2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2017,2018,2019, 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.
36 // FIXME: remove the "__" prefix in front of the group def when we move the
37 // nonbonded code into separate dir.
39 /*! \libinternal \defgroup __module_nbnxm Short-range non-bonded interaction module
40 * \ingroup group_mdrun
42 * \brief Computes forces and energies for short-range pair-interactions
43 * based on the Verlet algorithm. The algorithm uses pair-lists generated
44 * at fixed intervals as well as various flavors of pair interaction kernels
45 * implemented for a wide range of CPU and GPU architectures.
47 * The module includes support for flavors of Coulomb and Lennard-Jones interaction
48 * treatment implemented for a large range of SIMD instruction sets for CPU
49 * architectures as well as in CUDA and OpenCL for GPU architectures.
50 * Additionally there is a reference CPU non-SIMD and a reference CPU
51 * for GPU pair-list setup interaction kernel.
53 * The implementation of the kernels is based on the cluster non-bonded algorithm
54 * which in the code is referred to as the NxM algorithms ("nbnxm_" prefix);
55 * for details of the algorithm see DOI:10.1016/j.cpc.2013.06.003.
57 * Algorithmically, the non-bonded computation has two different modes:
58 * A "classical" mode: generate a list every nstlist steps containing at least
59 * all atom pairs up to a distance of rlistOuter and compute pair interactions
60 * for all pairs that are within the interaction cut-off.
61 * A "dynamic pruning" mode: generate an "outer-list" up to cut-off rlistOuter
62 * every nstlist steps and prune the outer-list using a cut-off of rlistInner
63 * every nstlistPrune steps to obtain a, smaller, "inner-list". This
64 * results in fewer interaction computations and allows for a larger nstlist.
65 * On a GPU, this dynamic pruning is performed in a rolling fashion, pruning
66 * only a sub-part of the list each (second) step. This way it can often
67 * overlap with integration and constraints on the CPU.
68 * Currently a simple heuristic determines which mode will be used.
70 * TODO: add a summary list and brief descriptions of the different submodules:
71 * search, CPU kernels, GPU glue code + kernels.
73 * \author Berk Hess <hess@kth.se>
74 * \author Szilárd Páll <pall.szilard@gmail.com>
75 * \author Mark Abraham <mark.j.abraham@gmail.com>
76 * \author Anca Hamuraru <anca@streamcomputing.eu>
77 * \author Teemu Virolainen <teemu@streamcomputing.eu>
78 * \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
80 * TODO: add more authors!
83 /*! \libinternal \file
85 * \brief This file contains the public interface of the nbnxm module
86 * that implements the NxM atom cluster non-bonded algorithm to efficiently
87 * compute pair forces.
90 * \author Berk Hess <hess@kth.se>
91 * \author Szilárd Páll <pall.szilard@gmail.com>
94 * \ingroup __module_nbnxm
98 #ifndef GMX_NBNXM_NBNXM_H
99 #define GMX_NBNXM_NBNXM_H
103 #include "gromacs/math/vectypes.h"
104 #include "gromacs/utility/arrayref.h"
105 #include "gromacs/utility/enumerationhelpers.h"
106 #include "gromacs/utility/real.h"
108 #include "locality.h"
110 // TODO: Remove this include and the two nbnxm includes above
111 #include "nbnxm_gpu.h"
113 struct gmx_device_info_t;
114 struct gmx_domdec_zones_t;
115 struct gmx_enerdata_t;
116 struct gmx_hw_info_t;
118 struct gmx_wallcycle;
119 struct interaction_const_t;
120 struct nonbonded_verlet_t;
131 /*! \brief Switch for whether to use GPU for buffer ops*/
132 enum class BufferOpsUseGpu
138 class GpuEventSynchronizer;
143 class UpdateGroupsCog;
148 enum class KernelType;
154 /*! \brief Nonbonded NxN kernel types: plain C, CPU SIMD, GPU, GPU emulation */
155 enum class KernelType : int
166 /*! \brief Ewald exclusion types */
167 enum class EwaldExclusionType : int
175 /* \brief The non-bonded setup, also affects the pairlist construction kernel */
178 //! The non-bonded type, also affects the pairlist construction kernel
179 KernelType kernelType = KernelType::NotSet;
180 //! Ewald exclusion computation handling type, currently only used for CPU
181 EwaldExclusionType ewaldExclusionType = EwaldExclusionType::NotSet;
184 /*! \brief Return a string identifying the kernel type.
186 * \param [in] kernelType nonbonded kernel type, takes values from the nbnxn_kernel_type enum
187 * \returns a string identifying the kernel corresponding to the type passed as argument
189 const char *lookup_kernel_name(Nbnxm::KernelType kernelType);
193 /*! \brief Flag to tell the nonbonded kernels whether to clear the force output buffers */
195 enbvClearFNo, enbvClearFYes
199 * \brief Top-level non-bonded data structure for the Verlet-type cut-off scheme. */
200 struct nonbonded_verlet_t
203 //! Constructs an object from its components
204 nonbonded_verlet_t(std::unique_ptr<PairlistSets> pairlistSets,
205 std::unique_ptr<PairSearch> pairSearch,
206 std::unique_ptr<nbnxn_atomdata_t> nbat,
207 const Nbnxm::KernelSetup &kernelSetup,
208 gmx_nbnxn_gpu_t *gpu_nbv,
209 gmx_wallcycle *wcycle);
211 ~nonbonded_verlet_t();
213 //! Returns whether a GPU is use for the non-bonded calculations
216 return kernelSetup_.kernelType == Nbnxm::KernelType::Gpu8x8x8;
219 //! Returns whether a GPU is emulated for the non-bonded calculations
220 bool emulateGpu() const
222 return kernelSetup_.kernelType == Nbnxm::KernelType::Cpu8x8x8_PlainC;
225 //! Return whether the pairlist is of simple, CPU type
226 bool pairlistIsSimple() const
228 return !useGpu() && !emulateGpu();
231 //! Initialize the pair list sets, TODO this should be private
232 void initPairlistSets(bool haveMultipleDomains);
234 //! Returns the order of the local atoms on the grid
235 gmx::ArrayRef<const int> getLocalAtomOrder() const;
237 //! Sets the order of the local atoms to the order grid atom ordering
238 void setLocalAtomOrder();
240 //! Returns the index position of the atoms on the search grid
241 gmx::ArrayRef<const int> getGridIndices() const;
243 //! Constructs the pairlist for the given locality
244 void constructPairlist(Nbnxm::InteractionLocality iLocality,
245 const t_blocka *excl,
249 //! Updates all the atom properties in Nbnxm
250 void setAtomProperties(const t_mdatoms &mdatoms,
251 gmx::ArrayRef<const int> atomInfo);
253 //! Updates the coordinates in Nbnxm for the given locality
254 void setCoordinates(Nbnxm::AtomLocality locality,
256 gmx::ArrayRef<const gmx::RVec> x,
257 BufferOpsUseGpu useGpu,
258 void *xPmeDevicePtr);
260 //! Init for GPU version of setup coordinates in Nbnxm
261 void atomdata_init_copy_x_to_nbat_x_gpu();
263 //! Sync the nonlocal GPU stream with dependent tasks in the local queue.
264 void insertNonlocalGpuDependency(Nbnxm::InteractionLocality interactionLocality);
266 //! Returns a reference to the pairlist sets
267 const PairlistSets &pairlistSets() const
269 return *pairlistSets_;
272 //! Returns whether step is a dynamic list pruning step, for CPU lists
273 bool isDynamicPruningStepCpu(int64_t step) const;
275 //! Returns whether step is a dynamic list pruning step, for GPU lists
276 bool isDynamicPruningStepGpu(int64_t step) const;
278 //! Dispatches the dynamic pruning kernel for the given locality, for CPU lists
279 void dispatchPruneKernelCpu(Nbnxm::InteractionLocality iLocality,
280 const rvec *shift_vec);
282 //! Dispatches the dynamic pruning kernel for GPU lists
283 void dispatchPruneKernelGpu(int64_t step);
285 //! \brief Executes the non-bonded kernel of the GPU or launches it on the GPU
286 void dispatchNonbondedKernel(Nbnxm::InteractionLocality iLocality,
287 const interaction_const_t &ic,
290 const t_forcerec &fr,
291 gmx_enerdata_t *enerd,
294 //! Executes the non-bonded free-energy kernel, always runs on the CPU
295 void dispatchFreeEnergyKernel(Nbnxm::InteractionLocality iLocality,
299 const t_mdatoms &mdatoms,
302 gmx_enerdata_t *enerd,
306 /*! \brief Add the forces stored in nbat to f, zeros the forces in nbat
307 * \param [in] locality Local or non-local
308 * \param [inout] f Force to be added to
310 void atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality locality,
313 /*! \brief Add the forces stored in nbat to f, allowing for possibility that GPU buffer ops are active
314 * \param [in] locality Local or non-local
315 * \param [inout] f Force to be added to
316 * \param [in] fPme Force from PME calculation
317 * \param [in] pmeForcesReady Event triggered when PME force calculation has completed
318 * \param [in] useGpu Whether GPU buffer ops are active
319 * \param [in] useGpuFPmeReduction Whether PME force reduction is on GPU
320 * \param [in] accumulateForce Whether force should be accumulated or stored
322 void atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality locality,
325 GpuEventSynchronizer *pmeForcesReady,
326 BufferOpsUseGpu useGpu,
327 bool useGpuFPmeReduction,
328 bool accumulateForce);
330 /*! \brief Outer body of function to perform initialization for F buffer operations on GPU. */
331 void atomdata_init_add_nbat_f_to_f_gpu();
333 /*! \brief H2D transfer of force buffer*/
334 void launch_copy_f_to_gpu(rvec *f, Nbnxm::AtomLocality locality);
336 /*! \brief D2H transfer of force buffer*/
337 void launch_copy_f_from_gpu(rvec *f, Nbnxm::AtomLocality locality);
339 /*! \brief Wait for GPU force reduction task and D2H transfer of its results to complete
341 * FIXME: need more details: when should be called / after which operation, etc.
343 void wait_for_gpu_force_reduction(Nbnxm::AtomLocality locality);
345 //! Return the kernel setup
346 const Nbnxm::KernelSetup &kernelSetup() const
351 //! Returns the outer radius for the pair list
352 real pairlistInnerRadius() const;
354 //! Returns the outer radius for the pair list
355 real pairlistOuterRadius() const;
357 //! Changes the pair-list outer and inner radius
358 void changePairlistRadii(real rlistOuter,
361 //! Set up internal flags that indicate what type of short-range work there is.
362 void setupGpuShortRangeWork(const gmx::GpuBonded *gpuBonded,
363 const Nbnxm::InteractionLocality iLocality)
365 if (useGpu() && !emulateGpu())
367 Nbnxm::setupGpuShortRangeWork(gpu_nbv, gpuBonded, iLocality);
371 //! Returns true if there is GPU short-range work for the given atom locality.
372 bool haveGpuShortRangeWork(const Nbnxm::AtomLocality aLocality)
374 return ((useGpu() && !emulateGpu()) &&
375 Nbnxm::haveGpuShortRangeWork(gpu_nbv, aLocality));
378 // TODO: Make all data members private
380 //! All data related to the pair lists
381 std::unique_ptr<PairlistSets> pairlistSets_;
382 //! Working data for constructing the pairlists
383 std::unique_ptr<PairSearch> pairSearch_;
385 std::unique_ptr<nbnxn_atomdata_t> nbat;
387 //! The non-bonded setup, also affects the pairlist construction kernel
388 Nbnxm::KernelSetup kernelSetup_;
389 //! \brief Pointer to wallcycle structure.
390 gmx_wallcycle *wcycle_;
392 //! GPU Nbnxm data, only used with a physical GPU (TODO: use unique_ptr)
393 gmx_nbnxn_gpu_t *gpu_nbv;
399 /*! \brief Creates an Nbnxm object */
400 std::unique_ptr<nonbonded_verlet_t>
401 init_nb_verlet(const gmx::MDLogger &mdlog,
402 gmx_bool bFEP_NonBonded,
403 const t_inputrec *ir,
404 const t_forcerec *fr,
406 const gmx_hw_info_t &hardwareInfo,
407 const gmx_device_info_t *deviceInfo,
408 const gmx_mtop_t *mtop,
410 gmx_wallcycle *wcycle);
414 /*! \brief Put the atoms on the pair search grid.
416 * Only atoms atomStart to atomEnd in x are put on the grid.
417 * The atom_density is used to determine the grid size.
418 * When atomDensity<=0, the density is determined from atomEnd-atomStart and the corners.
419 * With domain decomposition part of the n particles might have migrated,
420 * but have not been removed yet. This count is given by nmoved.
421 * When move[i] < 0 particle i has migrated and will not be put on the grid.
422 * Without domain decomposition move will be NULL.
424 void nbnxn_put_on_grid(nonbonded_verlet_t *nb_verlet,
427 const rvec lowerCorner,
428 const rvec upperCorner,
429 const gmx::UpdateGroupsCog *updateGroupsCog,
433 gmx::ArrayRef<const int> atomInfo,
434 gmx::ArrayRef<const gmx::RVec> x,
438 /*! \brief As nbnxn_put_on_grid, but for the non-local atoms
440 * with domain decomposition. Should be called after calling
441 * nbnxn_search_put_on_grid for the local atoms / home zone.
443 void nbnxn_put_on_grid_nonlocal(nonbonded_verlet_t *nb_verlet,
444 const struct gmx_domdec_zones_t *zones,
445 gmx::ArrayRef<const int> atomInfo,
446 gmx::ArrayRef<const gmx::RVec> x);
448 #endif // GMX_NBNXN_NBNXM_H