2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2004, The GROMACS development team.
6 * Copyright (c) 2013,2014,2015,2016,2017 by the GROMACS development team.
7 * Copyright (c) 2018,2019,2020,2021, by the GROMACS development team, led by
8 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
9 * and including many others, as listed in the AUTHORS file in the
10 * top-level source directory and at http://www.gromacs.org.
12 * GROMACS is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public License
14 * as published by the Free Software Foundation; either version 2.1
15 * of the License, or (at your option) any later version.
17 * GROMACS is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with GROMACS; if not, see
24 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 * If you want to redistribute modifications to GROMACS, please
28 * consider that scientific software is very special. Version
29 * control is crucial - bugs must be traceable. We will be happy to
30 * consider code for inclusion in the official distribution, but
31 * derived work must not be called official GROMACS. Details are found
32 * in the README & COPYING files - if they are missing, get the
33 * official version at http://www.gromacs.org.
35 * To help us fund GROMACS development, we humbly ask that you cite
36 * the research papers on the package. Check out http://www.gromacs.org.
38 /*! \libinternal \file
40 * \brief This file contains function declarations necessary for
41 * computing energies and forces for the PME long-ranged part (Coulomb
44 * \author Berk Hess <hess@kth.se>
46 * \ingroup module_ewald
49 #ifndef GMX_EWALD_PME_H
50 #define GMX_EWALD_PME_H
55 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
56 #include "gromacs/gpu_utils/gpu_macros.h"
57 #include "gromacs/math/vectypes.h"
58 #include "gromacs/utility/basedefinitions.h"
59 #include "gromacs/utility/real.h"
66 struct gmx_wallclock_gpu_pme_t;
67 struct gmx_enerdata_t;
75 enum class GpuTaskCompletion;
77 class GpuEventSynchronizer;
83 class ForceWithVirial;
85 enum class PinningPolicy : int;
88 /*! \libinternal \brief Class for managing usage of separate PME-only ranks
90 * Used for checking if some parts of the code could not use PME-only ranks
93 class SeparatePmeRanksPermitted
96 //! Disables PME ranks permitted flag with a reason
97 void disablePmeRanks(const std::string& reason);
98 //! Return status of PME ranks usage
99 bool permitSeparatePmeRanks() const;
100 //! Returns all reasons, for not using PME ranks
101 std::string reasonsWhyDisabled() const;
104 //! Flag that informs whether simualtion could use dedicated PME ranks
105 bool permitSeparatePmeRanks_ = true;
106 //! Storage for all reasons, why PME ranks could not be used
107 std::vector<std::string> reasons_;
114 GMX_SUM_GRID_FORWARD,
115 GMX_SUM_GRID_BACKWARD
118 /*! \brief Possible PME codepaths on a rank.
119 * \todo: make this enum class with gmx_pme_t C++ refactoring
121 enum class PmeRunMode
123 None, //!< No PME task is done
124 CPU, //!< Whole PME computation is done on CPU
125 GPU, //!< Whole PME computation is done on GPU
126 Mixed, //!< Mixed mode: only spread and gather run on GPU; FFT and solving are done on CPU.
129 /*! \brief Return the smallest allowed PME grid size for \p pmeOrder */
130 int minimalPmeGridSize(int pmeOrder);
132 //! Return whether the grid of \c pme is identical to \c grid_size.
133 bool gmx_pme_grid_matches(const gmx_pme_t& pme, const ivec grid_size);
135 /*! \brief Check restrictions on pme_order and the PME grid nkx,nky,nkz.
137 * With errorsAreFatal=true, an exception or fatal error is generated
138 * on violation of restrictions.
139 * With errorsAreFatal=false, false is returned on violation of restrictions.
140 * When all restrictions are obeyed, true is returned.
141 * Argument useThreads tells if any MPI rank doing PME uses more than 1 threads.
142 * If at calling useThreads is unknown, pass true for conservative checking.
144 * The PME GPU restrictions are checked separately during pme_gpu_init().
146 bool gmx_pme_check_restrictions(int pme_order,
150 int numPmeDomainsAlongX,
152 bool errorsAreFatal);
154 /*! \brief Construct PME data
156 * \throws gmx::InconsistentInputError if input grid sizes/PME order are inconsistent.
157 * \returns Pointer to newly allocated and initialized PME data.
159 * \todo We should evolve something like a \c GpuManager that holds \c
160 * DeviceInformation* and \c PmeGpuProgram* and perhaps other
161 * related things whose lifetime can/should exceed that of a task (or
162 * perhaps task manager). See Issue #2522.
164 gmx_pme_t* gmx_pme_init(const t_commrec* cr,
165 const NumPmeDomains& numPmeDomains,
166 const t_inputrec* ir,
167 gmx_bool bFreeEnergy_q,
168 gmx_bool bFreeEnergy_lj,
169 gmx_bool bReproducible,
175 const DeviceContext* deviceContext,
176 const DeviceStream* deviceStream,
177 const PmeGpuProgram* pmeGpuProgram,
178 const gmx::MDLogger& mdlog);
180 /*! \brief As gmx_pme_init, but takes most settings, except the grid/Ewald coefficients, from
181 * pme_src. This is only called when the PME cut-off/grid size changes.
183 void gmx_pme_reinit(gmx_pme_t** pmedata,
186 const t_inputrec* ir,
187 const ivec grid_size,
191 /*! \brief Destroys the PME data structure.*/
192 void gmx_pme_destroy(gmx_pme_t* pme);
194 /*! \brief Do a PME calculation on a CPU for the long range electrostatics and/or LJ.
196 * Computes the PME forces and the energy and viral, when requested,
197 * for all atoms in \p coordinates. Forces, when requested, are added
198 * to the buffer \p forces, which is allowed to contain more elements
199 * than the number of elements in \p coordinates.
200 * The meaning of \p flags is defined above, and determines which
201 * parts of the calculation are performed.
203 * \return 0 indicates all well, non zero is an error code.
205 int gmx_pme_do(struct gmx_pme_t* pme,
206 gmx::ArrayRef<const gmx::RVec> coordinates,
207 gmx::ArrayRef<gmx::RVec> forces,
219 gmx_wallcycle* wcycle,
228 const gmx::StepWorkload& stepWork);
230 /*! \brief Calculate the PME grid energy V for n charges.
232 * The potential (found in \p pme) must have been found already with a
233 * call to gmx_pme_do(). Note that the charges are not spread on the grid in the
234 * pme struct. Currently does not work in parallel or with free
237 void gmx_pme_calc_energy(gmx_pme_t* pme, gmx::ArrayRef<const gmx::RVec> x, gmx::ArrayRef<const real> q, real* V);
240 * This function updates the local atom data on GPU after DD (charges, coordinates, etc.).
241 * TODO: it should update the PME CPU atom data as well.
242 * (currently PME CPU call gmx_pme_do() gets passed the input pointers for each computation).
244 * \param[in,out] pme The PME structure.
245 * \param[in] numAtoms The number of particles.
246 * \param[in] chargesA The pointer to the array of particle charges in the normal state or FEP
247 * state A. Can be nullptr if PME is not performed on the GPU.
248 * \param[in] chargesB The pointer to the array of particle charges in state B. Only used if
249 * charges are perturbed and can otherwise be nullptr.
251 void gmx_pme_reinit_atoms(gmx_pme_t* pme, int numAtoms, const real* chargesA, const real* chargesB);
253 /* A block of PME GPU functions */
255 /*! \brief Checks whether the GROMACS build allows to run PME on GPU.
256 * TODO: this partly duplicates an internal PME assert function
257 * pme_gpu_check_restrictions(), except that works with a
258 * formed gmx_pme_t structure. Should that one go away/work with inputrec?
260 * \param[out] error If non-null, the error message when PME is not supported on GPU.
262 * \returns true if PME can run on GPU on this build, false otherwise.
264 bool pme_gpu_supports_build(std::string* error);
266 /*! \brief Checks whether the detected (GPU) hardware allows to run PME on GPU.
268 * \param[in] hwinfo Information about the detected hardware
269 * \param[out] error If non-null, the error message when PME is not supported on GPU.
271 * \returns true if PME can run on GPU on this build, false otherwise.
273 bool pme_gpu_supports_hardware(const gmx_hw_info_t& hwinfo, std::string* error);
275 /*! \brief Checks whether the input system allows to run PME on GPU.
276 * TODO: this partly duplicates an internal PME assert function
277 * pme_gpu_check_restrictions(), except that works with a
278 * formed gmx_pme_t structure. Should that one go away/work with inputrec?
280 * \param[in] ir Input system.
281 * \param[out] error If non-null, the error message if the input is not supported on GPU.
283 * \returns true if PME can run on GPU with this input, false otherwise.
285 bool pme_gpu_supports_input(const t_inputrec& ir, std::string* error);
288 * Returns the active PME codepath (CPU, GPU, mixed).
289 * \todo This is a rather static data that should be managed by the higher level task scheduler.
291 * \param[in] pme The PME data structure.
292 * \returns active PME codepath.
294 PmeRunMode pme_run_mode(const gmx_pme_t* pme);
296 /*! \libinternal \brief
297 * Return the pinning policy appropriate for this build configuration
298 * for relevant buffers used for PME task on this rank (e.g. running
300 gmx::PinningPolicy pme_get_pinning_policy();
303 * Tells if PME is enabled to run on GPU (not necessarily active at the moment).
304 * \todo This is a rather static data that should be managed by the hardware assignment manager.
305 * For now, it is synonymous with the active PME codepath (in the absence of dynamic switching).
307 * \param[in] pme The PME data structure.
308 * \returns true if PME can run on GPU, false otherwise.
310 inline bool pme_gpu_task_enabled(const gmx_pme_t* pme)
312 return (pme != nullptr) && (pme_run_mode(pme) != PmeRunMode::CPU);
315 /*! \brief Returns the block size requirement
317 * The GPU version of PME requires that the coordinates array have a
318 * size divisible by the returned number.
320 * \param[in] pme The PME data structure.
322 GPU_FUNC_QUALIFIER int pme_gpu_get_block_size(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
323 GPU_FUNC_TERM_WITH_RETURN(0);
325 // The following functions are all the PME GPU entry points,
326 // currently inlining to nothing on non-CUDA builds.
329 * Resets the PME GPU timings. To be called at the reset step.
331 * \param[in] pme The PME structure.
333 GPU_FUNC_QUALIFIER void pme_gpu_reset_timings(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme)) GPU_FUNC_TERM;
336 * Copies the PME GPU timings to the gmx_wallclock_gpu_pme_t structure (for log output). To be called at the run end.
338 * \param[in] pme The PME structure.
339 * \param[in] timings The gmx_wallclock_gpu_pme_t structure.
341 GPU_FUNC_QUALIFIER void pme_gpu_get_timings(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
342 gmx_wallclock_gpu_pme_t* GPU_FUNC_ARGUMENT(timings)) GPU_FUNC_TERM;
344 /* The main PME GPU functions */
347 * Prepares PME on GPU computation (updating the box if needed)
348 * \param[in] pme The PME data structure.
349 * \param[in] box The unit cell box.
350 * \param[in] wcycle The wallclock counter.
351 * \param[in] stepWork The required work for this simulation step
353 GPU_FUNC_QUALIFIER void pme_gpu_prepare_computation(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
354 const matrix GPU_FUNC_ARGUMENT(box),
355 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
356 const gmx::StepWorkload& GPU_FUNC_ARGUMENT(stepWork)) GPU_FUNC_TERM;
359 * Launches first stage of PME on GPU - spreading kernel.
361 * \param[in] pme The PME data structure.
362 * \param[in] xReadyOnDevice Event synchronizer indicating that the coordinates
363 * are ready in the device memory; nullptr allowed only on separate PME ranks.
364 * \param[in] wcycle The wallclock counter.
365 * \param[in] lambdaQ The Coulomb lambda of the current state of the
366 * system. Only used if FEP of Coulomb is active.
368 GPU_FUNC_QUALIFIER void pme_gpu_launch_spread(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
369 GpuEventSynchronizer* GPU_FUNC_ARGUMENT(xReadyOnDevice),
370 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
371 real GPU_FUNC_ARGUMENT(lambdaQ)) GPU_FUNC_TERM;
374 * Launches middle stages of PME (FFT R2C, solving, FFT C2R) either on GPU or on CPU, depending on the run mode.
376 * \param[in] pme The PME data structure.
377 * \param[in] wcycle The wallclock counter.
378 * \param[in] stepWork The required work for this simulation step
380 GPU_FUNC_QUALIFIER void
381 pme_gpu_launch_complex_transforms(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
382 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
383 const gmx::StepWorkload& GPU_FUNC_ARGUMENT(stepWork)) GPU_FUNC_TERM;
386 * Launches last stage of PME on GPU - force gathering and D2H force transfer.
388 * \param[in] pme The PME data structure.
389 * \param[in] wcycle The wallclock counter.
390 * \param[in] lambdaQ The Coulomb lambda to use when calculating the results.
392 GPU_FUNC_QUALIFIER void pme_gpu_launch_gather(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
393 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
394 real GPU_FUNC_ARGUMENT(lambdaQ)) GPU_FUNC_TERM;
397 * Attempts to complete PME GPU tasks.
399 * The \p completionKind argument controls whether the function blocks until all
400 * PME GPU tasks enqueued completed (as pme_gpu_wait_finish_task() does) or only
401 * checks and returns immediately if they did not.
402 * When blocking or the tasks have completed it also gets the output forces
403 * by assigning the ArrayRef to the \p forces pointer passed in.
404 * Virial/energy are also outputs if they were to be computed.
406 * \param[in] pme The PME data structure.
407 * \param[in] stepWork The required work for this simulation step
408 * \param[in] wcycle The wallclock counter.
409 * \param[out] forceWithVirial The output force and virial
410 * \param[out] enerd The output energies
411 * \param[in] lambdaQ The Coulomb lambda to use when calculating the results.
412 * \param[in] completionKind Indicates whether PME task completion should only be checked rather
414 * \returns True if the PME GPU tasks have completed
416 GPU_FUNC_QUALIFIER bool pme_gpu_try_finish_task(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
417 const gmx::StepWorkload& GPU_FUNC_ARGUMENT(stepWork),
418 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
419 gmx::ForceWithVirial* GPU_FUNC_ARGUMENT(forceWithVirial),
420 gmx_enerdata_t* GPU_FUNC_ARGUMENT(enerd),
421 real GPU_FUNC_ARGUMENT(lambdaQ),
422 GpuTaskCompletion GPU_FUNC_ARGUMENT(completionKind))
423 GPU_FUNC_TERM_WITH_RETURN(false);
426 * Blocks until PME GPU tasks are completed, and gets the output forces and virial/energy
427 * (if they were to be computed).
429 * \param[in] pme The PME data structure.
430 * \param[in] stepWork The required work for this simulation step
431 * \param[in] wcycle The wallclock counter.
432 * \param[out] forceWithVirial The output force and virial
433 * \param[out] enerd The output energies
434 * \param[in] lambdaQ The Coulomb lambda to use when calculating the results.
436 GPU_FUNC_QUALIFIER void pme_gpu_wait_and_reduce(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
437 const gmx::StepWorkload& GPU_FUNC_ARGUMENT(stepWork),
438 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
439 gmx::ForceWithVirial* GPU_FUNC_ARGUMENT(forceWithVirial),
440 gmx_enerdata_t* GPU_FUNC_ARGUMENT(enerd),
441 real GPU_FUNC_ARGUMENT(lambdaQ)) GPU_FUNC_TERM;
444 * The PME GPU reinitialization function that is called both at the end of any PME computation and on any load balancing.
446 * Clears the internal grid and energy/virial buffers; it is not safe to start
447 * the PME computation without calling this.
448 * Note that unlike in the nbnxn module, the force buffer does not need clearing.
450 * \todo Rename this function to *clear* -- it clearly only does output resetting
451 * and we should be clear about what the function does..
453 * \param[in] pme The PME data structure.
454 * \param[in] wcycle The wallclock counter.
456 GPU_FUNC_QUALIFIER void pme_gpu_reinit_computation(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
457 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
459 /*! \brief Set pointer to device copy of coordinate data.
460 * \param[in] pme The PME data structure.
461 * \param[in] d_x The pointer to the positions buffer to be set
463 GPU_FUNC_QUALIFIER void pme_gpu_set_device_x(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
464 DeviceBuffer<gmx::RVec> GPU_FUNC_ARGUMENT(d_x)) GPU_FUNC_TERM;
466 /*! \brief Get pointer to device copy of force data.
467 * \param[in] pme The PME data structure.
468 * \returns Pointer to force data
470 GPU_FUNC_QUALIFIER void* pme_gpu_get_device_f(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
471 GPU_FUNC_TERM_WITH_RETURN(nullptr);
473 /*! \brief Get pointer to the device synchronizer object that allows syncing on PME force calculation completion
474 * \param[in] pme The PME data structure.
475 * \returns Pointer to sychronizer
477 GPU_FUNC_QUALIFIER GpuEventSynchronizer* pme_gpu_get_f_ready_synchronizer(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
478 GPU_FUNC_TERM_WITH_RETURN(nullptr);