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/arrayref.h"
58 #include "gromacs/utility/basedefinitions.h"
59 #include "gromacs/utility/real.h"
66 struct gmx_wallclock_gpu_pme_t;
67 struct DeviceInformation;
68 struct gmx_enerdata_t;
74 enum class GpuTaskCompletion;
76 class GpuEventSynchronizer;
80 class ForceWithVirial;
82 enum class PinningPolicy : int;
91 /*! \brief Possible PME codepaths on a rank.
92 * \todo: make this enum class with gmx_pme_t C++ refactoring
96 None, //!< No PME task is done
97 CPU, //!< Whole PME computation is done on CPU
98 GPU, //!< Whole PME computation is done on GPU
99 Mixed, //!< Mixed mode: only spread and gather run on GPU; FFT and solving are done on CPU.
102 /*! \brief Return the smallest allowed PME grid size for \p pmeOrder */
103 int minimalPmeGridSize(int pmeOrder);
105 //! Return whether the grid of \c pme is identical to \c grid_size.
106 bool gmx_pme_grid_matches(const gmx_pme_t& pme, const ivec grid_size);
108 /*! \brief Check restrictions on pme_order and the PME grid nkx,nky,nkz.
110 * With errorsAreFatal=true, an exception or fatal error is generated
111 * on violation of restrictions.
112 * With errorsAreFatal=false, false is returned on violation of restrictions.
113 * When all restrictions are obeyed, true is returned.
114 * Argument useThreads tells if any MPI rank doing PME uses more than 1 threads.
115 * If at calling useThreads is unknown, pass true for conservative checking.
117 * The PME GPU restrictions are checked separately during pme_gpu_init().
119 bool gmx_pme_check_restrictions(int pme_order,
123 int numPmeDomainsAlongX,
125 bool errorsAreFatal);
127 /*! \brief Construct PME data
129 * \throws gmx::InconsistentInputError if input grid sizes/PME order are inconsistent.
130 * \returns Pointer to newly allocated and initialized PME data.
132 * \todo We should evolve something like a \c GpuManager that holds \c
133 * DeviceInformation* and \c PmeGpuProgram* and perhaps other
134 * related things whose lifetime can/should exceed that of a task (or
135 * perhaps task manager). See Redmine #2522.
137 gmx_pme_t* gmx_pme_init(const t_commrec* cr,
138 const NumPmeDomains& numPmeDomains,
139 const t_inputrec* ir,
140 gmx_bool bFreeEnergy_q,
141 gmx_bool bFreeEnergy_lj,
142 gmx_bool bReproducible,
148 const DeviceInformation* deviceInfo,
149 const PmeGpuProgram* pmeGpuProgram,
150 const gmx::MDLogger& mdlog);
152 /*! \brief As gmx_pme_init, but takes most settings, except the grid/Ewald coefficients, from
153 * pme_src. This is only called when the PME cut-off/grid size changes.
155 void gmx_pme_reinit(gmx_pme_t** pmedata,
158 const t_inputrec* ir,
159 const ivec grid_size,
163 /*! \brief Destroys the PME data structure.*/
164 void gmx_pme_destroy(gmx_pme_t* pme);
167 /*! \brief Flag values that control what gmx_pme_do() will calculate
169 * These can be combined with bitwise-OR if more than one thing is required.
171 #define GMX_PME_SPREAD (1 << 0)
172 #define GMX_PME_SOLVE (1 << 1)
173 #define GMX_PME_CALC_F (1 << 2)
174 #define GMX_PME_CALC_ENER_VIR (1 << 3)
175 /* This forces the grid to be backtransformed even without GMX_PME_CALC_F */
176 #define GMX_PME_CALC_POT (1 << 4)
178 #define GMX_PME_DO_ALL_F (GMX_PME_SPREAD | GMX_PME_SOLVE | GMX_PME_CALC_F)
181 /*! \brief Do a PME calculation on a CPU for the long range electrostatics and/or LJ.
183 * Computes the PME forces and the energy and viral, when requested,
184 * for all atoms in \p coordinates. Forces, when requested, are added
185 * to the buffer \p forces, which is allowed to contain more elements
186 * than the number of elements in \p coordinates.
187 * The meaning of \p flags is defined above, and determines which
188 * parts of the calculation are performed.
190 * \return 0 indicates all well, non zero is an error code.
192 int gmx_pme_do(struct gmx_pme_t* pme,
193 gmx::ArrayRef<const gmx::RVec> coordinates,
194 gmx::ArrayRef<gmx::RVec> forces,
206 gmx_wallcycle* wcycle,
217 /*! \brief Calculate the PME grid energy V for n charges.
219 * The potential (found in \p pme) must have been found already with a
220 * call to gmx_pme_do() with at least GMX_PME_SPREAD and GMX_PME_SOLVE
221 * specified. Note that the charges are not spread on the grid in the
222 * pme struct. Currently does not work in parallel or with free
225 void gmx_pme_calc_energy(gmx_pme_t* pme, gmx::ArrayRef<const gmx::RVec> x, gmx::ArrayRef<const real> q, real* V);
228 * This function updates the local atom data on GPU after DD (charges, coordinates, etc.).
229 * TODO: it should update the PME CPU atom data as well.
230 * (currently PME CPU call gmx_pme_do() gets passed the input pointers for each computation).
232 * \param[in,out] pme The PME structure.
233 * \param[in] numAtoms The number of particles.
234 * \param[in] charges The pointer to the array of particle charges.
236 void gmx_pme_reinit_atoms(gmx_pme_t* pme, int numAtoms, const real* charges);
238 /* A block of PME GPU functions */
240 /*! \brief Checks whether the GROMACS build allows to run PME on GPU.
241 * TODO: this partly duplicates an internal PME assert function
242 * pme_gpu_check_restrictions(), except that works with a
243 * formed gmx_pme_t structure. Should that one go away/work with inputrec?
245 * \param[out] error If non-null, the error message when PME is not supported on GPU.
247 * \returns true if PME can run on GPU on this build, false otherwise.
249 bool pme_gpu_supports_build(std::string* error);
251 /*! \brief Checks whether the detected (GPU) hardware allows to run PME on GPU.
253 * \param[in] hwinfo Information about the detected hardware
254 * \param[out] error If non-null, the error message when PME is not supported on GPU.
256 * \returns true if PME can run on GPU on this build, false otherwise.
258 bool pme_gpu_supports_hardware(const gmx_hw_info_t& hwinfo, std::string* error);
260 /*! \brief Checks whether the input system allows to run PME on GPU.
261 * TODO: this partly duplicates an internal PME assert function
262 * pme_gpu_check_restrictions(), except that works with a
263 * formed gmx_pme_t structure. Should that one go away/work with inputrec?
265 * \param[in] ir Input system.
266 * \param[in] mtop Complete system topology to check if an FE simulation perturbs charges.
267 * \param[out] error If non-null, the error message if the input is not supported on GPU.
269 * \returns true if PME can run on GPU with this input, false otherwise.
271 bool pme_gpu_supports_input(const t_inputrec& ir, const gmx_mtop_t& mtop, std::string* error);
274 * Returns the active PME codepath (CPU, GPU, mixed).
275 * \todo This is a rather static data that should be managed by the higher level task scheduler.
277 * \param[in] pme The PME data structure.
278 * \returns active PME codepath.
280 PmeRunMode pme_run_mode(const gmx_pme_t* pme);
282 /*! \libinternal \brief
283 * Return the pinning policy appropriate for this build configuration
284 * for relevant buffers used for PME task on this rank (e.g. running
286 gmx::PinningPolicy pme_get_pinning_policy();
289 * Tells if PME is enabled to run on GPU (not necessarily active at the moment).
290 * \todo This is a rather static data that should be managed by the hardware assignment manager.
291 * For now, it is synonymous with the active PME codepath (in the absence of dynamic switching).
293 * \param[in] pme The PME data structure.
294 * \returns true if PME can run on GPU, false otherwise.
296 inline bool pme_gpu_task_enabled(const gmx_pme_t* pme)
298 return (pme != nullptr) && (pme_run_mode(pme) != PmeRunMode::CPU);
301 /*! \brief Returns the size of the padding needed by GPU version of PME in the coordinates array.
303 * \param[in] pme The PME data structure.
305 GPU_FUNC_QUALIFIER int pme_gpu_get_padding_size(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
306 GPU_FUNC_TERM_WITH_RETURN(0);
308 // The following functions are all the PME GPU entry points,
309 // currently inlining to nothing on non-CUDA builds.
312 * Resets the PME GPU timings. To be called at the reset step.
314 * \param[in] pme The PME structure.
316 GPU_FUNC_QUALIFIER void pme_gpu_reset_timings(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme)) GPU_FUNC_TERM;
319 * Copies the PME GPU timings to the gmx_wallclock_gpu_pme_t structure (for log output). To be called at the run end.
321 * \param[in] pme The PME structure.
322 * \param[in] timings The gmx_wallclock_gpu_pme_t structure.
324 GPU_FUNC_QUALIFIER void pme_gpu_get_timings(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
325 gmx_wallclock_gpu_pme_t* GPU_FUNC_ARGUMENT(timings)) GPU_FUNC_TERM;
327 /* The main PME GPU functions */
330 * Prepares PME on GPU computation (updating the box if needed)
331 * \param[in] pme The PME data structure.
332 * \param[in] needToUpdateBox Tells if the stored unit cell parameters should be updated from \p box.
333 * \param[in] box The unit cell box.
334 * \param[in] wcycle The wallclock counter.
335 * \param[in] flags The combination of flags to affect this PME computation.
336 * The flags are the GMX_PME_ flags from pme.h.
337 * \param[in] useGpuForceReduction Whether PME forces are reduced on GPU this step or should be downloaded for CPU reduction
339 GPU_FUNC_QUALIFIER void pme_gpu_prepare_computation(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
340 bool GPU_FUNC_ARGUMENT(needToUpdateBox),
341 const matrix GPU_FUNC_ARGUMENT(box),
342 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
343 int GPU_FUNC_ARGUMENT(flags),
344 bool GPU_FUNC_ARGUMENT(useGpuForceReduction)) GPU_FUNC_TERM;
347 * Launches first stage of PME on GPU - spreading kernel.
349 * \param[in] pme The PME data structure.
350 * \param[in] xReadyOnDevice Event synchronizer indicating that the coordinates are ready in the device memory; nullptr allowed only on separate PME ranks.
351 * \param[in] wcycle The wallclock counter.
353 GPU_FUNC_QUALIFIER void pme_gpu_launch_spread(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
354 GpuEventSynchronizer* GPU_FUNC_ARGUMENT(xReadyOnDevice),
355 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
358 * Launches middle stages of PME (FFT R2C, solving, FFT C2R) either on GPU or on CPU, depending on the run mode.
360 * \param[in] pme The PME data structure.
361 * \param[in] wcycle The wallclock counter.
363 GPU_FUNC_QUALIFIER void pme_gpu_launch_complex_transforms(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
364 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
367 * Launches last stage of PME on GPU - force gathering and D2H force transfer.
369 * \param[in] pme The PME data structure.
370 * \param[in] wcycle The wallclock counter.
372 GPU_FUNC_QUALIFIER void pme_gpu_launch_gather(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
373 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
376 * Attempts to complete PME GPU tasks.
378 * The \p completionKind argument controls whether the function blocks until all
379 * PME GPU tasks enqueued completed (as pme_gpu_wait_finish_task() does) or only
380 * checks and returns immediately if they did not.
381 * When blocking or the tasks have completed it also gets the output forces
382 * by assigning the ArrayRef to the \p forces pointer passed in.
383 * Virial/energy are also outputs if they were to be computed.
385 * \param[in] pme The PME data structure.
386 * \param[in] flags The combination of flags to affect this PME computation.
387 * The flags are the GMX_PME_ flags from pme.h.
388 * \param[in] wcycle The wallclock counter.
389 * \param[out] forceWithVirial The output force and virial
390 * \param[out] enerd The output energies
391 * \param[in] flags The combination of flags to affect this PME computation.
392 * The flags are the GMX_PME_ flags from pme.h.
393 * \param[in] completionKind Indicates whether PME task completion should only be checked rather
394 * than waited for \returns True if the PME GPU tasks have completed
396 GPU_FUNC_QUALIFIER bool pme_gpu_try_finish_task(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
397 int GPU_FUNC_ARGUMENT(flags),
398 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
399 gmx::ForceWithVirial* GPU_FUNC_ARGUMENT(forceWithVirial),
400 gmx_enerdata_t* GPU_FUNC_ARGUMENT(enerd),
401 GpuTaskCompletion GPU_FUNC_ARGUMENT(completionKind))
402 GPU_FUNC_TERM_WITH_RETURN(false);
405 * Blocks until PME GPU tasks are completed, and gets the output forces and virial/energy
406 * (if they were to be computed).
408 * \param[in] pme The PME data structure.
409 * \param[in] flags The combination of flags to affect this PME computation.
410 * The flags are the GMX_PME_ flags from pme.h.
411 * \param[in] wcycle The wallclock counter.
412 * \param[out] forceWithVirial The output force and virial
413 * \param[out] enerd The output energies
415 GPU_FUNC_QUALIFIER void pme_gpu_wait_and_reduce(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
416 int GPU_FUNC_ARGUMENT(flags),
417 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
418 gmx::ForceWithVirial* GPU_FUNC_ARGUMENT(forceWithVirial),
419 gmx_enerdata_t* GPU_FUNC_ARGUMENT(enerd)) GPU_FUNC_TERM;
422 * The PME GPU reinitialization function that is called both at the end of any PME computation and on any load balancing.
424 * Clears the internal grid and energy/virial buffers; it is not safe to start
425 * the PME computation without calling this.
426 * Note that unlike in the nbnxn module, the force buffer does not need clearing.
428 * \todo Rename this function to *clear* -- it clearly only does output resetting
429 * and we should be clear about what the function does..
431 * \param[in] pme The PME data structure.
432 * \param[in] wcycle The wallclock counter.
434 GPU_FUNC_QUALIFIER void pme_gpu_reinit_computation(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
435 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
437 /*! \brief Set pointer to device copy of coordinate data.
438 * \param[in] pme The PME data structure.
439 * \param[in] d_x The pointer to the positions buffer to be set
441 GPU_FUNC_QUALIFIER void pme_gpu_set_device_x(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
442 DeviceBuffer<gmx::RVec> GPU_FUNC_ARGUMENT(d_x)) GPU_FUNC_TERM;
444 /*! \brief Get pointer to device copy of force data.
445 * \param[in] pme The PME data structure.
446 * \returns Pointer to force data
448 GPU_FUNC_QUALIFIER void* pme_gpu_get_device_f(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
449 GPU_FUNC_TERM_WITH_RETURN(nullptr);
451 /*! \brief Returns the pointer to the GPU stream.
452 * \param[in] pme The PME data structure.
453 * \returns Pointer to GPU stream object.
455 GPU_FUNC_QUALIFIER void* pme_gpu_get_device_stream(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
456 GPU_FUNC_TERM_WITH_RETURN(nullptr);
458 /*! \brief Returns the pointer to the GPU context.
459 * \param[in] pme The PME data structure.
460 * \returns Pointer to GPU context object.
462 GPU_FUNC_QUALIFIER void* pme_gpu_get_device_context(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
463 GPU_FUNC_TERM_WITH_RETURN(nullptr);
465 /*! \brief Get pointer to the device synchronizer object that allows syncing on PME force calculation completion
466 * \param[in] pme The PME data structure.
467 * \returns Pointer to sychronizer
469 GPU_FUNC_QUALIFIER GpuEventSynchronizer* pme_gpu_get_f_ready_synchronizer(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
470 GPU_FUNC_TERM_WITH_RETURN(nullptr);