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, 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/gpu_macros.h"
54 #include "gromacs/math/vectypes.h"
55 #include "gromacs/timing/walltime_accounting.h"
56 #include "gromacs/utility/arrayref.h"
57 #include "gromacs/utility/basedefinitions.h"
58 #include "gromacs/utility/real.h"
61 struct interaction_const_t;
66 struct gmx_wallclock_gpu_pme_t;
67 struct gmx_device_info_t;
68 struct gmx_enerdata_t;
74 enum class GpuTaskCompletion;
77 using PmeGpuProgramHandle = const PmeGpuProgram *;
81 class ForceWithVirial;
83 enum class PinningPolicy : int;
87 GMX_SUM_GRID_FORWARD, GMX_SUM_GRID_BACKWARD
90 /*! \brief Possible PME codepaths on a rank.
91 * \todo: make this enum class with gmx_pme_t C++ refactoring
95 None, //!< No PME task is done
96 CPU, //!< Whole PME computation is done on CPU
97 GPU, //!< Whole PME computation is done on GPU
98 Mixed, //!< Mixed mode: only spread and gather run on GPU; FFT and solving are done on CPU.
101 //! PME gathering output forces treatment
102 enum class PmeForceOutputHandling
104 Set, /**< Gather simply writes into provided force buffer */
105 ReduceWithInput, /**< Gather adds its output to the buffer.
106 On GPU, that means additional H2D copy before the kernel launch. */
109 /*! \brief Return the smallest allowed PME grid size for \p pmeOrder */
110 int minimalPmeGridSize(int pmeOrder);
112 /*! \brief Check restrictions on pme_order and the PME grid nkx,nky,nkz.
114 * With errorsAreFatal=true, an exception or fatal error is generated
115 * on violation of restrictions.
116 * With errorsAreFatal=false, false is returned on violation of restrictions.
117 * When all restrictions are obeyed, true is returned.
118 * Argument useThreads tells if any MPI rank doing PME uses more than 1 threads.
119 * If at calling useThreads is unknown, pass true for conservative checking.
121 * The PME GPU restrictions are checked separately during pme_gpu_init().
123 bool gmx_pme_check_restrictions(int pme_order,
124 int nkx, int nky, int nkz,
125 int numPmeDomainsAlongX,
127 bool errorsAreFatal);
129 /*! \brief Construct PME data
131 * \throws gmx::InconsistentInputError if input grid sizes/PME order are inconsistent.
132 * \returns Pointer to newly allocated and initialized PME data.
134 * \todo We should evolve something like a \c GpuManager that holds \c
135 * gmx_device_info_t * and \c PmeGpuProgramHandle and perhaps other
136 * related things whose lifetime can/should exceed that of a task (or
137 * perhaps task manager). See Redmine #2522.
139 gmx_pme_t *gmx_pme_init(const t_commrec *cr,
140 const NumPmeDomains &numPmeDomains,
141 const t_inputrec *ir, int homenr,
142 gmx_bool bFreeEnergy_q, gmx_bool bFreeEnergy_lj,
143 gmx_bool bReproducible,
144 real ewaldcoeff_q, real ewaldcoeff_lj,
148 const gmx_device_info_t *gpuInfo,
149 PmeGpuProgramHandle pmeGpuProgram,
150 const gmx::MDLogger &mdlog);
152 /*! \brief Destroys the PME data structure.*/
153 void gmx_pme_destroy(gmx_pme_t *pme);
156 /*! \brief Flag values that control what gmx_pme_do() will calculate
158 * These can be combined with bitwise-OR if more than one thing is required.
160 #define GMX_PME_SPREAD (1<<0)
161 #define GMX_PME_SOLVE (1<<1)
162 #define GMX_PME_CALC_F (1<<2)
163 #define GMX_PME_CALC_ENER_VIR (1<<3)
164 /* This forces the grid to be backtransformed even without GMX_PME_CALC_F */
165 #define GMX_PME_CALC_POT (1<<4)
167 #define GMX_PME_DO_ALL_F (GMX_PME_SPREAD | GMX_PME_SOLVE | GMX_PME_CALC_F)
170 /*! \brief Do a PME calculation on a CPU for the long range electrostatics and/or LJ.
172 * The meaning of \p flags is defined above, and determines which
173 * parts of the calculation are performed.
175 * \return 0 indicates all well, non zero is an error code.
177 int gmx_pme_do(struct gmx_pme_t *pme,
178 int start, int homenr,
180 real chargeA[], real chargeB[],
181 real c6A[], real c6B[],
182 real sigmaA[], real sigmaB[],
183 matrix box, const t_commrec *cr,
184 int maxshift_x, int maxshift_y,
185 t_nrnb *nrnb, gmx_wallcycle *wcycle,
186 matrix vir_q, matrix vir_lj,
187 real *energy_q, real *energy_lj,
188 real lambda_q, real lambda_lj,
189 real *dvdlambda_q, real *dvdlambda_lj,
192 /*! \brief Called on the nodes that do PME exclusively */
193 int gmx_pmeonly(struct gmx_pme_t *pme,
194 const t_commrec *cr, t_nrnb *mynrnb,
195 gmx_wallcycle *wcycle,
196 gmx_walltime_accounting_t walltime_accounting,
197 t_inputrec *ir, PmeRunMode runMode);
199 /*! \brief Calculate the PME grid energy V for n charges.
201 * The potential (found in \p pme) must have been found already with a
202 * call to gmx_pme_do() with at least GMX_PME_SPREAD and GMX_PME_SOLVE
203 * specified. Note that the charges are not spread on the grid in the
204 * pme struct. Currently does not work in parallel or with free
207 void gmx_pme_calc_energy(struct gmx_pme_t *pme, int n, rvec *x, real *q, real *V);
209 /*! \brief Send the charges and maxshift to out PME-only node. */
210 void gmx_pme_send_parameters(const t_commrec *cr,
211 const interaction_const_t *ic,
212 gmx_bool bFreeEnergy_q, gmx_bool bFreeEnergy_lj,
213 real *chargeA, real *chargeB,
214 real *sqrt_c6A, real *sqrt_c6B,
215 real *sigmaA, real *sigmaB,
216 int maxshift_x, int maxshift_y);
218 /*! \brief Send the coordinates to our PME-only node and request a PME calculation */
219 void gmx_pme_send_coordinates(const t_commrec *cr, matrix box, rvec *x,
220 real lambda_q, real lambda_lj,
222 int64_t step, gmx_wallcycle *wcycle);
224 /*! \brief Tell our PME-only node to finish */
225 void gmx_pme_send_finish(const t_commrec *cr);
227 /*! \brief Tell our PME-only node to reset all cycle and flop counters */
228 void gmx_pme_send_resetcounters(const t_commrec *cr, int64_t step);
230 /*! \brief PP nodes receive the long range forces from the PME nodes */
231 void gmx_pme_receive_f(const t_commrec *cr,
232 gmx::ForceWithVirial *forceWithVirial,
233 real *energy_q, real *energy_lj,
234 real *dvdlambda_q, real *dvdlambda_lj,
238 * This function updates the local atom data on GPU after DD (charges, coordinates, etc.).
239 * TODO: it should update the PME CPU atom data as well.
240 * (currently PME CPU call gmx_pme_do() gets passed the input pointers for each computation).
242 * \param[in] pme The PME structure.
243 * \param[in] nAtoms The number of particles.
244 * \param[in] charges The pointer to the array of particle charges.
246 void gmx_pme_reinit_atoms(const gmx_pme_t *pme, int nAtoms, const real *charges);
248 /* A block of PME GPU functions */
250 /*! \brief Checks whether the GROMACS build allows to run PME on GPU.
251 * TODO: this partly duplicates an internal PME assert function
252 * pme_gpu_check_restrictions(), except that works with a
253 * formed gmx_pme_t structure. Should that one go away/work with inputrec?
255 * \param[in] hwinfo Information about the detected hardware
256 * \param[out] error If non-null, the error message when PME is not supported on GPU.
258 * \returns true if PME can run on GPU on this build, false otherwise.
260 bool pme_gpu_supports_build(const gmx_hw_info_t &hwinfo,
263 /*! \brief Checks whether the input system allows to run PME on GPU.
264 * TODO: this partly duplicates an internal PME assert function
265 * pme_gpu_check_restrictions(), except that works with a
266 * formed gmx_pme_t structure. Should that one go away/work with inputrec?
268 * \param[in] ir Input system.
269 * \param[in] mtop Complete system topology to check if an FE simulation perturbs charges.
270 * \param[out] error If non-null, the error message if the input is not supported on GPU.
272 * \returns true if PME can run on GPU with this input, false otherwise.
274 bool pme_gpu_supports_input(const t_inputrec &ir, const gmx_mtop_t &mtop, std::string *error);
277 * Returns the active PME codepath (CPU, GPU, mixed).
278 * \todo This is a rather static data that should be managed by the higher level task scheduler.
280 * \param[in] pme The PME data structure.
281 * \returns active PME codepath.
283 PmeRunMode pme_run_mode(const gmx_pme_t *pme);
285 /*! \libinternal \brief
286 * Return the pinning policy appropriate for this build configuration
287 * for relevant buffers used for PME task on this rank (e.g. running
289 gmx::PinningPolicy pme_get_pinning_policy();
292 * Tells if PME is enabled to run on GPU (not necessarily active at the moment).
293 * \todo This is a rather static data that should be managed by the hardware assignment manager.
294 * For now, it is synonymous with the active PME codepath (in the absence of dynamic switching).
296 * \param[in] pme The PME data structure.
297 * \returns true if PME can run on GPU, false otherwise.
299 inline bool pme_gpu_task_enabled(const gmx_pme_t *pme)
301 return (pme != nullptr) && (pme_run_mode(pme) != PmeRunMode::CPU);
304 // The following functions are all the PME GPU entry points,
305 // currently inlining to nothing on non-CUDA builds.
308 * Resets the PME GPU timings. To be called at the reset step.
310 * \param[in] pme The PME structure.
312 GPU_FUNC_QUALIFIER void pme_gpu_reset_timings(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme)) GPU_FUNC_TERM
315 * Copies the PME GPU timings to the gmx_wallclock_gpu_pme_t structure (for log output). To be called at the run end.
317 * \param[in] pme The PME structure.
318 * \param[in] timings The gmx_wallclock_gpu_pme_t structure.
320 GPU_FUNC_QUALIFIER void pme_gpu_get_timings(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
321 gmx_wallclock_gpu_pme_t *GPU_FUNC_ARGUMENT(timings)) GPU_FUNC_TERM
323 /* The main PME GPU functions */
326 * Prepares PME on GPU computation (updating the box if needed)
327 * \param[in] pme The PME data structure.
328 * \param[in] needToUpdateBox Tells if the stored unit cell parameters should be updated from \p box.
329 * \param[in] box The unit cell box.
330 * \param[in] wcycle The wallclock counter.
331 * \param[in] flags The combination of flags to affect this PME computation.
332 * The flags are the GMX_PME_ flags from pme.h.
334 GPU_FUNC_QUALIFIER void pme_gpu_prepare_computation(gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
335 bool GPU_FUNC_ARGUMENT(needToUpdateBox),
336 const matrix GPU_FUNC_ARGUMENT(box),
337 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle),
338 int GPU_FUNC_ARGUMENT(flags)) GPU_FUNC_TERM
341 * Launches first stage of PME on GPU - H2D input transfers, spreading kernel, and D2H grid transfer if needed.
343 * \param[in] pme The PME data structure.
344 * \param[in] x The array of local atoms' coordinates.
345 * \param[in] wcycle The wallclock counter.
347 GPU_FUNC_QUALIFIER void pme_gpu_launch_spread(gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
348 const rvec *GPU_FUNC_ARGUMENT(x),
349 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM
352 * Launches middle stages of PME (FFT R2C, solving, FFT C2R) either on GPU or on CPU, depending on the run mode.
354 * \param[in] pme The PME data structure.
355 * \param[in] wcycle The wallclock counter.
357 GPU_FUNC_QUALIFIER void pme_gpu_launch_complex_transforms(gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
358 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM
361 * Launches last stage of PME on GPU - force gathering and D2H force transfer.
363 * \param[in] pme The PME data structure.
364 * \param[in] wcycle The wallclock counter.
365 * \param[in] forceTreatment Tells how data should be treated. The gathering kernel either stores
366 * the output reciprocal forces into the host array, or copies its contents to the GPU first
367 * and accumulates. The reduction is non-atomic.
369 GPU_FUNC_QUALIFIER void pme_gpu_launch_gather(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
370 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle),
371 PmeForceOutputHandling GPU_FUNC_ARGUMENT(forceTreatment)) GPU_FUNC_TERM
374 * Attempts to complete PME GPU tasks.
376 * The \p completionKind argument controls whether the function blocks until all
377 * PME GPU tasks enqueued completed (as pme_gpu_wait_finish_task() does) or only
378 * checks and returns immediately if they did not.
379 * When blocking or the tasks have completed it also gets the output forces
380 * by assigning the ArrayRef to the \p forces pointer passed in.
381 * Virial/energy are also outputs if they were to be computed.
383 * \param[in] pme The PME data structure.
384 * \param[in] flags The combination of flags to affect this PME computation.
385 * The flags are the GMX_PME_ flags from pme.h.
386 * \param[in] wcycle The wallclock counter.
387 * \param[out] forceWithVirial The output force and virial
388 * \param[out] enerd The output energies
389 * \param[in] flags The combination of flags to affect this PME computation.
390 * The flags are the GMX_PME_ flags from pme.h.
391 * \param[in] completionKind Indicates whether PME task completion should only be checked rather than waited for
392 * \returns True if the PME GPU tasks have completed
394 GPU_FUNC_QUALIFIER bool
395 pme_gpu_try_finish_task(gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
396 int GPU_FUNC_ARGUMENT(flags),
397 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle),
398 gmx::ForceWithVirial *GPU_FUNC_ARGUMENT(forceWithVirial),
399 gmx_enerdata_t *GPU_FUNC_ARGUMENT(enerd),
400 GpuTaskCompletion GPU_FUNC_ARGUMENT(completionKind)) GPU_FUNC_TERM_WITH_RETURN(false)
403 * Blocks until PME GPU tasks are completed, and gets the output forces and virial/energy
404 * (if they were to be computed).
406 * \param[in] pme The PME data structure.
407 * \param[in] flags The combination of flags to affect this PME computation.
408 * The flags are the GMX_PME_ flags from pme.h.
409 * \param[in] wcycle The wallclock counter.
410 * \param[out] forceWithVirial The output force and virial
411 * \param[out] enerd The output energies
413 GPU_FUNC_QUALIFIER void
414 pme_gpu_wait_and_reduce(gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
415 int GPU_FUNC_ARGUMENT(flags),
416 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle),
417 gmx::ForceWithVirial *GPU_FUNC_ARGUMENT(forceWithVirial),
418 gmx_enerdata_t *GPU_FUNC_ARGUMENT(enerd)) GPU_FUNC_TERM
421 * The PME GPU reinitialization function that is called both at the end of any PME computation and on any load balancing.
423 * Clears the internal grid and energy/virial buffers; it is not safe to start
424 * the PME computation without calling this.
425 * Note that unlike in the nbnxn module, the force buffer does not need clearing.
427 * \todo Rename this function to *clear* -- it clearly only does output resetting
428 * and we should be clear about what the function does..
430 * \param[in] pme The PME data structure.
431 * \param[in] wcycle The wallclock counter.
433 GPU_FUNC_QUALIFIER void pme_gpu_reinit_computation(const gmx_pme_t *GPU_FUNC_ARGUMENT(pme),
434 gmx_wallcycle *GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM