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
52 #include "gromacs/utility/arrayref.h"
54 #include "pme_gpu_types_host.h"
55 #include "pme_output.h"
57 class GpuEventSynchronizer;
58 struct DeviceInformation;
61 struct gmx_pme_t; // only used in pme_gpu_reinit
64 enum class PmeForceOutputHandling;
68 struct PmeGpuSettings;
76 //! Type of spline data
77 enum class PmeSplineDataType
80 Derivatives, // dtheta
81 }; // TODO move this into new and shiny pme.h (pme-types.h?)
83 //! PME grid dimension ordering (from major to minor)
84 enum class GridOrdering
90 /*! \libinternal \brief
91 * Returns the size of the block size requirement
93 * The GPU version of PME requires that the coordinates array have a
94 * size divisible by the returned number.
96 * \returns Number of atoms in a single GPU atom data chunk, which
97 * determines a minimum divisior of the size of the memory allocated.
99 int pme_gpu_get_atom_data_block_size();
101 /*! \libinternal \brief
102 * Synchronizes the current computation, waiting for the GPU kernels/transfers to finish.
104 * \param[in] pmeGpu The PME GPU structure.
106 GPU_FUNC_QUALIFIER void pme_gpu_synchronize(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu)) GPU_FUNC_TERM;
108 /*! \libinternal \brief
109 * Allocates the fixed size energy and virial buffer both on GPU and CPU.
111 * \param[in,out] pmeGpu The PME GPU structure.
113 void pme_gpu_alloc_energy_virial(PmeGpu* pmeGpu);
115 /*! \libinternal \brief
116 * Frees the energy and virial memory both on GPU and CPU.
118 * \param[in] pmeGpu The PME GPU structure.
120 void pme_gpu_free_energy_virial(PmeGpu* pmeGpu);
122 /*! \libinternal \brief
123 * Clears the energy and virial memory on GPU with 0.
124 * Should be called at the end of PME computation which returned energy/virial.
126 * \param[in] pmeGpu The PME GPU structure.
128 void pme_gpu_clear_energy_virial(const PmeGpu* pmeGpu);
130 /*! \libinternal \brief
131 * Reallocates and copies the pre-computed B-spline values to the GPU.
133 * \param[in,out] pmeGpu The PME GPU structure.
135 void pme_gpu_realloc_and_copy_bspline_values(PmeGpu* pmeGpu);
137 /*! \libinternal \brief
138 * Frees the pre-computed B-spline values on the GPU (and the transfer CPU buffers).
140 * \param[in] pmeGpu The PME GPU structure.
142 void pme_gpu_free_bspline_values(const PmeGpu* pmeGpu);
144 /*! \libinternal \brief
145 * Reallocates the GPU buffer for the PME forces.
147 * \param[in] pmeGpu The PME GPU structure.
149 void pme_gpu_realloc_forces(PmeGpu* pmeGpu);
151 /*! \libinternal \brief
152 * Frees the GPU buffer for the PME forces.
154 * \param[in] pmeGpu The PME GPU structure.
156 void pme_gpu_free_forces(const PmeGpu* pmeGpu);
158 /*! \libinternal \brief
159 * Copies the forces from the CPU buffer to the GPU (to reduce them with the PME GPU gathered
160 * forces). To be called e.g. after the bonded calculations.
162 * \param[in] pmeGpu The PME GPU structure.
164 void pme_gpu_copy_input_forces(PmeGpu* pmeGpu);
166 /*! \libinternal \brief
167 * Copies the forces from the GPU to the CPU buffer. To be called after the gathering stage.
169 * \param[in] pmeGpu The PME GPU structure.
171 void pme_gpu_copy_output_forces(PmeGpu* pmeGpu);
173 /*! \libinternal \brief
174 * Checks whether work in the PME GPU stream has completed.
176 * \param[in] pmeGpu The PME GPU structure.
178 * \returns True if work in the PME stream has completed.
180 bool pme_gpu_stream_query(const PmeGpu* pmeGpu);
182 /*! \libinternal \brief
183 * Reallocates the buffer on the GPU and copies the charges/coefficients from the CPU buffer.
184 * Clears the padded part if needed.
186 * \param[in] pmeGpu The PME GPU structure.
187 * \param[in] h_coefficients The input atom charges/coefficients.
189 * Does not need to be done for every PME computation, only whenever the local charges change.
190 * (So, in the beginning of the run, or on DD step).
192 void pme_gpu_realloc_and_copy_input_coefficients(PmeGpu* pmeGpu, const float* h_coefficients);
194 /*! \libinternal \brief
195 * Frees the charges/coefficients on the GPU.
197 * \param[in] pmeGpu The PME GPU structure.
199 void pme_gpu_free_coefficients(const PmeGpu* pmeGpu);
201 /*! \libinternal \brief
202 * Reallocates the buffers on the GPU and the host for the atoms spline data.
204 * \param[in,out] pmeGpu The PME GPU structure.
206 void pme_gpu_realloc_spline_data(PmeGpu* pmeGpu);
208 /*! \libinternal \brief
209 * Frees the buffers on the GPU for the atoms spline data.
211 * \param[in] pmeGpu The PME GPU structure.
213 void pme_gpu_free_spline_data(const PmeGpu* pmeGpu);
215 /*! \libinternal \brief
216 * Reallocates the buffers on the GPU and the host for the particle gridline indices.
218 * \param[in,out] pmeGpu The PME GPU structure.
220 void pme_gpu_realloc_grid_indices(PmeGpu* pmeGpu);
222 /*! \libinternal \brief
223 * Frees the buffer on the GPU for the particle gridline indices.
225 * \param[in] pmeGpu The PME GPU structure.
227 void pme_gpu_free_grid_indices(const PmeGpu* pmeGpu);
229 /*! \libinternal \brief
230 * Reallocates the real space grid and the complex reciprocal grid (if needed) on the GPU.
232 * \param[in] pmeGpu The PME GPU structure.
234 void pme_gpu_realloc_grids(PmeGpu* pmeGpu);
236 /*! \libinternal \brief
237 * Frees the real space grid and the complex reciprocal grid (if needed) on the GPU.
239 * \param[in] pmeGpu The PME GPU structure.
241 void pme_gpu_free_grids(const PmeGpu* pmeGpu);
243 /*! \libinternal \brief
244 * Clears the real space grid on the GPU.
245 * Should be called at the end of each computation.
247 * \param[in] pmeGpu The PME GPU structure.
249 void pme_gpu_clear_grids(const PmeGpu* pmeGpu);
251 /*! \libinternal \brief
252 * Reallocates and copies the pre-computed fractional coordinates' shifts to the GPU.
254 * \param[in] pmeGpu The PME GPU structure.
256 void pme_gpu_realloc_and_copy_fract_shifts(PmeGpu* pmeGpu);
258 /*! \libinternal \brief
259 * Frees the pre-computed fractional coordinates' shifts on the GPU.
261 * \param[in] pmeGpu The PME GPU structure.
263 void pme_gpu_free_fract_shifts(const PmeGpu* pmeGpu);
265 /*! \libinternal \brief
266 * Copies the input real-space grid from the host to the GPU.
268 * \param[in] pmeGpu The PME GPU structure.
269 * \param[in] h_grid The host-side grid buffer.
271 void pme_gpu_copy_input_gather_grid(const PmeGpu* pmeGpu, float* h_grid);
273 /*! \libinternal \brief
274 * Copies the output real-space grid from the GPU to the host.
276 * \param[in] pmeGpu The PME GPU structure.
277 * \param[out] h_grid The host-side grid buffer.
279 void pme_gpu_copy_output_spread_grid(const PmeGpu* pmeGpu, float* h_grid);
281 /*! \libinternal \brief
282 * Copies the spread output spline data and gridline indices from the GPU to the host.
284 * \param[in] pmeGpu The PME GPU structure.
286 void pme_gpu_copy_output_spread_atom_data(const PmeGpu* pmeGpu);
288 /*! \libinternal \brief
289 * Copies the gather input spline data and gridline indices from the host to the GPU.
291 * \param[in] pmeGpu The PME GPU structure.
293 void pme_gpu_copy_input_gather_atom_data(const PmeGpu* pmeGpu);
295 /*! \libinternal \brief
296 * Waits for the grid copying to the host-side buffer after spreading to finish.
298 * \param[in] pmeGpu The PME GPU structure.
300 void pme_gpu_sync_spread_grid(const PmeGpu* pmeGpu);
302 /*! \libinternal \brief
303 * Does the one-time GPU-framework specific PME initialization.
304 * For CUDA, the PME stream is created with the highest priority.
306 * \param[in] pmeGpu The PME GPU structure.
308 void pme_gpu_init_internal(PmeGpu* pmeGpu);
310 /*! \libinternal \brief
311 * Initializes the CUDA FFT structures.
313 * \param[in] pmeGpu The PME GPU structure.
315 void pme_gpu_reinit_3dfft(const PmeGpu* pmeGpu);
317 /*! \libinternal \brief
318 * Destroys the CUDA FFT structures.
320 * \param[in] pmeGpu The PME GPU structure.
322 void pme_gpu_destroy_3dfft(const PmeGpu* pmeGpu);
324 /* The PME stages themselves */
326 /*! \libinternal \brief
327 * A GPU spline computation and charge spreading function.
329 * \param[in] pmeGpu The PME GPU structure.
330 * \param[in] xReadyOnDevice Event synchronizer indicating that the coordinates are ready in the device memory;
331 * can be nullptr when invoked on a separate PME rank or from PME tests.
332 * \param[in] gridIndex Index of the PME grid - unused, assumed to be 0.
333 * \param[out] h_grid The host-side grid buffer (used only if the result of the spread is expected on the host,
334 * e.g. testing or host-side FFT)
335 * \param[in] computeSplines Should the computation of spline parameters and gridline indices be performed.
336 * \param[in] spreadCharges Should the charges/coefficients be spread on the grid.
338 GPU_FUNC_QUALIFIER void pme_gpu_spread(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
339 GpuEventSynchronizer* GPU_FUNC_ARGUMENT(xReadyOnDevice),
340 int GPU_FUNC_ARGUMENT(gridIndex),
341 real* GPU_FUNC_ARGUMENT(h_grid),
342 bool GPU_FUNC_ARGUMENT(computeSplines),
343 bool GPU_FUNC_ARGUMENT(spreadCharges)) GPU_FUNC_TERM;
345 /*! \libinternal \brief
346 * 3D FFT R2C/C2R routine.
348 * \param[in] pmeGpu The PME GPU structure.
349 * \param[in] direction Transform direction (real-to-complex or complex-to-real)
350 * \param[in] gridIndex Index of the PME grid - unused, assumed to be 0.
352 void pme_gpu_3dfft(const PmeGpu* pmeGpu, enum gmx_fft_direction direction, int gridIndex);
354 /*! \libinternal \brief
355 * A GPU Fourier space solving function.
357 * \param[in] pmeGpu The PME GPU structure.
358 * \param[in,out] h_grid The host-side input and output Fourier grid buffer (used only with testing or host-side FFT)
359 * \param[in] gridOrdering Specifies the dimenion ordering of the complex grid. TODO: store this information?
360 * \param[in] computeEnergyAndVirial Tells if the energy and virial computation should be performed.
362 GPU_FUNC_QUALIFIER void pme_gpu_solve(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
363 t_complex* GPU_FUNC_ARGUMENT(h_grid),
364 GridOrdering GPU_FUNC_ARGUMENT(gridOrdering),
365 bool GPU_FUNC_ARGUMENT(computeEnergyAndVirial)) GPU_FUNC_TERM;
367 /*! \libinternal \brief
368 * A GPU force gathering function.
370 * \param[in] pmeGpu The PME GPU structure.
371 * reductions. \param[in] h_grid The host-side grid buffer (used only in testing mode)
373 GPU_FUNC_QUALIFIER void pme_gpu_gather(PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
374 const float* GPU_FUNC_ARGUMENT(h_grid)) GPU_FUNC_TERM;
376 /*! \brief Sets the device pointer to coordinate data
377 * \param[in] pmeGpu The PME GPU structure.
378 * \param[in] d_x Pointer to coordinate data
380 GPU_FUNC_QUALIFIER void pme_gpu_set_kernelparam_coordinates(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
381 DeviceBuffer<gmx::RVec> GPU_FUNC_ARGUMENT(d_x)) GPU_FUNC_TERM;
383 /*! \brief Return pointer to device copy of force data.
384 * \param[in] pmeGpu The PME GPU structure.
385 * \returns Pointer to force data
387 GPU_FUNC_QUALIFIER void* pme_gpu_get_kernelparam_forces(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu))
388 GPU_FUNC_TERM_WITH_RETURN(nullptr);
390 /*! \brief Return pointer to GPU stream.
391 * \param[in] pmeGpu The PME GPU structure.
392 * \returns Pointer to stream object.
394 GPU_FUNC_QUALIFIER const DeviceStream* pme_gpu_get_stream(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu))
395 GPU_FUNC_TERM_WITH_RETURN(nullptr);
397 /*! \brief Return pointer to the sync object triggered after the PME force calculation completion
398 * \param[in] pmeGpu The PME GPU structure.
399 * \returns Pointer to sync object
401 GPU_FUNC_QUALIFIER GpuEventSynchronizer* pme_gpu_get_forces_ready_synchronizer(
402 const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu)) GPU_FUNC_TERM_WITH_RETURN(nullptr);
404 /*! \libinternal \brief
405 * Returns the PME GPU settings
407 * \param[in] pmeGpu The PME GPU structure.
408 * \returns The settings for PME on GPU
410 inline const PmeGpuSettings& pme_gpu_settings(const PmeGpu* pmeGpu)
412 return pmeGpu->settings;
415 /*! \libinternal \brief
416 * Returns the PME GPU staging object
418 * \param[in] pmeGpu The PME GPU structure.
419 * \returns The staging object for PME on GPU
421 inline const PmeGpuStaging& pme_gpu_staging(const PmeGpu* pmeGpu)
423 return pmeGpu->staging;
426 /*! \libinternal \brief
427 * Sets whether the PME module is running in testing mode
429 * \param[in] pmeGpu The PME GPU structure.
430 * \param[in] testing Whether testing mode is on.
432 inline void pme_gpu_set_testing(PmeGpu* pmeGpu, bool testing)
436 pmeGpu->settings.copyAllOutputs = testing;
437 pmeGpu->settings.transferKind = testing ? GpuApiCallBehavior::Sync : GpuApiCallBehavior::Async;
441 /* A block of C++ functions that live in pme_gpu_internal.cpp */
443 /*! \libinternal \brief
444 * Returns the energy and virial GPU outputs, useful for testing.
446 * It is the caller's responsibility to be aware of whether the GPU
447 * handled the solve stage.
449 * \param[in] pme The PME structure.
450 * \param[out] output Pointer to output where energy and virial should be stored.
452 GPU_FUNC_QUALIFIER void pme_gpu_getEnergyAndVirial(const gmx_pme_t& GPU_FUNC_ARGUMENT(pme),
453 PmeOutput* GPU_FUNC_ARGUMENT(output)) GPU_FUNC_TERM;
455 /*! \libinternal \brief
456 * Returns the GPU outputs (forces, energy and virial)
458 * \param[in] pme The PME structure.
459 * \param[in] computeEnergyAndVirial Whether the energy and virial are being computed
460 * \returns The output object.
462 GPU_FUNC_QUALIFIER PmeOutput pme_gpu_getOutput(const gmx_pme_t& GPU_FUNC_ARGUMENT(pme),
463 bool GPU_FUNC_ARGUMENT(computeEnergyAndVirial))
464 GPU_FUNC_TERM_WITH_RETURN(PmeOutput{});
466 /*! \libinternal \brief
467 * Updates the unit cell parameters. Does not check if update is necessary - that is done in pme_gpu_prepare_computation().
469 * \param[in] pmeGpu The PME GPU structure.
470 * \param[in] box The unit cell box.
472 GPU_FUNC_QUALIFIER void pme_gpu_update_input_box(PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
473 const matrix GPU_FUNC_ARGUMENT(box)) GPU_FUNC_TERM;
475 /*! \libinternal \brief
476 * Finishes the PME GPU computation, waiting for the output forces and/or energy/virial to be copied to the host.
477 * If forces were computed, they will have arrived at the external host buffer provided to gather.
478 * If virial/energy were computed, they will have arrived into the internal staging buffer
479 * (even though that should have already happened before even launching the gather).
480 * Finally, cudaEvent_t based GPU timers get updated if enabled. They also need stream synchronization for correctness.
481 * Additionally, device-side buffers are cleared asynchronously for the next computation.
483 * \param[in] pmeGpu The PME GPU structure.
485 void pme_gpu_finish_computation(const PmeGpu* pmeGpu);
487 /*! \libinternal \brief
488 * Get the normal/padded grid dimensions of the real-space PME grid on GPU. Only used in tests.
490 * \param[in] pmeGpu The PME GPU structure.
491 * \param[out] gridSize Pointer to the grid dimensions to fill in.
492 * \param[out] paddedGridSize Pointer to the padded grid dimensions to fill in.
494 GPU_FUNC_QUALIFIER void pme_gpu_get_real_grid_sizes(const PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
495 gmx::IVec* GPU_FUNC_ARGUMENT(gridSize),
496 gmx::IVec* GPU_FUNC_ARGUMENT(paddedGridSize)) GPU_FUNC_TERM;
498 /*! \libinternal \brief
499 * (Re-)initializes the PME GPU data at the beginning of the run or on DLB.
501 * \param[in,out] pme The PME structure.
502 * \param[in] deviceInfo The GPU device information structure.
503 * \param[in] pmeGpuProgram The PME GPU program data
504 * \throws gmx::NotImplementedError if this generally valid PME structure is not valid for GPU runs.
506 GPU_FUNC_QUALIFIER void pme_gpu_reinit(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
507 const DeviceInformation* GPU_FUNC_ARGUMENT(deviceInfo),
508 const PmeGpuProgram* GPU_FUNC_ARGUMENT(pmeGpuProgram)) GPU_FUNC_TERM;
510 /*! \libinternal \brief
511 * Destroys the PME GPU data at the end of the run.
513 * \param[in] pmeGpu The PME GPU structure.
515 GPU_FUNC_QUALIFIER void pme_gpu_destroy(PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu)) GPU_FUNC_TERM;
517 /*! \libinternal \brief
518 * Reallocates the local atoms data (charges, coordinates, etc.). Copies the charges to the GPU.
520 * \param[in] pmeGpu The PME GPU structure.
521 * \param[in] nAtoms The number of particles.
522 * \param[in] charges The pointer to the host-side array of particle charges.
524 * This is a function that should only be called in the beginning of the run and on domain
525 * decomposition. Should be called before the pme_gpu_set_io_ranges.
527 GPU_FUNC_QUALIFIER void pme_gpu_reinit_atoms(PmeGpu* GPU_FUNC_ARGUMENT(pmeGpu),
528 int GPU_FUNC_ARGUMENT(nAtoms),
529 const real* GPU_FUNC_ARGUMENT(charges)) GPU_FUNC_TERM;
531 /*! \brief \libinternal
532 * The PME GPU reinitialization function that is called both at the end of any PME computation and on any load balancing.
534 * This clears the device-side working buffers in preparation for new computation.
536 * \param[in] pmeGpu The PME GPU structure.
538 void pme_gpu_reinit_computation(const PmeGpu* pmeGpu);
541 * Blocks until PME GPU tasks are completed, and gets the output forces and virial/energy
542 * (if they were to be computed).
544 * \param[in] pme The PME data structure.
545 * \param[in] computeEnergyAndVirial Tells if the energy and virial computation should be performed.
546 * \param[out] wcycle The wallclock counter.
547 * \return The output forces, energy and virial
549 GPU_FUNC_QUALIFIER PmeOutput pme_gpu_wait_finish_task(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
550 bool GPU_FUNC_ARGUMENT(computeEnergyAndVirial),
551 gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle))
552 GPU_FUNC_TERM_WITH_RETURN(PmeOutput{});