2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 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 /*! \libinternal \file
37 * \brief Defines the host-side PME GPU data structures.
38 * \todo Some renaming/refactoring, which does not impair the performance:
39 * -- bringing the function names up to guidelines
40 * -- PmeGpuSettings -> PmeGpuTasks
41 * -- refining GPU notation application (#2053)
42 * -- renaming coefficients to charges (?)
44 * \author Aleksei Iupinov <a.yupinov@gmail.com>
45 * \ingroup module_ewald
48 #ifndef GMX_EWALD_PME_GPU_TYPES_HOST_H
49 #define GMX_EWALD_PME_GPU_TYPES_HOST_H
56 #include "gromacs/ewald/pme.h"
57 #include "gromacs/ewald/pme_gpu_program.h"
58 #include "gromacs/gpu_utils/clfftinitializer.h"
59 #include "gromacs/gpu_utils/gpu_utils.h" // for GpuApiCallBehavior
60 #include "gromacs/gpu_utils/hostallocator.h"
61 #include "gromacs/math/vectypes.h"
63 #if GMX_GPU != GMX_GPU_NONE
64 struct PmeGpuSpecific;
66 /*! \brief A dummy typedef for the GPU host data placeholder on non-GPU builds */
67 typedef int PmeGpuSpecific;
70 #if GMX_GPU == GMX_GPU_CUDA
71 struct PmeGpuCudaKernelParams;
72 /*! \brief A typedef for including the GPU kernel arguments data by pointer */
73 typedef PmeGpuCudaKernelParams PmeGpuKernelParams;
74 #elif GMX_GPU == GMX_GPU_OPENCL
75 struct PmeGpuKernelParamsBase;
76 /*! \brief A typedef for including the GPU kernel arguments data by pointer */
77 typedef PmeGpuKernelParamsBase PmeGpuKernelParams;
79 /*! \brief A dummy typedef for the GPU kernel arguments data placeholder on non-GPU builds */
80 typedef int PmeGpuKernelParams;
83 struct gmx_device_info_t;
86 * The PME GPU settings structure, included in the main PME GPU structure by value.
90 /* Permanent settings set on initialization */
91 /*! \brief A boolean which tells if the solving is performed on GPU. Currently always true */
93 /*! \brief A boolean which tells if the gathering is performed on GPU. Currently always true */
94 bool performGPUGather;
95 /*! \brief A boolean which tells if the FFT is performed on GPU. Currently true for a single MPI rank. */
97 /*! \brief A convenience boolean which tells if PME decomposition is used. */
98 bool useDecomposition;
99 /*! \brief True if PME forces are reduced on-GPU, false if reduction is done on the CPU;
100 * in the former case transfer does not need to happen.
102 * Note that this flag may change per-step.
104 bool useGpuForceReduction;
106 /*! \brief A boolean which tells if any PME GPU stage should copy all of its outputs to the
107 * host. Only intended to be used by the test framework.
110 /*! \brief An enum which tells whether most PME GPU D2H/H2D data transfers should be synchronous. */
111 GpuApiCallBehavior transferKind;
112 /*! \brief Various flags for the current PME computation, corresponding to the GMX_PME_ flags in pme.h. */
115 * Currently only supported by CUDA.
116 * Controls if we should use order (i.e. 4) threads per atom for the GPU
117 * or order*order (i.e. 16) threads per atom.
119 bool useOrderThreadsPerAtom;
121 * Currently only supported by CUDA.
122 * Controls if we should recalculate the splines in the gather or
123 * save the values in the spread and reload in the gather.
125 bool recalculateSplines;
128 // TODO There's little value in computing the Coulomb and LJ virial
129 // separately, so we should simplify that.
130 // TODO The matrices might be best as a view, but not currently
131 // possible. Use mdspan?
134 gmx::ArrayRef<gmx::RVec> forces_; //!< Host staging area for PME forces
135 bool haveForceOutput_ =
136 false; //!< True if forces have been staged other false (when forces are reduced on the GPU).
137 real coulombEnergy_ = 0; //!< Host staging area for PME coulomb energy
138 matrix coulombVirial_ = { { 0 } }; //!< Host staging area for PME coulomb virial contributions
139 real lennardJonesEnergy_ = 0; //!< Host staging area for PME LJ energy
140 matrix lennardJonesVirial_ = { { 0 } }; //!< Host staging area for PME LJ virial contributions
144 * The PME GPU intermediate buffers structure, included in the main PME GPU structure by value.
145 * Buffers are managed by the PME GPU module.
149 //! Host-side force buffer
150 gmx::PaddedHostVector<gmx::RVec> h_forces;
152 /*! \brief Virial and energy intermediate host-side buffer. Size is PME_GPU_VIRIAL_AND_ENERGY_COUNT. */
153 float* h_virialAndEnergy;
154 /*! \brief B-spline values intermediate host-side buffer. */
155 float* h_splineModuli;
157 /*! \brief Pointer to the host memory with B-spline values. Only used for host-side gather, or unit tests */
159 /*! \brief Pointer to the host memory with B-spline derivative values. Only used for host-side gather, or unit tests */
161 /*! \brief Pointer to the host memory with ivec atom gridline indices. Only used for host-side gather, or unit tests */
162 int* h_gridlineIndices;
166 * The PME GPU structure for all the data copied directly from the CPU PME structure.
167 * The copying is done when the CPU PME structure is already (re-)initialized
168 * (pme_gpu_reinit is called at the end of gmx_pme_init).
169 * All the variables here are named almost the same way as in gmx_pme_t.
170 * The types are different: pointers are replaced by vectors.
171 * TODO: use the shared data with the PME CPU.
172 * Included in the main PME GPU structure by value.
176 /*! \brief Grid count - currently always 1 on GPU */
178 /*! \brief Grid dimensions - nkx, nky, nkz */
180 /*! \brief PME interpolation order */
182 /*! \brief Ewald splitting coefficient for Coulomb */
184 /*! \brief Electrostatics parameter */
186 /*! \brief Gridline indices - nnx, nny, nnz */
188 /*! \brief Fractional shifts - fshx, fshy, fshz */
189 std::vector<real> fsh;
190 /*! \brief Precomputed B-spline values */
191 std::vector<real> bsp_mod[DIM];
192 /*! \brief The PME codepath being taken */
194 /*! \brief Whether PME execution is happening on a PME-only rank (from gmx_pme_t.bPPnode). */
196 /*! \brief The box scaler based on inputrec - created in pme_init and managed by CPU structure */
197 class EwaldBoxZScaler* boxScaler;
198 /*! \brief The previous computation box to know if we even need to update the current box params.
199 * \todo Manage this on higher level.
200 * \todo Alternatively, when this structure is used by CPU PME code, make use of this field there as well.
206 * The main PME GPU host structure, included in the PME CPU structure by pointer.
210 /*! \brief The information copied once per reinit from the CPU structure. */
211 std::shared_ptr<PmeShared> common; // TODO: make the CPU structure use the same type
213 //! A handle to the program created by buildPmeGpuProgram()
214 PmeGpuProgramHandle programHandle_;
216 //! Handle that ensures the clFFT library has been initialized once per process.
217 std::unique_ptr<gmx::ClfftInitializer> initializedClfftLibrary_;
219 /*! \brief The settings. */
220 PmeGpuSettings settings;
222 /*! \brief The host-side buffers.
223 * The device-side buffers are buried in kernelParams, but that will have to change.
225 PmeGpuStaging staging;
227 /*! \brief Number of local atoms, padded to be divisible by c_pmeAtomDataAlignment.
228 * Used for kernel scheduling.
229 * kernelParams.atoms.nAtoms is the actual atom count to be used for data copying.
230 * TODO: this and the next member represent a memory allocation/padding properties -
231 * what a container type should do ideally.
234 /*! \brief Number of local atoms, padded to be divisible by c_pmeAtomDataAlignment
235 * if c_usePadding is true.
236 * Used only as a basic size for almost all the atom data allocations
237 * (spline parameter data is also aligned by PME_SPREADGATHER_PARTICLES_PER_WARP).
238 * This should be the same as (c_usePadding ? nAtomsPadded : kernelParams.atoms.nAtoms).
239 * kernelParams.atoms.nAtoms is the actual atom count to be used for most data copying.
243 /*! \brief A pointer to the device used during the execution. */
244 const gmx_device_info_t* deviceInfo;
246 /*! \brief Kernel scheduling grid width limit in X - derived from deviceinfo compute capability in CUDA.
247 * Declared as very large int to make it useful in computations with type promotion, to avoid overflows.
248 * OpenCL seems to not have readily available global work size limit, so we just assign a large arbitrary constant to this instead.
249 * TODO: this should be in PmeGpuProgram(Impl)
251 std::intmax_t maxGridWidthX;
253 /*! \brief A single structure encompassing all the PME data used on GPU.
254 * Its value is the only argument to all the PME GPU kernels.
255 * \todo Test whether this should be copied to the constant GPU memory once for each computation
256 * (or even less often with no box updates) instead of being an argument.
258 std::shared_ptr<PmeGpuKernelParams> kernelParams;
260 /*! \brief The pointer to GPU-framework specific host-side data, such as CUDA streams and events. */
261 std::shared_ptr<PmeGpuSpecific> archSpecific; /* FIXME: make it an unique_ptr */