ASSERT(pmeGpu->common->ngrids == 1 || (pmeGpu->common->ngrids == 2 && pme->bFEP_q),
208 "If not decoupling Coulomb interactions there should only be one FEP grid. If "
209 "decoupling Coulomb interactions there should be two grids.");
211 /* PME on GPU can currently manage two grids:
212 * grid_index=0: Coulomb PME with charges in the normal state or from FEP state A.
213 * grid_index=1: Coulomb PME with charges from FEP state B.
215 real** fftgrids = pme->fftgrid;
216 /* Spread the coefficients on a grid */
217 const bool computeSplines = true;
218 const bool spreadCharges = true;
219 wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
220 wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuPme);
222 pmeGpu, xReadyOnDevice, fftgrids, computeSplines, spreadCharges, lambdaQ, useGpuDirectComm, pmeCoordinateReceiverGpu);
223 wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuPme);
224 wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
227 void pme_gpu_launch_complex_transforms(gmx_pme_t* pme, gmx_wallcycle* wcycle, const gmx::StepWorkload& stepWork)
229 PmeGpu* pmeGpu = pme->gpu;
230 const auto& settings = pmeGpu->settings;
231 // There's no support for computing energy without virial, or vice versa
232 const bool computeEnergyAndVirial = stepWork.computeEnergy || stepWork.computeVirial;
233 if (!settings.performGPUFFT)
235 wallcycle_start(wcycle, WallCycleCounter::WaitGpuPmeSpread);
236 pme_gpu_sync_spread_grid(pme->gpu);
237 wallcycle_stop(wcycle, WallCycleCounter::WaitGpuPmeSpread);
242 /* The 3dffts and the solve are done in a loop to simplify things, even if this means that
243 * there will be two kernel launches for solve. */
244 for (int gridIndex = 0; gridIndex < pmeGpu->common->ngrids; gridIndex++)
247 t_complex* cfftgrid = pme->cfftgrid[gridIndex];
248 parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_REAL_TO_COMPLEX, wcycle);
250 /* solve in k-space for our local cells */
251 if (settings.performGPUSolve)
253 const auto gridOrdering =
254 settings.useDecomposition ? GridOrdering::YZX : GridOrdering::XYZ;
255 wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
256 wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuPme);
257 pme_gpu_solve(pmeGpu, gridIndex, cfftgrid, gridOrdering, computeEnergyAndVirial);
258 wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuPme);
259 wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
263 wallcycle_start(wcycle, WallCycleCounter::PmeSolveMixedMode);
264 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
265 for (int thread = 0; thread < pme->nthread; thread++)
267 solve_pme_yzx(pme, cfftgrid, pme->boxVolume, computeEnergyAndVirial, pme->nthread, thread);
269 wallcycle_stop(wcycle, WallCycleCounter::PmeSolveMixedMode);
272 parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_COMPLEX_TO_REAL, wcycle);
275 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
278 void pme_gpu_launch_gather(const gmx_pme_t* pme, gmx_wallcycle gmx_unused* wcycle, const real lambdaQ)
280 GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
282 if (!pme_gpu_settings(pme->gpu).performGPUGather)
287 wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
288 wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuPme);
290 float** fftgrids = pme->fftgrid;
291 pme_gpu_gather(pme->gpu, fftgrids, lambdaQ);
292 wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuPme);
293 wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
296 //! Accumulate the \c forcesToAdd to \c f, using the available threads.
297 static void sum_forces(gmx::ArrayRef<gmx::RVec> f, gmx::ArrayRef<const gmx::RVec> forceToAdd)
299 const int end = forceToAdd.size();
301 int gmx_unused nt = gmx_omp_nthreads_get(ModuleMultiThread::Pme);
302 #pragma omp parallel for num_threads(nt) schedule(static)
303 for (int i = 0; i < end; i++)
305 f[i] += forceToAdd[i];
309 //! Reduce quantities from \c output to \c forceWithVirial and \c enerd.
310 static void pme_gpu_reduce_outputs(const bool computeEnergyAndVirial,
311 const PmeOutput& output,
312 gmx_wallcycle* wcycle,
313 gmx::ForceWithVirial* forceWithVirial,
314 gmx_enerdata_t* enerd)
316 wallcycle_start(wcycle, WallCycleCounter::PmeGpuFReduction);
317 GMX_ASSERT(forceWithVirial, "Invalid force pointer");
319 if (computeEnergyAndVirial)
321 GMX_ASSERT(enerd, "Invalid energy output manager");
322 forceWithVirial->addVirialContribution(output.coulombVirial_);
323 enerd->term[F_COUL_RECIP] += output.coulombEnergy_;
324 enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Coul] += output.coulombDvdl_;
326 if (output.haveForceOutput_)
328 sum_forces(forceWithVirial->force_, output.forces_);
330 wallcycle_stop(wcycle, WallCycleCounter::PmeGpuFReduction);
333 bool pme_gpu_try_finish_task(gmx_pme_t* pme,
334 const gmx::StepWorkload& stepWork,
335 gmx_wallcycle* wcycle,
336 gmx::ForceWithVirial* forceWithVirial,
337 gmx_enerdata_t* enerd,
339 GpuTaskCompletion completionKind)
341 GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
342 GMX_ASSERT(!pme->gpu->settings.useGpuForceReduction,
343 "GPU force reduction should not be active on the pme_gpu_try_finish_task() path");
345 // First, if possible, check whether all tasks on the stream have
346 // completed, and return fast if not. Accumulate to wcycle the
347 // time needed for that checking, but do not yet record that the
348 // gather has occured.
349 bool needToSynchronize = true;
350 constexpr bool c_streamQuerySupported = GMX_GPU_CUDA;
352 // TODO: implement c_streamQuerySupported with an additional GpuEventSynchronizer per stream (#2521)
353 if ((completionKind == GpuTaskCompletion::Check) && c_streamQuerySupported)
355 wallcycle_start_nocount(wcycle, WallCycleCounter::WaitGpuPmeGather);
356 // Query the PME stream for completion of all tasks enqueued and
357 // if we're not done, stop the timer before early return.
358 const bool pmeGpuDone = pme_gpu_stream_query(pme->gpu);
359 wallcycle_stop(wcycle, WallCycleCounter::WaitGpuPmeGather);
365 needToSynchronize = false;
368 wallcycle_start(wcycle, WallCycleCounter::WaitGpuPmeGather);
369 // If the above check passed, then there is no need to make an
370 // explicit synchronization call.
371 if (needToSynchronize)
373 // Synchronize the whole PME stream at once, including D2H result transfers.
374 pme_gpu_synchronize(pme->gpu);
376 pme_gpu_update_timings(pme->gpu);
377 // There's no support for computing energy without virial, or vice versa
378 const bool computeEnergyAndVirial = stepWork.computeEnergy || stepWork.computeVirial;
379 PmeOutput output = pme_gpu_getOutput(
380 *pme, computeEnergyAndVirial, pme->gpu->common->ngrids > 1 ? lambdaQ : 1.0);
381 wallcycle_stop(wcycle, WallCycleCounter::WaitGpuPmeGather);
383 GMX_ASSERT(pme->gpu->settings.useGpuForceReduction == !output.haveForceOutput_,
384 "When forces are reduced on the CPU, there needs to be force output");
385 pme_gpu_reduce_outputs(computeEnergyAndVirial, output, wcycle, forceWithVirial, enerd);
390 // This is used by PME-only ranks
391 PmeOutput pme_gpu_wait_finish_task(gmx_pme_t* pme,
392 const bool computeEnergyAndVirial,
394 gmx_wallcycle* wcycle)
396 GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
398 wallcycle_start(wcycle, WallCycleCounter::WaitGpuPmeGather);
400 // Synchronize the whole PME stream at once, including D2H result transfers
401 // if there are outputs we need to wait for at this step; we still call getOutputs
402 // for uniformity and because it sets the PmeOutput.haveForceOutput_.
403 if (!pme->gpu->settings.useGpuForceReduction || computeEnergyAndVirial)
405 pme_gpu_synchronize(pme->gpu);
408 PmeOutput output = pme_gpu_getOutput(
409 *pme, computeEnergyAndVirial, pme->gpu->common->ngrids > 1 ? lambdaQ : 1.0);
410 wallcycle_stop(wcycle, WallCycleCounter::WaitGpuPmeGather);
414 // This is used when not using the alternate-waiting reduction
415 void pme_gpu_wait_and_reduce(gmx_pme_t* pme,
416 const gmx::StepWorkload& stepWork,
417 gmx_wallcycle* wcycle,
418 gmx::ForceWithVirial* forceWithVirial,
419 gmx_enerdata_t* enerd,
422 // There's no support for computing energy without virial, or vice versa
423 const bool computeEnergyAndVirial = stepWork.computeEnergy || stepWork.computeVirial;
424 PmeOutput output = pme_gpu_wait_finish_task(
425 pme, computeEnergyAndVirial, pme->gpu->common->ngrids > 1 ? lambdaQ : 1.0, wcycle);
426 GMX_ASSERT(pme->gpu->settings.useGpuForceReduction == !output.haveForceOutput_,
427 "When forces are reduced on the CPU, there needs to be force output");
428 pme_gpu_reduce_outputs(computeEnergyAndVirial, output, wcycle, forceWithVirial, enerd);
431 void pme_gpu_reinit_computation(const gmx_pme_t* pme, gmx_wallcycle* wcycle)
433 GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
435 wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
436 wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuPme);
438 pme_gpu_update_timings(pme->gpu);
440 pme_gpu_clear_grids(pme->gpu);
441 pme_gpu_clear_energy_virial(pme->gpu);
443 wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuPme);
444 wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
447 DeviceBuffer<gmx::RVec> pme_gpu_get_device_f(const gmx_pme_t* pme)
449 if (!pme || !pme_gpu_active(pme))
451 return DeviceBuffer<gmx::RVec>{};
453 return pme_gpu_get_kernelparam_forces(pme->gpu);
456 void pme_gpu_set_device_x(const gmx_pme_t* pme, DeviceBuffer<gmx::RVec> d_x)
458 GMX_ASSERT(pme != nullptr, "Null pointer is passed as a PME to the set coordinates function.");
459 GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
461 pme_gpu_set_kernelparam_coordinates(pme->gpu, d_x);
464 GpuEventSynchronizer* pme_gpu_get_f_ready_synchronizer(const gmx_pme_t* pme)
466 if (!pme || !pme_gpu_active(pme))
471 return pme_gpu_get_forces_ready_synchronizer(pme->gpu);