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/timing/walltime_accounting.h"
58 #include "gromacs/utility/arrayref.h"
59 #include "gromacs/utility/basedefinitions.h"
60 #include "gromacs/utility/real.h"
63 struct interaction_const_t;
69 struct gmx_wallclock_gpu_pme_t;
70 struct gmx_device_info_t;
71 struct gmx_enerdata_t;
77 enum class GpuTaskCompletion;
79 class GpuEventSynchronizer;
81 using PmeGpuProgramHandle = const PmeGpuProgram*;
86 class ForceWithVirial;
88 enum class PinningPolicy : int;
97 /*! \brief Possible PME codepaths on a rank.
98 * \todo: make this enum class with gmx_pme_t C++ refactoring
100 enum class PmeRunMode
102 None, //!< No PME task is done
103 CPU, //!< Whole PME computation is done on CPU
104 GPU, //!< Whole PME computation is done on GPU
105 Mixed, //!< Mixed mode: only spread and gather run on GPU; FFT and solving are done on CPU.
108 //! PME gathering output forces treatment
109 enum class PmeForceOutputHandling
111 Set, /**< Gather simply writes into provided force buffer */
112 ReduceWithInput, /**< Gather adds its output to the buffer.
113 On GPU, that means additional H2D copy before the kernel launch. */
116 /*! \brief Return the smallest allowed PME grid size for \p pmeOrder */
117 int minimalPmeGridSize(int pmeOrder);
119 /*! \brief Check restrictions on pme_order and the PME grid nkx,nky,nkz.
121 * With errorsAreFatal=true, an exception or fatal error is generated
122 * on violation of restrictions.
123 * With errorsAreFatal=false, false is returned on violation of restrictions.
124 * When all restrictions are obeyed, true is returned.
125 * Argument useThreads tells if any MPI rank doing PME uses more than 1 threads.
126 * If at calling useThreads is unknown, pass true for conservative checking.
128 * The PME GPU restrictions are checked separately during pme_gpu_init().
130 bool gmx_pme_check_restrictions(int pme_order,
134 int numPmeDomainsAlongX,
136 bool errorsAreFatal);
138 /*! \brief Construct PME data
140 * \throws gmx::InconsistentInputError if input grid sizes/PME order are inconsistent.
141 * \returns Pointer to newly allocated and initialized PME data.
143 * \todo We should evolve something like a \c GpuManager that holds \c
144 * gmx_device_info_t * and \c PmeGpuProgramHandle and perhaps other
145 * related things whose lifetime can/should exceed that of a task (or
146 * perhaps task manager). See Redmine #2522.
148 gmx_pme_t* gmx_pme_init(const t_commrec* cr,
149 const NumPmeDomains& numPmeDomains,
150 const t_inputrec* ir,
151 gmx_bool bFreeEnergy_q,
152 gmx_bool bFreeEnergy_lj,
153 gmx_bool bReproducible,
159 const gmx_device_info_t* gpuInfo,
160 PmeGpuProgramHandle pmeGpuProgram,
161 const gmx::MDLogger& mdlog);
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 Called on the nodes that do PME exclusively */
218 int gmx_pmeonly(struct gmx_pme_t* pme,
221 gmx_wallcycle* wcycle,
222 gmx_walltime_accounting_t walltime_accounting,
226 /*! \brief Calculate the PME grid energy V for n charges.
228 * The potential (found in \p pme) must have been found already with a
229 * call to gmx_pme_do() with at least GMX_PME_SPREAD and GMX_PME_SOLVE
230 * specified. Note that the charges are not spread on the grid in the
231 * pme struct. Currently does not work in parallel or with free
234 void gmx_pme_calc_energy(gmx_pme_t* pme, gmx::ArrayRef<const gmx::RVec> x, gmx::ArrayRef<const real> q, real* V);
236 /*! \brief Send the charges and maxshift to out PME-only node. */
237 void gmx_pme_send_parameters(const t_commrec* cr,
238 const interaction_const_t* ic,
239 gmx_bool bFreeEnergy_q,
240 gmx_bool bFreeEnergy_lj,
250 /*! \brief Send the coordinates to our PME-only node and request a PME calculation */
251 void gmx_pme_send_coordinates(t_forcerec* fr,
259 bool useGpuPmePpComms,
260 bool reinitGpuPmePpComms,
261 bool sendCoordinatesFromGpu,
262 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent,
263 gmx_wallcycle* wcycle);
265 /*! \brief Tell our PME-only node to finish */
266 void gmx_pme_send_finish(const t_commrec* cr);
268 /*! \brief Tell our PME-only node to reset all cycle and flop counters */
269 void gmx_pme_send_resetcounters(const t_commrec* cr, int64_t step);
271 /*! \brief PP nodes receive the long range forces from the PME nodes */
272 void gmx_pme_receive_f(gmx::PmePpCommGpu* pmePpCommGpu,
274 gmx::ForceWithVirial* forceWithVirial,
279 bool useGpuPmePpComms,
280 bool receivePmeForceToGpu,
284 * This function updates the local atom data on GPU after DD (charges, coordinates, etc.).
285 * TODO: it should update the PME CPU atom data as well.
286 * (currently PME CPU call gmx_pme_do() gets passed the input pointers for each computation).
288 * \param[in,out] pme The PME structure.
289 * \param[in] numAtoms The number of particles.
290 * \param[in] charges The pointer to the array of particle charges.
292 void gmx_pme_reinit_atoms(gmx_pme_t* pme, int numAtoms, const real* charges);
294 /* A block of PME GPU functions */
296 /*! \brief Checks whether the GROMACS build allows to run PME on GPU.
297 * TODO: this partly duplicates an internal PME assert function
298 * pme_gpu_check_restrictions(), except that works with a
299 * formed gmx_pme_t structure. Should that one go away/work with inputrec?
301 * \param[out] error If non-null, the error message when PME is not supported on GPU.
303 * \returns true if PME can run on GPU on this build, false otherwise.
305 bool pme_gpu_supports_build(std::string* error);
307 /*! \brief Checks whether the detected (GPU) hardware allows to run PME on GPU.
309 * \param[in] hwinfo Information about the detected hardware
310 * \param[out] error If non-null, the error message when PME is not supported on GPU.
312 * \returns true if PME can run on GPU on this build, false otherwise.
314 bool pme_gpu_supports_hardware(const gmx_hw_info_t& hwinfo, std::string* error);
316 /*! \brief Checks whether the input system allows to run PME on GPU.
317 * TODO: this partly duplicates an internal PME assert function
318 * pme_gpu_check_restrictions(), except that works with a
319 * formed gmx_pme_t structure. Should that one go away/work with inputrec?
321 * \param[in] ir Input system.
322 * \param[in] mtop Complete system topology to check if an FE simulation perturbs charges.
323 * \param[out] error If non-null, the error message if the input is not supported on GPU.
325 * \returns true if PME can run on GPU with this input, false otherwise.
327 bool pme_gpu_supports_input(const t_inputrec& ir, const gmx_mtop_t& mtop, std::string* error);
330 * Returns the active PME codepath (CPU, GPU, mixed).
331 * \todo This is a rather static data that should be managed by the higher level task scheduler.
333 * \param[in] pme The PME data structure.
334 * \returns active PME codepath.
336 PmeRunMode pme_run_mode(const gmx_pme_t* pme);
338 /*! \libinternal \brief
339 * Return the pinning policy appropriate for this build configuration
340 * for relevant buffers used for PME task on this rank (e.g. running
342 gmx::PinningPolicy pme_get_pinning_policy();
345 * Tells if PME is enabled to run on GPU (not necessarily active at the moment).
346 * \todo This is a rather static data that should be managed by the hardware assignment manager.
347 * For now, it is synonymous with the active PME codepath (in the absence of dynamic switching).
349 * \param[in] pme The PME data structure.
350 * \returns true if PME can run on GPU, false otherwise.
352 inline bool pme_gpu_task_enabled(const gmx_pme_t* pme)
354 return (pme != nullptr) && (pme_run_mode(pme) != PmeRunMode::CPU);
357 /*! \brief Returns the size of the padding needed by GPU version of PME in the coordinates array.
359 * \param[in] pme The PME data structure.
361 GPU_FUNC_QUALIFIER int pme_gpu_get_padding_size(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
362 GPU_FUNC_TERM_WITH_RETURN(0);
364 // The following functions are all the PME GPU entry points,
365 // currently inlining to nothing on non-CUDA builds.
368 * Resets the PME GPU timings. To be called at the reset step.
370 * \param[in] pme The PME structure.
372 GPU_FUNC_QUALIFIER void pme_gpu_reset_timings(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme)) GPU_FUNC_TERM;
375 * Copies the PME GPU timings to the gmx_wallclock_gpu_pme_t structure (for log output). To be called at the run end.
377 * \param[in] pme The PME structure.
378 * \param[in] timings The gmx_wallclock_gpu_pme_t structure.
380 GPU_FUNC_QUALIFIER void pme_gpu_get_timings(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
381 gmx_wallclock_gpu_pme_t* GPU_FUNC_ARGUMENT(timings)) GPU_FUNC_TERM;
383 /* The main PME GPU functions */
386 * Prepares PME on GPU computation (updating the box if needed)
387 * \param[in] pme The PME data structure.
388 * \param[in] needToUpdateBox Tells if the stored unit cell parameters should be updated from \p box.
389 * \param[in] box The unit cell box.
390 * \param[in] wcycle The wallclock counter.
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] useGpuForceReduction Whether PME forces are reduced on GPU this step or should be downloaded for CPU reduction
395 GPU_FUNC_QUALIFIER void pme_gpu_prepare_computation(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
396 bool GPU_FUNC_ARGUMENT(needToUpdateBox),
397 const matrix GPU_FUNC_ARGUMENT(box),
398 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
399 int GPU_FUNC_ARGUMENT(flags),
400 bool GPU_FUNC_ARGUMENT(useGpuForceReduction)) GPU_FUNC_TERM;
403 * Launches first stage of PME on GPU - spreading kernel.
405 * \param[in] pme The PME data structure.
406 * \param[in] xReadyOnDevice Event synchronizer indicating that the coordinates are ready in the device memory; nullptr allowed only on separate PME ranks.
407 * \param[in] wcycle The wallclock counter.
409 GPU_FUNC_QUALIFIER void pme_gpu_launch_spread(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
410 GpuEventSynchronizer* GPU_FUNC_ARGUMENT(xReadyOnDevice),
411 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
414 * Launches middle stages of PME (FFT R2C, solving, FFT C2R) either on GPU or on CPU, depending on the run mode.
416 * \param[in] pme The PME data structure.
417 * \param[in] wcycle The wallclock counter.
419 GPU_FUNC_QUALIFIER void pme_gpu_launch_complex_transforms(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
420 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
423 * Launches last stage of PME on GPU - force gathering and D2H force transfer.
425 * \param[in] pme The PME data structure.
426 * \param[in] wcycle The wallclock counter.
427 * \param[in] forceTreatment Tells how data should be treated. The gathering kernel either
428 * stores the output reciprocal forces into the host array, or copies its contents to the GPU first
429 * and accumulates. The reduction is non-atomic.
431 GPU_FUNC_QUALIFIER void pme_gpu_launch_gather(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
432 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
433 PmeForceOutputHandling GPU_FUNC_ARGUMENT(forceTreatment)) GPU_FUNC_TERM;
436 * Attempts to complete PME GPU tasks.
438 * The \p completionKind argument controls whether the function blocks until all
439 * PME GPU tasks enqueued completed (as pme_gpu_wait_finish_task() does) or only
440 * checks and returns immediately if they did not.
441 * When blocking or the tasks have completed it also gets the output forces
442 * by assigning the ArrayRef to the \p forces pointer passed in.
443 * Virial/energy are also outputs if they were to be computed.
445 * \param[in] pme The PME data structure.
446 * \param[in] flags The combination of flags to affect this PME computation.
447 * The flags are the GMX_PME_ flags from pme.h.
448 * \param[in] wcycle The wallclock counter.
449 * \param[out] forceWithVirial The output force and virial
450 * \param[out] enerd The output energies
451 * \param[in] flags The combination of flags to affect this PME computation.
452 * The flags are the GMX_PME_ flags from pme.h.
453 * \param[in] completionKind Indicates whether PME task completion should only be checked rather
454 * than waited for \returns True if the PME GPU tasks have completed
456 GPU_FUNC_QUALIFIER bool pme_gpu_try_finish_task(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
457 int GPU_FUNC_ARGUMENT(flags),
458 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
459 gmx::ForceWithVirial* GPU_FUNC_ARGUMENT(forceWithVirial),
460 gmx_enerdata_t* GPU_FUNC_ARGUMENT(enerd),
461 GpuTaskCompletion GPU_FUNC_ARGUMENT(completionKind))
462 GPU_FUNC_TERM_WITH_RETURN(false);
465 * Blocks until PME GPU tasks are completed, and gets the output forces and virial/energy
466 * (if they were to be computed).
468 * \param[in] pme The PME data structure.
469 * \param[in] flags The combination of flags to affect this PME computation.
470 * The flags are the GMX_PME_ flags from pme.h.
471 * \param[in] wcycle The wallclock counter.
472 * \param[out] forceWithVirial The output force and virial
473 * \param[out] enerd The output energies
475 GPU_FUNC_QUALIFIER void pme_gpu_wait_and_reduce(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
476 int GPU_FUNC_ARGUMENT(flags),
477 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
478 gmx::ForceWithVirial* GPU_FUNC_ARGUMENT(forceWithVirial),
479 gmx_enerdata_t* GPU_FUNC_ARGUMENT(enerd)) GPU_FUNC_TERM;
482 * The PME GPU reinitialization function that is called both at the end of any PME computation and on any load balancing.
484 * Clears the internal grid and energy/virial buffers; it is not safe to start
485 * the PME computation without calling this.
486 * Note that unlike in the nbnxn module, the force buffer does not need clearing.
488 * \todo Rename this function to *clear* -- it clearly only does output resetting
489 * and we should be clear about what the function does..
491 * \param[in] pme The PME data structure.
492 * \param[in] wcycle The wallclock counter.
494 GPU_FUNC_QUALIFIER void pme_gpu_reinit_computation(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
495 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
498 /*! \brief Get pointer to device copy of coordinate data.
499 * \param[in] pme The PME data structure.
500 * \returns Pointer to coordinate data
502 GPU_FUNC_QUALIFIER DeviceBuffer<float> pme_gpu_get_device_x(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
503 GPU_FUNC_TERM_WITH_RETURN(DeviceBuffer<float>{});
505 /*! \brief Set pointer to device copy of coordinate data.
506 * \param[in] pme The PME data structure.
507 * \param[in] d_x The pointer to the positions buffer to be set
509 GPU_FUNC_QUALIFIER void pme_gpu_set_device_x(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
510 DeviceBuffer<float> GPU_FUNC_ARGUMENT(d_x)) GPU_FUNC_TERM;
512 /*! \brief Get pointer to device copy of force data.
513 * \param[in] pme The PME data structure.
514 * \returns Pointer to force data
516 GPU_FUNC_QUALIFIER void* pme_gpu_get_device_f(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
517 GPU_FUNC_TERM_WITH_RETURN(nullptr);
519 /*! \brief Returns the pointer to the GPU stream.
520 * \param[in] pme The PME data structure.
521 * \returns Pointer to GPU stream object.
523 GPU_FUNC_QUALIFIER void* pme_gpu_get_device_stream(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
524 GPU_FUNC_TERM_WITH_RETURN(nullptr);
526 /*! \brief Returns the pointer to the GPU context.
527 * \param[in] pme The PME data structure.
528 * \returns Pointer to GPU context object.
530 GPU_FUNC_QUALIFIER void* pme_gpu_get_device_context(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
531 GPU_FUNC_TERM_WITH_RETURN(nullptr);
533 /*! \brief Get pointer to the device synchronizer object that allows syncing on PME force calculation completion
534 * \param[in] pme The PME data structure.
535 * \returns Pointer to sychronizer
537 GPU_FUNC_QUALIFIER GpuEventSynchronizer* pme_gpu_get_f_ready_synchronizer(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
538 GPU_FUNC_TERM_WITH_RETURN(nullptr);