2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2016,2017,2018,2019,2020, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
38 * \brief This file contains internal function definitions for performing the PME calculations on GPU.
39 * These are not meant to be exposed outside of the PME GPU code.
40 * As of now, their bodies are still in the common pme_gpu.cpp files.
42 * \author Aleksei Iupinov <a.yupinov@gmail.com>
43 * \ingroup module_ewald
46 #ifndef GMX_EWALD_PME_GPU_INTERNAL_H
47 #define GMX_EWALD_PME_GPU_INTERNAL_H
49 #include "gromacs/fft/fft.h" // for the gmx_fft_direction enum
50 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
51 #include "gromacs/gpu_utils/gpu_macros.h" // for the GPU_FUNC_ macros
53 #include "pme_gpu_types_host.h"
54 #include "pme_output.h"
57 struct DeviceInformation;
59 class GpuEventSynchronizer;
62 struct gmx_pme_t; // only used in pme_gpu_reinit
65 enum class PmeForceOutputHandling;
69 struct PmeGpuSettings;
79 //! Type of spline data
80 enum class PmeSplineDataType
83 Derivatives, // dtheta
84 }; // TODO move this into new and shiny pme.h (pme-types.h?)
86 //! PME grid dimension ordering (from major to minor)
87 enum class GridOrdering
93 /*! \libinternal \brief
94 * Returns the size of the block size requirement
96 * The GPU version of PME requires that the coordinates array have a
97 * size divisible by the returned number.
99 * \returns Number of atoms in a single GPU atom data chunk, which
100 * determines a minimum divisior of the size of the memory allocated.
102 int pme_gpu_get_atom_data_block_size();
104 /*! \libinternal \brief
105 * Synchronizes the current computation, waiting for the GPU kernels/transfers to finish.
107 * \param[in] pmeGpu The PME GPU structure.
109 GPU_FUNC_QUALIFIER void pme_gpu_synchronize(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu)) GPU_FUNC_TERM;
111 /*! \libinternal \brief
112 * Allocates the fixed size energy and virial buffer both on GPU and CPU.
114 * \param[in,out] pmeGpu The PME GPU structure.
116 void pme_gpu_alloc_energy_virial(PmeGpu* pmeGpu);
118 /*! \libinternal \brief
119 * Frees the energy and virial memory both on GPU and CPU.
121 * \param[in] pmeGpu The PME GPU structure.
123 void pme_gpu_free_energy_virial(PmeGpu* pmeGpu);
125 /*! \libinternal \brief
126 * Clears the energy and virial memory on GPU with 0.
127 * Should be called at the end of PME computation which returned energy/virial.
129 * \param[in] pmeGpu The PME GPU structure.
131 void pme_gpu_clear_energy_virial(const PmeGpu* pmeGpu);
133 /*! \libinternal \brief
134 * Reallocates and copies the pre-computed B-spline values to the GPU.
136 * \param[in,out] pmeGpu The PME GPU structure.
138 void pme_gpu_realloc_and_copy_bspline_values(PmeGpu* pmeGpu);
140 /*! \libinternal \brief
141 * Frees the pre-computed B-spline values on the GPU (and the transfer CPU buffers).
143 * \param[in] pmeGpu The PME GPU structure.
145 void pme_gpu_free_bspline_values(const PmeGpu* pmeGpu);
147 /*! \libinternal \brief
148 * Reallocates the GPU buffer for the PME forces.
150 * \param[in] pmeGpu The PME GPU structure.
152 void pme_gpu_realloc_forces(PmeGpu* pmeGpu);
154 /*! \libinternal \brief
155 * Frees the GPU buffer for the PME forces.
157 * \param[in] pmeGpu The PME GPU structure.
159 void pme_gpu_free_forces(const PmeGpu* pmeGpu);
161 /*! \libinternal \brief
162 * Copies the forces from the CPU buffer to the GPU (to reduce them with the PME GPU gathered
163 * forces). To be called e.g. after the bonded calculations.
165 * \param[in] pmeGpu The PME GPU structure.
167 void pme_gpu_copy_input_forces(PmeGpu* pmeGpu);
169 /*! \libinternal \brief
170 * Copies the forces from the GPU to the CPU buffer. To be called after the gathering stage.
172 * \param[in] pmeGpu The PME GPU structure.
174 void pme_gpu_copy_output_forces(PmeGpu* pmeGpu);
176 /*! \libinternal \brief
177 * Checks whether work in the PME GPU stream has completed.
179 * \param[in] pmeGpu The PME GPU structure.
181 * \returns True if work in the PME stream has completed.
183 bool pme_gpu_stream_query(const PmeGpu* pmeGpu);
185 /*! \libinternal \brief
186 * Reallocates the buffer on the GPU and copies the charges/coefficients from the CPU buffer.
187 * Clears the padded part if needed.
189 * \param[in] pmeGpu The PME GPU structure.
190 * \param[in] h_coefficients The input atom charges/coefficients.
192 * Does not need to be done for every PME computation, only whenever the local charges change.
193 * (So, in the beginning of the run, or on DD step).
195 void pme_gpu_realloc_and_copy_input_coefficients(PmeGpu* pmeGpu, const float* h_coefficients);
197 /*! \libinternal \brief
198 * Frees the charges/coefficients on the GPU.
200 * \param[in] pmeGpu The PME GPU structure.
202 void pme_gpu_free_coefficients(const PmeGpu* pmeGpu);
204 /*! \libinternal \brief
205 * Reallocates the buffers on the GPU and the host for the atoms spline data.
207 * \param[in,out] pmeGpu The PME GPU structure.
209 void pme_gpu_realloc_spline_data(PmeGpu* pmeGpu);
211 /*! \libinternal \brief
212 * Frees the buffers on the GPU for the atoms spline data.
214 * \param[in] pmeGpu The PME GPU structure.
216 void pme_gpu_free_spline_data(const PmeGpu* pmeGpu);
218 /*! \libinternal \brief
219 * Reallocates the buffers on the GPU and the host for the particle gridline indices.
221 * \param[in,out] pmeGpu The PME GPU structure.
223 void pme_gpu_realloc_grid_indices(PmeGpu* pmeGpu);
225 /*! \libinternal \brief
226 * Frees the buffer on the GPU for the particle gridline indices.
228 * \param[in] pmeGpu The PME GPU structure.
230 void pme_gpu_free_grid_indices(const PmeGpu* pmeGpu);
232 /*! \libinternal \brief
233 * Reallocates the real space grid and the complex reciprocal grid (if needed) on the GPU.
235 * \param[in] pmeGpu The PME GPU structure.
237 void pme_gpu_realloc_grids(PmeGpu* pmeGpu);
239 /*! \libinternal \brief
240 * Frees the real space grid and the complex reciprocal grid (if needed) on the GPU.
242 * \param[in] pmeGpu The PME GPU structure.
244 void pme_gpu_free_grids(const PmeGpu* pmeGpu);
246 /*! \libinternal \brief
247 * Clears the real space grid on the GPU.
248 * Should be called at the end of each computation.
250 * \param[in] pmeGpu The PME GPU structure.
252 void pme_gpu_clear_grids(const PmeGpu* pmeGpu);
254 /*! \libinternal \brief
255 * Reallocates and copies the pre-computed fractional coordinates' shifts to the GPU.
257 * \param[in] pmeGpu The PME GPU structure.
259 void pme_gpu_realloc_and_copy_fract_shifts(PmeGpu* pmeGpu);
261 /*! \libinternal \brief
262 * Frees the pre-computed fractional coordinates' shifts on the GPU.
264 * \param[in] pmeGpu The PME GPU structure.
266 void pme_gpu_free_fract_shifts(const PmeGpu* pmeGpu);
268 /*! \libinternal \brief
269 * Copies the input real-space grid from the host to the GPU.
271 * \param[in] pmeGpu The PME GPU structure.
272 * \param[in] h_grid The host-side grid buffer.
274 void pme_gpu_copy_input_gather_grid(const PmeGpu* pmeGpu, float* h_grid);
276 /*! \libinternal \brief
277 * Copies the output real-space grid from the GPU to the host.
279 * \param[in] pmeGpu The PME GPU structure.
280 * \param[out] h_grid The host-side grid buffer.
282 void pme_gpu_copy_output_spread_grid(const PmeGpu* pmeGpu, float* h_grid);
284 /*! \libinternal \brief
285 * Copies the spread output spline data and gridline indices from the GPU to the host.
287 * \param[in] pmeGpu The PME GPU structure.
289 void pme_gpu_copy_output_spread_atom_data(const PmeGpu* pmeGpu);
291 /*! \libinternal \brief
292 * Copies the gather input spline data and gridline indices from the host to the GPU.
294 * \param[in] pmeGpu The PME GPU structure.
296 void pme_gpu_copy_input_gather_atom_data(const PmeGpu* pmeGpu);
298 /*! \libinternal \brief
299 * Waits for the grid copying to the host-side buffer after spreading to finish.
301 * \param[in] pmeGpu The PME GPU structure.
303 void pme_gpu_sync_spread_grid(const PmeGpu* pmeGpu);
305 /*! \libinternal \brief
306 * Initializes the CUDA FFT structures.
308 * \param[in] pmeGpu The PME GPU structure.
310 void pme_gpu_reinit_3dfft(const PmeGpu* pmeGpu);
312 /*! \libinternal \brief
313 * Destroys the CUDA FFT structures.
315 * \param[in] pmeGpu The PME GPU structure.
317 void pme_gpu_destroy_3dfft(const PmeGpu* pmeGpu);
319 /* The PME stages themselves */
321 /*! \libinternal \brief
322 * A GPU spline computation and charge spreading function.
324 * \param[in] pmeGpu The PME GPU structure.
325 * \param[in] xReadyOnDevice Event synchronizer indicating that the coordinates are ready in the device memory;
326 * can be nullptr when invoked on a separate PME rank or from PME tests.
327 * \param[in] gridIndex Index of the PME grid - unused, assumed to be 0.
328 * \param[out] h_grid The host-side grid buffer (used only if the result of the spread is expected on the host,
329 * e.g. testing or host-side FFT)
330 * \param[in] computeSplines Should the computation of spline parameters and gridline indices be performed.
331 * \param[in] spreadCharges Should the charges/coefficients be spread on the grid.
333 GPU_FUNC_QUALIFIER void pme_gpu_spread(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
334 GpuEventSynchronizer* GPU_FUNC_ARGUMENT(xReadyOnDevice),
335 int GPU_FUNC_ARGUMENT(gridIndex),
336 real* GPU_FUNC_ARGUMENT(h_grid),
337 bool GPU_FUNC_ARGUMENT(computeSplines),
338 bool GPU_FUNC_ARGUMENT(spreadCharges)) GPU_FUNC_TERM;
340 /*! \libinternal \brief
341 * 3D FFT R2C/C2R routine.
343 * \param[in] pmeGpu The PME GPU structure.
344 * \param[in] direction Transform direction (real-to-complex or complex-to-real)
345 * \param[in] gridIndex Index of the PME grid - unused, assumed to be 0.
347 void pme_gpu_3dfft(const PmeGpu* pmeGpu, enum gmx_fft_direction direction, int gridIndex);
349 /*! \libinternal \brief
350 * A GPU Fourier space solving function.
352 * \param[in] pmeGpu The PME GPU structure.
353 * \param[in,out] h_grid The host-side input and output Fourier grid buffer (used only with testing or host-side FFT)
354 * \param[in] gridOrdering Specifies the dimenion ordering of the complex grid. TODO: store this information?
355 * \param[in] computeEnergyAndVirial Tells if the energy and virial computation should be performed.
357 GPU_FUNC_QUALIFIER void pme_gpu_solve(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
358 t_complex* GPU_FUNC_ARGUMENT(h_grid),
359 GridOrdering GPU_FUNC_ARGUMENT(gridOrdering),
360 bool GPU_FUNC_ARGUMENT(computeEnergyAndVirial)) GPU_FUNC_TERM;
362 /*! \libinternal \brief
363 * A GPU force gathering function.
365 * \param[in] pmeGpu The PME GPU structure.
366 * reductions. \param[in] h_grid The host-side grid buffer (used only in testing mode)
368 GPU_FUNC_QUALIFIER void pme_gpu_gather(PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
369 const float* GPU_FUNC_ARGUMENT(h_grid)) GPU_FUNC_TERM;
371 /*! \brief Sets the device pointer to coordinate data
372 * \param[in] pmeGpu The PME GPU structure.
373 * \param[in] d_x Pointer to coordinate data
375 GPU_FUNC_QUALIFIER void pme_gpu_set_kernelparam_coordinates(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
376 DeviceBuffer<gmx::RVec> GPU_FUNC_ARGUMENT(d_x)) GPU_FUNC_TERM;
378 /*! \brief Return pointer to device copy of force data.
379 * \param[in] pmeGpu The PME GPU structure.
380 * \returns Pointer to force data
382 GPU_FUNC_QUALIFIER void* pme_gpu_get_kernelparam_forces(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu))
383 GPU_FUNC_TERM_WITH_RETURN(nullptr);
385 /*! \brief Return pointer to the sync object triggered after the PME force calculation completion
386 * \param[in] pmeGpu The PME GPU structure.
387 * \returns Pointer to sync object
389 GPU_FUNC_QUALIFIER GpuEventSynchronizer* pme_gpu_get_forces_ready_synchronizer(
390 const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu)) GPU_FUNC_TERM_WITH_RETURN(nullptr);
392 /*! \libinternal \brief
393 * Returns the PME GPU settings
395 * \param[in] pmeGpu The PME GPU structure.
396 * \returns The settings for PME on GPU
398 inline const PmeGpuSettings& pme_gpu_settings(const PmeGpu* pmeGpu)
400 return pmeGpu->settings;
403 /*! \libinternal \brief
404 * Returns the PME GPU staging object
406 * \param[in] pmeGpu The PME GPU structure.
407 * \returns The staging object for PME on GPU
409 inline const PmeGpuStaging& pme_gpu_staging(const PmeGpu* pmeGpu)
411 return pmeGpu->staging;
414 /*! \libinternal \brief
415 * Sets whether the PME module is running in testing mode
417 * \param[in] pmeGpu The PME GPU structure.
418 * \param[in] testing Whether testing mode is on.
420 inline void pme_gpu_set_testing(PmeGpu* pmeGpu, bool testing)
424 pmeGpu->settings.copyAllOutputs = testing;
425 pmeGpu->settings.transferKind = testing ? GpuApiCallBehavior::Sync : GpuApiCallBehavior::Async;
429 /* A block of C++ functions that live in pme_gpu_internal.cpp */
431 /*! \libinternal \brief
432 * Returns the energy and virial GPU outputs, useful for testing.
434 * It is the caller's responsibility to be aware of whether the GPU
435 * handled the solve stage.
437 * \param[in] pme The PME structure.
438 * \param[out] output Pointer to output where energy and virial should be stored.
440 GPU_FUNC_QUALIFIER void pme_gpu_getEnergyAndVirial(const gmx_pme_t& GPU_FUNC_ARGUMENT(pme),
441 PmeOutput* GPU_FUNC_ARGUMENT(output)) GPU_FUNC_TERM;
443 /*! \libinternal \brief
444 * Returns the GPU outputs (forces, energy and virial)
446 * \param[in] pme The PME structure.
447 * \param[in] computeEnergyAndVirial Whether the energy and virial are being computed
448 * \returns The output object.
450 GPU_FUNC_QUALIFIER PmeOutput pme_gpu_getOutput(const gmx_pme_t& GPU_FUNC_ARGUMENT(pme),
451 bool GPU_FUNC_ARGUMENT(computeEnergyAndVirial))
452 GPU_FUNC_TERM_WITH_RETURN(PmeOutput{});
454 /*! \libinternal \brief
455 * Updates the unit cell parameters. Does not check if update is necessary - that is done in pme_gpu_prepare_computation().
457 * \param[in] pmeGpu The PME GPU structure.
458 * \param[in] box The unit cell box.
460 GPU_FUNC_QUALIFIER void pme_gpu_update_input_box(PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
461 const matrix GPU_FUNC_ARGUMENT(box)) GPU_FUNC_TERM;
463 /*! \libinternal \brief
464 * Finishes the PME GPU computation, waiting for the output forces and/or energy/virial to be copied to the host.
465 * If forces were computed, they will have arrived at the external host buffer provided to gather.
466 * If virial/energy were computed, they will have arrived into the internal staging buffer
467 * (even though that should have already happened before even launching the gather).
468 * Finally, cudaEvent_t based GPU timers get updated if enabled. They also need stream synchronization for correctness.
469 * Additionally, device-side buffers are cleared asynchronously for the next computation.
471 * \param[in] pmeGpu The PME GPU structure.
473 void pme_gpu_finish_computation(const PmeGpu* pmeGpu);
475 /*! \libinternal \brief
476 * Get the normal/padded grid dimensions of the real-space PME grid on GPU. Only used in tests.
478 * \param[in] pmeGpu The PME GPU structure.
479 * \param[out] gridSize Pointer to the grid dimensions to fill in.
480 * \param[out] paddedGridSize Pointer to the padded grid dimensions to fill in.
482 GPU_FUNC_QUALIFIER void pme_gpu_get_real_grid_sizes(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
483 gmx::IVec* GPU_FUNC_ARGUMENT(gridSize),
484 gmx::IVec* GPU_FUNC_ARGUMENT(paddedGridSize)) GPU_FUNC_TERM;
486 /*! \libinternal \brief
487 * (Re-)initializes the PME GPU data at the beginning of the run or on DLB.
489 * \param[in,out] pme The PME structure.
490 * \param[in] deviceContext The GPU context.
491 * \param[in] deviceStream The GPU stream.
492 * \param[in,out] pmeGpuProgram The handle to the program/kernel data created outside (e.g. in unit tests/runner)
494 * \throws gmx::NotImplementedError if this generally valid PME structure is not valid for GPU runs.
496 GPU_FUNC_QUALIFIER void pme_gpu_reinit(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
497 const DeviceContext* GPU_FUNC_ARGUMENT(deviceContext),
498 const DeviceStream* GPU_FUNC_ARGUMENT(deviceStream),
499 const PmeGpuProgram* GPU_FUNC_ARGUMENT(pmeGpuProgram)) GPU_FUNC_TERM;
501 /*! \libinternal \brief
502 * Destroys the PME GPU data at the end of the run.
504 * \param[in] pmeGpu The PME GPU structure.
506 GPU_FUNC_QUALIFIER void pme_gpu_destroy(PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu)) GPU_FUNC_TERM;
508 /*! \libinternal \brief
509 * Reallocates the local atoms data (charges, coordinates, etc.). Copies the charges to the GPU.
511 * \param[in] pmeGpu The PME GPU structure.
512 * \param[in] nAtoms The number of particles.
513 * \param[in] charges The pointer to the host-side array of particle charges.
515 * This is a function that should only be called in the beginning of the run and on domain
516 * decomposition. Should be called before the pme_gpu_set_io_ranges.
518 GPU_FUNC_QUALIFIER void pme_gpu_reinit_atoms(PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
519 int GPU_FUNC_ARGUMENT(nAtoms),
520 const real* GPU_FUNC_ARGUMENT(charges)) GPU_FUNC_TERM;
522 /*! \brief \libinternal
523 * The PME GPU reinitialization function that is called both at the end of any PME computation and on any load balancing.
525 * This clears the device-side working buffers in preparation for new computation.
527 * \param[in] pmeGpu The PME GPU structure.
529 void pme_gpu_reinit_computation(const PmeGpu* pmeGpu);
532 * Blocks until PME GPU tasks are completed, and gets the output forces and virial/energy
533 * (if they were to be computed).
535 * \param[in] pme The PME data structure.
536 * \param[in] computeEnergyAndVirial Tells if the energy and virial computation should be performed.
537 * \param[out] wcycle The wallclock counter.
538 * \return The output forces, energy and virial
540 GPU_FUNC_QUALIFIER PmeOutput pme_gpu_wait_finish_task(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
541 bool GPU_FUNC_ARGUMENT(computeEnergyAndVirial),
542 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle))
543 GPU_FUNC_TERM_WITH_RETURN(PmeOutput{});