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, 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
54 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
55 #include "gromacs/gpu_utils/gpu_macros.h"
56 #include "gromacs/math/vectypes.h"
57 #include "gromacs/utility/basedefinitions.h"
58 #include "gromacs/utility/real.h"
65 struct gmx_wallclock_gpu_pme_t;
66 struct gmx_enerdata_t;
74 enum class GpuTaskCompletion;
76 class GpuEventSynchronizer;
82 class ForceWithVirial;
84 enum class PinningPolicy : int;
94 /*! \brief Possible PME codepaths on a rank.
95 * \todo: make this enum class with gmx_pme_t C++ refactoring
99 None, //!< No PME task is done
100 CPU, //!< Whole PME computation is done on CPU
101 GPU, //!< Whole PME computation is done on GPU
102 Mixed, //!< Mixed mode: only spread and gather run on GPU; FFT and solving are done on CPU.
105 /*! \brief Return the smallest allowed PME grid size for \p pmeOrder */
106 int minimalPmeGridSize(int pmeOrder);
108 //! Return whether the grid of \c pme is identical to \c grid_size.
109 bool gmx_pme_grid_matches(const gmx_pme_t& pme, const ivec grid_size);
111 /*! \brief Check restrictions on pme_order and the PME grid nkx,nky,nkz.
113 * With errorsAreFatal=true, an exception or fatal error is generated
114 * on violation of restrictions.
115 * With errorsAreFatal=false, false is returned on violation of restrictions.
116 * When all restrictions are obeyed, true is returned.
117 * Argument useThreads tells if any MPI rank doing PME uses more than 1 threads.
118 * If at calling useThreads is unknown, pass true for conservative checking.
120 * The PME GPU restrictions are checked separately during pme_gpu_init().
122 bool gmx_pme_check_restrictions(int pme_order,
126 int numPmeDomainsAlongX,
128 bool errorsAreFatal);
130 /*! \brief Construct PME data
132 * \throws gmx::InconsistentInputError if input grid sizes/PME order are inconsistent.
133 * \returns Pointer to newly allocated and initialized PME data.
135 * \todo We should evolve something like a \c GpuManager that holds \c
136 * DeviceInformation* and \c PmeGpuProgram* and perhaps other
137 * related things whose lifetime can/should exceed that of a task (or
138 * perhaps task manager). See Issue #2522.
140 gmx_pme_t* gmx_pme_init(const t_commrec* cr,
141 const NumPmeDomains& numPmeDomains,
142 const t_inputrec* ir,
143 gmx_bool bFreeEnergy_q,
144 gmx_bool bFreeEnergy_lj,
145 gmx_bool bReproducible,
151 const DeviceContext* deviceContext,
152 const DeviceStream* deviceStream,
153 const PmeGpuProgram* pmeGpuProgram,
154 const gmx::MDLogger& mdlog);
156 /*! \brief As gmx_pme_init, but takes most settings, except the grid/Ewald coefficients, from
157 * pme_src. This is only called when the PME cut-off/grid size changes.
159 void gmx_pme_reinit(gmx_pme_t** pmedata,
162 const t_inputrec* ir,
163 const ivec grid_size,
167 /*! \brief Destroys the PME data structure.*/
168 void gmx_pme_destroy(gmx_pme_t* pme);
170 /*! \brief Do a PME calculation on a CPU for the long range electrostatics and/or LJ.
172 * Computes the PME forces and the energy and viral, when requested,
173 * for all atoms in \p coordinates. Forces, when requested, are added
174 * to the buffer \p forces, which is allowed to contain more elements
175 * than the number of elements in \p coordinates.
176 * The meaning of \p flags is defined above, and determines which
177 * parts of the calculation are performed.
179 * \return 0 indicates all well, non zero is an error code.
181 int gmx_pme_do(struct gmx_pme_t* pme,
182 gmx::ArrayRef<const gmx::RVec> coordinates,
183 gmx::ArrayRef<gmx::RVec> forces,
195 gmx_wallcycle* wcycle,
204 const gmx::StepWorkload& stepWork);
206 /*! \brief Calculate the PME grid energy V for n charges.
208 * The potential (found in \p pme) must have been found already with a
209 * call to gmx_pme_do(). Note that the charges are not spread on the grid in the
210 * pme struct. Currently does not work in parallel or with free
213 void gmx_pme_calc_energy(gmx_pme_t* pme, gmx::ArrayRef<const gmx::RVec> x, gmx::ArrayRef<const real> q, real* V);
216 * This function updates the local atom data on GPU after DD (charges, coordinates, etc.).
217 * TODO: it should update the PME CPU atom data as well.
218 * (currently PME CPU call gmx_pme_do() gets passed the input pointers for each computation).
220 * \param[in,out] pme The PME structure.
221 * \param[in] numAtoms The number of particles.
222 * \param[in] charges The pointer to the array of particle charges.
224 void gmx_pme_reinit_atoms(gmx_pme_t* pme, int numAtoms, const real* charges);
226 /* A block of PME GPU functions */
228 /*! \brief Checks whether the GROMACS build allows to run PME on GPU.
229 * TODO: this partly duplicates an internal PME assert function
230 * pme_gpu_check_restrictions(), except that works with a
231 * formed gmx_pme_t structure. Should that one go away/work with inputrec?
233 * \param[out] error If non-null, the error message when PME is not supported on GPU.
235 * \returns true if PME can run on GPU on this build, false otherwise.
237 bool pme_gpu_supports_build(std::string* error);
239 /*! \brief Checks whether the detected (GPU) hardware allows to run PME on GPU.
241 * \param[in] hwinfo Information about the detected hardware
242 * \param[out] error If non-null, the error message when PME is not supported on GPU.
244 * \returns true if PME can run on GPU on this build, false otherwise.
246 bool pme_gpu_supports_hardware(const gmx_hw_info_t& hwinfo, std::string* error);
248 /*! \brief Checks whether the input system allows to run PME on GPU.
249 * TODO: this partly duplicates an internal PME assert function
250 * pme_gpu_check_restrictions(), except that works with a
251 * formed gmx_pme_t structure. Should that one go away/work with inputrec?
253 * \param[in] ir Input system.
254 * \param[in] mtop Complete system topology to check if an FE simulation perturbs charges.
255 * \param[out] error If non-null, the error message if the input is not supported on GPU.
257 * \returns true if PME can run on GPU with this input, false otherwise.
259 bool pme_gpu_supports_input(const t_inputrec& ir, const gmx_mtop_t& mtop, std::string* error);
262 * Returns the active PME codepath (CPU, GPU, mixed).
263 * \todo This is a rather static data that should be managed by the higher level task scheduler.
265 * \param[in] pme The PME data structure.
266 * \returns active PME codepath.
268 PmeRunMode pme_run_mode(const gmx_pme_t* pme);
270 /*! \libinternal \brief
271 * Return the pinning policy appropriate for this build configuration
272 * for relevant buffers used for PME task on this rank (e.g. running
274 gmx::PinningPolicy pme_get_pinning_policy();
277 * Tells if PME is enabled to run on GPU (not necessarily active at the moment).
278 * \todo This is a rather static data that should be managed by the hardware assignment manager.
279 * For now, it is synonymous with the active PME codepath (in the absence of dynamic switching).
281 * \param[in] pme The PME data structure.
282 * \returns true if PME can run on GPU, false otherwise.
284 inline bool pme_gpu_task_enabled(const gmx_pme_t* pme)
286 return (pme != nullptr) && (pme_run_mode(pme) != PmeRunMode::CPU);
289 /*! \brief Returns the block size requirement
291 * The GPU version of PME requires that the coordinates array have a
292 * size divisible by the returned number.
294 * \param[in] pme The PME data structure.
296 GPU_FUNC_QUALIFIER int pme_gpu_get_block_size(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
297 GPU_FUNC_TERM_WITH_RETURN(0);
299 // The following functions are all the PME GPU entry points,
300 // currently inlining to nothing on non-CUDA builds.
303 * Resets the PME GPU timings. To be called at the reset step.
305 * \param[in] pme The PME structure.
307 GPU_FUNC_QUALIFIER void pme_gpu_reset_timings(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme)) GPU_FUNC_TERM;
310 * Copies the PME GPU timings to the gmx_wallclock_gpu_pme_t structure (for log output). To be called at the run end.
312 * \param[in] pme The PME structure.
313 * \param[in] timings The gmx_wallclock_gpu_pme_t structure.
315 GPU_FUNC_QUALIFIER void pme_gpu_get_timings(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
316 gmx_wallclock_gpu_pme_t* GPU_FUNC_ARGUMENT(timings)) GPU_FUNC_TERM;
318 /* The main PME GPU functions */
321 * Prepares PME on GPU computation (updating the box if needed)
322 * \param[in] pme The PME data structure.
323 * \param[in] box The unit cell box.
324 * \param[in] wcycle The wallclock counter.
325 * \param[in] stepWork The required work for this simulation step
327 GPU_FUNC_QUALIFIER void pme_gpu_prepare_computation(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
328 const matrix GPU_FUNC_ARGUMENT(box),
329 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
330 const gmx::StepWorkload& GPU_FUNC_ARGUMENT(stepWork)) GPU_FUNC_TERM;
333 * Launches first stage of PME on GPU - spreading kernel.
335 * \param[in] pme The PME data structure.
336 * \param[in] xReadyOnDevice Event synchronizer indicating that the coordinates are ready in the device memory; nullptr allowed only on separate PME ranks.
337 * \param[in] wcycle The wallclock counter.
339 GPU_FUNC_QUALIFIER void pme_gpu_launch_spread(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
340 GpuEventSynchronizer* GPU_FUNC_ARGUMENT(xReadyOnDevice),
341 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
344 * Launches middle stages of PME (FFT R2C, solving, FFT C2R) either on GPU or on CPU, depending on the run mode.
346 * \param[in] pme The PME data structure.
347 * \param[in] wcycle The wallclock counter.
348 * \param[in] stepWork The required work for this simulation step
350 GPU_FUNC_QUALIFIER void
351 pme_gpu_launch_complex_transforms(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
352 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
353 const gmx::StepWorkload& GPU_FUNC_ARGUMENT(stepWork)) GPU_FUNC_TERM;
356 * Launches last stage of PME on GPU - force gathering and D2H force transfer.
358 * \param[in] pme The PME data structure.
359 * \param[in] wcycle The wallclock counter.
361 GPU_FUNC_QUALIFIER void pme_gpu_launch_gather(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
362 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
365 * Attempts to complete PME GPU tasks.
367 * The \p completionKind argument controls whether the function blocks until all
368 * PME GPU tasks enqueued completed (as pme_gpu_wait_finish_task() does) or only
369 * checks and returns immediately if they did not.
370 * When blocking or the tasks have completed it also gets the output forces
371 * by assigning the ArrayRef to the \p forces pointer passed in.
372 * Virial/energy are also outputs if they were to be computed.
374 * \param[in] pme The PME data structure.
375 * \param[in] stepWork The required work for this simulation step
376 * \param[in] wcycle The wallclock counter.
377 * \param[out] forceWithVirial The output force and virial
378 * \param[out] enerd The output energies
379 * \param[in] completionKind Indicates whether PME task completion should only be checked rather
381 * \returns True if the PME GPU tasks have completed
383 GPU_FUNC_QUALIFIER bool pme_gpu_try_finish_task(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
384 const gmx::StepWorkload& GPU_FUNC_ARGUMENT(stepWork),
385 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
386 gmx::ForceWithVirial* GPU_FUNC_ARGUMENT(forceWithVirial),
387 gmx_enerdata_t* GPU_FUNC_ARGUMENT(enerd),
388 GpuTaskCompletion GPU_FUNC_ARGUMENT(completionKind))
389 GPU_FUNC_TERM_WITH_RETURN(false);
392 * Blocks until PME GPU tasks are completed, and gets the output forces and virial/energy
393 * (if they were to be computed).
395 * \param[in] pme The PME data structure.
396 * \param[in] stepWork The required work for this simulation step
397 * \param[in] wcycle The wallclock counter.
398 * \param[out] forceWithVirial The output force and virial
399 * \param[out] enerd The output energies
401 GPU_FUNC_QUALIFIER void pme_gpu_wait_and_reduce(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
402 const gmx::StepWorkload& GPU_FUNC_ARGUMENT(stepWork),
403 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
404 gmx::ForceWithVirial* GPU_FUNC_ARGUMENT(forceWithVirial),
405 gmx_enerdata_t* GPU_FUNC_ARGUMENT(enerd)) GPU_FUNC_TERM;
408 * The PME GPU reinitialization function that is called both at the end of any PME computation and on any load balancing.
410 * Clears the internal grid and energy/virial buffers; it is not safe to start
411 * the PME computation without calling this.
412 * Note that unlike in the nbnxn module, the force buffer does not need clearing.
414 * \todo Rename this function to *clear* -- it clearly only does output resetting
415 * and we should be clear about what the function does..
417 * \param[in] pme The PME data structure.
418 * \param[in] wcycle The wallclock counter.
420 GPU_FUNC_QUALIFIER void pme_gpu_reinit_computation(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
421 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
423 /*! \brief Set pointer to device copy of coordinate data.
424 * \param[in] pme The PME data structure.
425 * \param[in] d_x The pointer to the positions buffer to be set
427 GPU_FUNC_QUALIFIER void pme_gpu_set_device_x(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
428 DeviceBuffer<gmx::RVec> GPU_FUNC_ARGUMENT(d_x)) GPU_FUNC_TERM;
430 /*! \brief Get pointer to device copy of force data.
431 * \param[in] pme The PME data structure.
432 * \returns Pointer to force data
434 GPU_FUNC_QUALIFIER void* pme_gpu_get_device_f(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
435 GPU_FUNC_TERM_WITH_RETURN(nullptr);
437 /*! \brief Get pointer to the device synchronizer object that allows syncing on PME force calculation completion
438 * \param[in] pme The PME data structure.
439 * \returns Pointer to sychronizer
441 GPU_FUNC_QUALIFIER GpuEventSynchronizer* pme_gpu_get_f_ready_synchronizer(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
442 GPU_FUNC_TERM_WITH_RETURN(nullptr);