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) 2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
37 /*! \libinternal \file
39 * \brief This file contains function declarations necessary for
40 * computing energies and forces for the PME long-ranged part (Coulomb
43 * \author Berk Hess <hess@kth.se>
45 * \ingroup module_ewald
48 #ifndef GMX_EWALD_PME_H
49 #define GMX_EWALD_PME_H
53 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
54 #include "gromacs/gpu_utils/gpu_macros.h"
55 #include "gromacs/math/vectypes.h"
56 #include "gromacs/timing/walltime_accounting.h"
57 #include "gromacs/utility/arrayref.h"
58 #include "gromacs/utility/basedefinitions.h"
59 #include "gromacs/utility/real.h"
62 struct interaction_const_t;
68 struct gmx_wallclock_gpu_pme_t;
69 struct gmx_device_info_t;
70 struct gmx_enerdata_t;
76 enum class GpuTaskCompletion;
78 class GpuEventSynchronizer;
80 using PmeGpuProgramHandle = const PmeGpuProgram *;
85 class ForceWithVirial;
87 enum class PinningPolicy : int;
91 GMX_SUM_GRID_FORWARD, GMX_SUM_GRID_BACKWARD
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 //! PME gathering output forces treatment
106 enum class PmeForceOutputHandling
108 Set, /**< Gather simply writes into provided force buffer */
109 ReduceWithInput, /**< Gather adds its output to the buffer.
110 On GPU, that means additional H2D copy before the kernel launch. */
113 /*! \brief Return the smallest allowed PME grid size for \p pmeOrder */
114 int minimalPmeGridSize(int pmeOrder);
116 /*! \brief Check restrictions on pme_order and the PME grid nkx,nky,nkz.
118 * With errorsAreFatal=true, an exception or fatal error is generated
119 * on violation of restrictions.
120 * With errorsAreFatal=false, false is returned on violation of restrictions.
121 * When all restrictions are obeyed, true is returned.
122 * Argument useThreads tells if any MPI rank doing PME uses more than 1 threads.
123 * If at calling useThreads is unknown, pass true for conservative checking.
125 * The PME GPU restrictions are checked separately during pme_gpu_init().
127 bool gmx_pme_check_restrictions(int pme_order,
128 int nkx, int nky, int nkz,
129 int numPmeDomainsAlongX,
131 bool errorsAreFatal);
133 /*! \brief Construct PME data
135 * \throws gmx::InconsistentInputError if input grid sizes/PME order are inconsistent.
136 * \returns Pointer to newly allocated and initialized PME data.
138 * \todo We should evolve something like a \c GpuManager that holds \c
139 * gmx_device_info_t * and \c PmeGpuProgramHandle and perhaps other
140 * related things whose lifetime can/should exceed that of a task (or
141 * perhaps task manager). See Redmine #2522.
143 gmx_pme_t *gmx_pme_init(const t_commrec *cr,
144 const NumPmeDomains &numPmeDomains,
145 const t_inputrec *ir,
146 gmx_bool bFreeEnergy_q, gmx_bool bFreeEnergy_lj,
147 gmx_bool bReproducible,
148 real ewaldcoeff_q, real ewaldcoeff_lj,
152 const gmx_device_info_t *gpuInfo,
153 PmeGpuProgramHandle pmeGpuProgram,
154 const gmx::MDLogger &mdlog);
156 /*! \brief Destroys the PME data structure.*/
157 void gmx_pme_destroy(gmx_pme_t *pme);
160 /*! \brief Flag values that control what gmx_pme_do() will calculate
162 * These can be combined with bitwise-OR if more than one thing is required.
164 #define GMX_PME_SPREAD (1<<0)
165 #define GMX_PME_SOLVE (1<<1)
166 #define GMX_PME_CALC_F (1<<2)
167 #define GMX_PME_CALC_ENER_VIR (1<<3)
168 /* This forces the grid to be backtransformed even without GMX_PME_CALC_F */
169 #define GMX_PME_CALC_POT (1<<4)
171 #define GMX_PME_DO_ALL_F (GMX_PME_SPREAD | GMX_PME_SOLVE | GMX_PME_CALC_F)
174 /*! \brief Do a PME calculation on a CPU for the long range electrostatics and/or LJ.
176 * Computes the PME forces and the energy and viral, when requested,
177 * for all atoms in \p coordinates. Forces, when requested, are added
178 * to the buffer \p forces, which is allowed to contain more elements
179 * than the number of elements in \p coordinates.
180 * The meaning of \p flags is defined above, and determines which
181 * parts of the calculation are performed.
183 * \return 0 indicates all well, non zero is an error code.
185 int gmx_pme_do(struct gmx_pme_t *pme,
186 gmx::ArrayRef<const gmx::RVec> coordinates,
187 gmx::ArrayRef<gmx::RVec> forces,
188 real chargeA[], real chargeB[],
189 real c6A[], real c6B[],
190 real sigmaA[], real sigmaB[],
191 const matrix box, const t_commrec *cr,
192 int maxshift_x, int maxshift_y,
193 t_nrnb *nrnb, gmx_wallcycle *wcycle,
194 matrix vir_q, matrix vir_lj,
195 real *energy_q, real *energy_lj,
196 real lambda_q, real lambda_lj,
197 real *dvdlambda_q, real *dvdlambda_lj,
200 /*! \brief Called on the nodes that do PME exclusively */
201 int gmx_pmeonly(struct gmx_pme_t *pme,
202 const t_commrec *cr, t_nrnb *mynrnb,
203 gmx_wallcycle *wcycle,
204 gmx_walltime_accounting_t walltime_accounting,
205 t_inputrec *ir, PmeRunMode runMode);
207 /*! \brief Calculate the PME grid energy V for n charges.
209 * The potential (found in \p pme) must have been found already with a
210 * call to gmx_pme_do() with at least GMX_PME_SPREAD and GMX_PME_SOLVE
211 * specified. Note that the charges are not spread on the grid in the
212 * pme struct. Currently does not work in parallel or with free
215 void gmx_pme_calc_energy(gmx_pme_t *pme,
216 gmx::ArrayRef<const gmx::RVec> x,
217 gmx::ArrayRef<const real> q,
220 /*! \brief Send the charges and maxshift to out PME-only node. */
221 void gmx_pme_send_parameters(const t_commrec *cr,
222 const interaction_const_t *ic,
223 gmx_bool bFreeEnergy_q, gmx_bool bFreeEnergy_lj,
224 real *chargeA, real *chargeB,
225 real *sqrt_c6A, real *sqrt_c6B,
226 real *sigmaA, real *sigmaB,
227 int maxshift_x, int maxshift_y);
229 /*! \brief Send the coordinates to our PME-only node and request a PME calculation */
230 void gmx_pme_send_coordinates(t_forcerec *fr, const t_commrec *cr, const matrix box, const rvec *x,
231 real lambda_q, real lambda_lj,
233 int64_t step, bool useGpuPmePpComms,
234 bool reinitGpuPmePpComms,
235 bool sendCoordinatesFromGpu,
236 GpuEventSynchronizer *coordinatesReadyOnDeviceEvent, gmx_wallcycle *wcycle);
238 /*! \brief Tell our PME-only node to finish */
239 void gmx_pme_send_finish(const t_commrec *cr);
241 /*! \brief Tell our PME-only node to reset all cycle and flop counters */
242 void gmx_pme_send_resetcounters(const t_commrec *cr, int64_t step);
244 /*! \brief PP nodes receive the long range forces from the PME nodes */
245 void gmx_pme_receive_f(gmx::PmePpCommGpu *pmePpCommGpu,
247 gmx::ForceWithVirial *forceWithVirial,
248 real *energy_q, real *energy_lj,
249 real *dvdlambda_q, real *dvdlambda_lj,
250 bool useGpuPmePpComms, bool receivePmeForceToGpu,
254 * This function updates the local atom data on GPU after DD (charges, coordinates, etc.).
255 * TODO: it should update the PME CPU atom data as well.
256 * (currently PME CPU call gmx_pme_do() gets passed the input pointers for each computation).
258 * \param[in,out] pme The PME structure.
259 * \param[in] numAtoms The number of particles.
260 * \param[in] charges The pointer to the array of particle charges.
262 void gmx_pme_reinit_atoms(gmx_pme_t *pme,
264 const real *charges);
266 /* A block of PME GPU functions */
268 /*! \brief Checks whether the GROMACS build allows to run PME on GPU.
269 * TODO: this partly duplicates an internal PME assert function
270 * pme_gpu_check_restrictions(), except that works with a
271 * formed gmx_pme_t structure. Should that one go away/work with inputrec?
273 * \param[out] error If non-null, the error message when PME is not supported on GPU.
275 * \returns true if PME can run on GPU on this build, false otherwise.
277 bool pme_gpu_supports_build(std::string *error);
279 /*! \brief Checks whether the detected (GPU) hardware allows to run PME on GPU.
281 * \param[in] hwinfo Information about the detected hardware
282 * \param[out] error If non-null, the error message when PME is not supported on GPU.
284 * \returns true if PME can run on GPU on this build, false otherwise.
286 bool pme_gpu_supports_hardware(const gmx_hw_info_t &hwinfo,
289 /*! \brief Checks whether the input system allows to run PME on GPU.
290 * TODO: this partly duplicates an internal PME assert function
291 * pme_gpu_check_restrictions(), except that works with a
292 * formed gmx_pme_t structure. Should that one go away/work with inputrec?
294 * \param[in] ir Input system.
295 * \param[in] mtop Complete system topology to check if an FE simulation perturbs charges.
296 * \param[out] error If non-null, the error message if the input is not supported on GPU.
298 * \returns true if PME can run on GPU with this input, false otherwise.
300 bool pme_gpu_supports_input(const t_inputrec &ir, const gmx_mtop_t &mtop, std::string *error);
303 * Returns the active PME codepath (CPU, GPU, mixed).
304 * \todo This is a rather static data that should be managed by the higher level task scheduler.
306 * \param[in] pme The PME data structure.
307 * \returns active PME codepath.
309 PmeRunMode pme_run_mode(const gmx_pme_t *pme);
311 /*! \libinternal \brief
312 * Return the pinning policy appropriate for this build configuration
313 * for relevant buffers used for PME task on this rank (e.g. running
315 gmx::PinningPolicy pme_get_pinning_policy();
318 * Tells if PME is enabled to run on GPU (not necessarily active at the moment).
319 * \todo This is a rather static data that should be managed by the hardware assignment manager.
320 * For now, it is synonymous with the active PME codepath (in the absence of dynamic switching).
322 * \param[in] pme The PME data structure.
323 * \returns true if PME can run on GPU, false otherwise.
325 inline bool pme_gpu_task_enabled(const gmx_pme_t *pme)
327 return (pme != nullptr) && (pme_run_mode(pme) != PmeRunMode::CPU);
330 /*! \brief Returns the size of the padding needed by GPU version of PME in the coordinates array.
332 * \param[in] pme The PME data structure.
334 GPU_FUNC_QUALIFIER int pme_gpu_get_padding_size(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme)) GPU_FUNC_TERM_WITH_RETURN(0);
336 // The following functions are all the PME GPU entry points,
337 // currently inlining to nothing on non-CUDA builds.
340 * Resets the PME GPU timings. To be called at the reset step.
342 * \param[in] pme The PME structure.
344 GPU_FUNC_QUALIFIER void pme_gpu_reset_timings(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme)) GPU_FUNC_TERM;
347 * Copies the PME GPU timings to the gmx_wallclock_gpu_pme_t structure (for log output). To be called at the run end.
349 * \param[in] pme The PME structure.
350 * \param[in] timings The gmx_wallclock_gpu_pme_t structure.
352 GPU_FUNC_QUALIFIER void pme_gpu_get_timings(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
353 gmx_wallclock_gpu_pme_t *GPU_FUNC_ARGUMENT(timings)) GPU_FUNC_TERM;
355 /* The main PME GPU functions */
358 * Prepares PME on GPU computation (updating the box if needed)
359 * \param[in] pme The PME data structure.
360 * \param[in] needToUpdateBox Tells if the stored unit cell parameters should be updated from \p box.
361 * \param[in] box The unit cell box.
362 * \param[in] wcycle The wallclock counter.
363 * \param[in] flags The combination of flags to affect this PME computation.
364 * The flags are the GMX_PME_ flags from pme.h.
365 * \param[in] useGpuForceReduction Whether PME forces are reduced on GPU this step or should be downloaded for CPU reduction
367 GPU_FUNC_QUALIFIER void pme_gpu_prepare_computation(gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
368 bool GPU_FUNC_ARGUMENT(needToUpdateBox),
369 const matrix GPU_FUNC_ARGUMENT(box),
370 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle),
371 int GPU_FUNC_ARGUMENT(flags),
372 bool GPU_FUNC_ARGUMENT(useGpuForceReduction)) GPU_FUNC_TERM;
375 * Launches first stage of PME on GPU - spreading kernel.
377 * \param[in] pme The PME data structure.
378 * \param[in] xReadyOnDevice Event synchronizer indicating that the coordinates are ready in the device memory; nullptr allowed only on separate PME ranks.
379 * \param[in] wcycle The wallclock counter.
381 GPU_FUNC_QUALIFIER void pme_gpu_launch_spread(gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
382 GpuEventSynchronizer *GPU_FUNC_ARGUMENT(xReadyOnDevice),
383 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
386 * Launches middle stages of PME (FFT R2C, solving, FFT C2R) either on GPU or on CPU, depending on the run mode.
388 * \param[in] pme The PME data structure.
389 * \param[in] wcycle The wallclock counter.
391 GPU_FUNC_QUALIFIER void pme_gpu_launch_complex_transforms(gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
392 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
395 * Launches last stage of PME on GPU - force gathering and D2H force transfer.
397 * \param[in] pme The PME data structure.
398 * \param[in] wcycle The wallclock counter.
399 * \param[in] forceTreatment Tells how data should be treated. The gathering kernel either stores
400 * the output reciprocal forces into the host array, or copies its contents to the GPU first
401 * and accumulates. The reduction is non-atomic.
403 GPU_FUNC_QUALIFIER void pme_gpu_launch_gather(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
404 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle),
405 PmeForceOutputHandling GPU_FUNC_ARGUMENT(forceTreatment)) GPU_FUNC_TERM;
408 * Attempts to complete PME GPU tasks.
410 * The \p completionKind argument controls whether the function blocks until all
411 * PME GPU tasks enqueued completed (as pme_gpu_wait_finish_task() does) or only
412 * checks and returns immediately if they did not.
413 * When blocking or the tasks have completed it also gets the output forces
414 * by assigning the ArrayRef to the \p forces pointer passed in.
415 * Virial/energy are also outputs if they were to be computed.
417 * \param[in] pme The PME data structure.
418 * \param[in] flags The combination of flags to affect this PME computation.
419 * The flags are the GMX_PME_ flags from pme.h.
420 * \param[in] wcycle The wallclock counter.
421 * \param[out] forceWithVirial The output force and virial
422 * \param[out] enerd The output energies
423 * \param[in] flags The combination of flags to affect this PME computation.
424 * The flags are the GMX_PME_ flags from pme.h.
425 * \param[in] completionKind Indicates whether PME task completion should only be checked rather than waited for
426 * \returns True if the PME GPU tasks have completed
428 GPU_FUNC_QUALIFIER bool
429 pme_gpu_try_finish_task(gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
430 int GPU_FUNC_ARGUMENT(flags),
431 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle),
432 gmx::ForceWithVirial *GPU_FUNC_ARGUMENT(forceWithVirial),
433 gmx_enerdata_t *GPU_FUNC_ARGUMENT(enerd),
434 GpuTaskCompletion GPU_FUNC_ARGUMENT(completionKind)) GPU_FUNC_TERM_WITH_RETURN(false);
437 * Blocks until PME GPU tasks are completed, and gets the output forces and virial/energy
438 * (if they were to be computed).
440 * \param[in] pme The PME data structure.
441 * \param[in] flags The combination of flags to affect this PME computation.
442 * The flags are the GMX_PME_ flags from pme.h.
443 * \param[in] wcycle The wallclock counter.
444 * \param[out] forceWithVirial The output force and virial
445 * \param[out] enerd The output energies
447 GPU_FUNC_QUALIFIER void
448 pme_gpu_wait_and_reduce(gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
449 int GPU_FUNC_ARGUMENT(flags),
450 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle),
451 gmx::ForceWithVirial *GPU_FUNC_ARGUMENT(forceWithVirial),
452 gmx_enerdata_t *GPU_FUNC_ARGUMENT(enerd)) GPU_FUNC_TERM;
455 * The PME GPU reinitialization function that is called both at the end of any PME computation and on any load balancing.
457 * Clears the internal grid and energy/virial buffers; it is not safe to start
458 * the PME computation without calling this.
459 * Note that unlike in the nbnxn module, the force buffer does not need clearing.
461 * \todo Rename this function to *clear* -- it clearly only does output resetting
462 * and we should be clear about what the function does..
464 * \param[in] pme The PME data structure.
465 * \param[in] wcycle The wallclock counter.
467 GPU_FUNC_QUALIFIER void pme_gpu_reinit_computation(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
468 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
471 /*! \brief Get pointer to device copy of coordinate data.
472 * \param[in] pme The PME data structure.
473 * \returns Pointer to coordinate data
475 GPU_FUNC_QUALIFIER DeviceBuffer<float> pme_gpu_get_device_x(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme)) GPU_FUNC_TERM_WITH_RETURN(DeviceBuffer<float> {});
477 /*! \brief Set pointer to device copy of coordinate data.
478 * \param[in] pme The PME data structure.
479 * \param[in] d_x The pointer to the positions buffer to be set
481 GPU_FUNC_QUALIFIER void pme_gpu_set_device_x(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
482 DeviceBuffer<float> GPU_FUNC_ARGUMENT(d_x)) GPU_FUNC_TERM;
484 /*! \brief Get pointer to device copy of force data.
485 * \param[in] pme The PME data structure.
486 * \returns Pointer to force data
488 GPU_FUNC_QUALIFIER void *pme_gpu_get_device_f(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme)) GPU_FUNC_TERM_WITH_RETURN(nullptr);
490 /*! \brief Returns the pointer to the GPU stream.
491 * \param[in] pme The PME data structure.
492 * \returns Pointer to GPU stream object.
494 GPU_FUNC_QUALIFIER void *pme_gpu_get_device_stream(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme)) GPU_FUNC_TERM_WITH_RETURN(nullptr);
496 /*! \brief Returns the pointer to the GPU context.
497 * \param[in] pme The PME data structure.
498 * \returns Pointer to GPU context object.
500 GPU_FUNC_QUALIFIER void *pme_gpu_get_device_context(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme)) GPU_FUNC_TERM_WITH_RETURN(nullptr);
502 /*! \brief Get pointer to the device synchronizer object that allows syncing on PME force calculation completion
503 * \param[in] pme The PME data structure.
504 * \returns Pointer to sychronizer
506 GPU_FUNC_QUALIFIER GpuEventSynchronizer *pme_gpu_get_f_ready_synchronizer(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme)) GPU_FUNC_TERM_WITH_RETURN(nullptr);