2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2016,2017, 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 GPU-agnostic PME GPU data structures
38 * (the host-side PME GPU data, and the GPU function parameters).
39 * \todo Due to Gerrit workflow and time constraints, some renaming/refactoring
40 * which does not impair the performance will be performed once
41 * most of the initial PME CUDA implementation is merged
42 * into the master branch (likely, after release 2017).
43 * This should include:
44 * -- bringing the function names up to guidelines
45 * -- PmeGpuSettings -> PmeGpuTasks
46 * -- refining GPU notation application (#2053)
47 * -- renaming coefficients to charges (?)
49 * \author Aleksei Iupinov <a.yupinov@gmail.com>
50 * \ingroup module_ewald
53 #ifndef GMX_EWALD_PME_GPU_TYPES_H
54 #define GMX_EWALD_PME_GPU_TYPES_H
61 #include "gromacs/ewald/pme.h"
62 #include "gromacs/math/vectypes.h"
63 #include "gromacs/utility/basedefinitions.h"
66 struct gmx_device_info_t;
68 #if GMX_GPU == GMX_GPU_CUDA
71 /*! \brief A typedef for including the GPU host data by pointer */
72 typedef PmeGpuCuda PmeGpuSpecific;
74 struct PmeGpuCudaKernelParams;
75 /*! \brief A typedef for including the GPU kernel arguments data by pointer */
76 typedef PmeGpuCudaKernelParams PmeGpuKernelParams;
80 /*! \brief A dummy typedef for the GPU host data placeholder on non-GPU builds */
81 typedef int PmeGpuSpecific;
82 /*! \brief A dummy typedef for the GPU kernel arguments data placeholder on non-GPU builds */
83 typedef int PmeGpuKernelParams;
87 /* What follows is all the PME GPU function arguments,
88 * sorted into several device-side structures depending on the update rate.
89 * This is GPU agnostic (float3 replaced by float[3], etc.).
90 * The GPU-framework specifics (e.g. cudaTextureObject_t handles) are described
91 * in the larger structure PmeGpuCudaKernelParams in the pme.cuh.
95 * A GPU data structure for storing the constant PME data.
96 * This only has to be initialized once.
98 struct PmeGpuConstParams
100 /*! \brief Electrostatics coefficient = ONE_4PI_EPS0 / pme->epsilon_r */
102 /*! \brief Virial and energy GPU array. Size is PME_GPU_ENERGY_AND_VIRIAL_COUNT (7) floats.
103 * The element order is virxx, viryy, virzz, virxy, virxz, viryz, energy. */
104 float *d_virialAndEnergy;
108 * A GPU data structure for storing the PME data related to the grid sizes and cut-off.
109 * This only has to be updated at every DD step.
111 struct PmeGpuGridParams
114 /*! \brief Real-space grid data dimensions. */
115 int realGridSize[DIM];
116 /*! \brief Real-space grid dimensions, only converted to floating point. */
117 float realGridSizeFP[DIM];
118 /*! \brief Real-space grid dimensions (padded). The padding as compared to realGridSize includes the (order - 1) overlap. */
119 int realGridSizePadded[DIM]; /* Is major dimension of this ever used in kernels? */
120 /*! \brief Fourier grid dimensions. This counts the complex numbers! */
121 int complexGridSize[DIM];
122 /*! \brief Fourier grid dimensions (padded). This counts the complex numbers! */
123 int complexGridSizePadded[DIM];
126 /*! \brief Real space grid. */
128 /*! \brief Complex grid - used in FFT/solve. If inplace cuFFT is used, then it is the same pointer as realGrid. */
129 float *d_fourierGrid;
131 /*! \brief Ewald solving factor = (M_PI / pme->ewaldcoeff_q)^2 */
134 /*! \brief Grid spline values as in pme->bsp_mod
135 * (laid out sequentially (XXX....XYYY......YZZZ.....Z))
137 float *d_splineModuli;
138 /*! \brief Offsets for X/Y/Z components of d_splineModuli */
139 int splineValuesOffset[DIM];
141 /*! \brief Fractional shifts lookup table as in pme->fshx/fshy/fshz, laid out sequentially (XXX....XYYY......YZZZ.....Z) */
142 float *d_fractShiftsTable;
143 /*! \brief Gridline indices lookup table
144 * (modulo lookup table as in pme->nnx/nny/nnz, laid out sequentially (XXX....XYYY......YZZZ.....Z)) */
145 int *d_gridlineIndicesTable;
146 /*! \brief Offsets for X/Y/Z components of d_fractShiftsTable and d_gridlineIndicesTable */
147 int tablesOffsets[DIM];
151 * A GPU data structure for storing the PME data of the atoms, local to this process' domain partition.
152 * This only has to be updated every DD step.
154 struct PmeGpuAtomParams
156 /*! \brief Number of local atoms */
158 /*! \brief Pointer to the global GPU memory with input rvec atom coordinates.
159 * The coordinates themselves change and need to be copied to the GPU for every PME computation,
160 * but reallocation happens only at DD.
162 float *d_coordinates;
163 /*! \brief Pointer to the global GPU memory with input atom charges.
164 * The charges only need to be reallocated and copied to the GPU at DD step.
166 float *d_coefficients;
167 /*! \brief Pointer to the global GPU memory with input/output rvec atom forces.
168 * The forces change and need to be copied from (and possibly to) the GPU for every PME computation,
169 * but reallocation happens only at DD.
172 /*! \brief Pointer to the global GPU memory with ivec atom gridline indices.
173 * Computed on GPU in the spline calculation part.
175 int *d_gridlineIndices;
177 /* B-spline parameters are computed entirely on GPU for every PME computation, not copied.
178 * Unless we want to try something like GPU spread + CPU gather?
180 /*! \brief Pointer to the global GPU memory with B-spline values */
182 /*! \brief Pointer to the global GPU memory with B-spline derivative values */
187 * A GPU data structure for storing the PME data which might change for each new PME computation.
189 struct PmeGpuDynamicParams
191 /* The box parameters. The box only changes size with pressure coupling enabled. */
193 * Reciprocal (inverted unit cell) box.
195 * The box is transposed as compared to the CPU pme->recipbox.
196 * Basically, spread uses matrix columns (while solve and gather use rows).
197 * This storage format might be not the most optimal since the box is always triangular so there are zeroes.
199 float recipBox[DIM][DIM];
200 /*! \brief The unit cell volume for solving. */
205 * A single structure encompassing almost all the PME data used in GPU kernels on device.
206 * This is inherited by the GPU framework-specific structure
207 * (PmeGpuCudaKernelParams in pme.cuh).
208 * This way, most code preparing the kernel parameters can be GPU-agnostic by casting
209 * the kernel parameter data pointer to PmeGpuKernelParamsBase.
211 struct PmeGpuKernelParamsBase
213 /*! \brief Constant data that is set once. */
214 PmeGpuConstParams constants;
215 /*! \brief Data dependent on the grid size/cutoff. */
216 PmeGpuGridParams grid;
217 /*! \brief Data dependent on the DD and local atoms. */
218 PmeGpuAtomParams atoms;
219 /*! \brief Data that possibly changes for every new PME computation.
220 * This should be kept up-to-date by calling pme_gpu_prepare_computation(...)
221 * before launching spreading.
223 PmeGpuDynamicParams current;
226 /* Here are the host-side structures */
229 * The PME GPU settings structure, included in the main PME GPU structure by value.
231 struct PmeGpuSettings
233 /* Permanent settings set on initialization */
234 /*! \brief A boolean which tells if the solving is performed on GPU. Currently always true */
235 bool performGPUSolve;
236 /*! \brief A boolean which tells if the gathering is performed on GPU. Currently always true */
237 bool performGPUGather;
238 /*! \brief A boolean which tells if the FFT is performed on GPU. Currently true for a single MPI rank. */
240 /*! \brief A convenience boolean which tells if PME decomposition is used. */
241 bool useDecomposition;
242 /*! \brief A boolean which tells if any PME GPU stage should copy all of its outputs to the host.
243 * Only intended to be used by the test framework.
246 /*! \brief Various flags for the current PME computation, corresponding to the GMX_PME_ flags in pme.h. */
251 * The PME GPU intermediate buffers structure, included in the main PME GPU structure by value.
252 * Buffers are managed by the PME GPU module.
256 /*! \brief Virial and energy intermediate host-side buffer. Size is PME_GPU_VIRIAL_AND_ENERGY_COUNT. */
257 float *h_virialAndEnergy;
258 /*! \brief B-spline values intermediate host-side buffer. */
259 float *h_splineModuli;
261 /*! \brief Pointer to the host memory with B-spline values. Only used for host-side gather, or unit tests */
263 /*! \brief Pointer to the host memory with B-spline derivative values. Only used for host-side gather, or unit tests */
265 /*! \brief Pointer to the host memory with ivec atom gridline indices. Only used for host-side gather, or unit tests */
266 int *h_gridlineIndices;
270 * The PME GPU structure for all the data copied directly from the CPU PME structure.
271 * The copying is done when the CPU PME structure is already (re-)initialized
272 * (pme_gpu_reinit is called at the end of gmx_pme_init).
273 * All the variables here are named almost the same way as in gmx_pme_t.
274 * The types are different: pointers are replaced by vectors.
275 * TODO: use the shared data with the PME CPU.
276 * Included in the main PME GPU structure by value.
280 /*! \brief Grid count - currently always 1 on GPU */
282 /*! \brief Grid dimensions - nkx, nky, nkz */
284 /*! \brief Padded grid dimensions - pmegrid_nx, pmegrid_ny, pmegrid_nz
285 * TODO: find out if these are really needed for the CPU FFT compatibility.
288 /*! \brief PME interpolation order */
290 /*! \brief Ewald splitting coefficient for Coulomb */
292 /*! \brief Electrostatics parameter */
294 /*! \brief Gridline indices - nnx, nny, nnz */
296 /*! \brief Fractional shifts - fshx, fshy, fshz */
297 std::vector<real> fsh;
298 /*! \brief Precomputed B-spline values */
299 std::vector<real> bsp_mod[DIM];
300 /*! \brief The PME codepath being taken */
302 /*! \brief The box scaler based on inputrec - created in pme_init and managed by CPU structure */
303 class EwaldBoxZScaler *boxScaler;
304 /*! \brief The previous computation box to know if we even need to update the current box params.
305 * \todo Manage this on higher level.
306 * \todo Alternatively, when this structure is used by CPU PME code, make use of this field there as well.
312 * The main PME GPU host structure, included in the PME CPU structure by pointer.
316 /*! \brief The information copied once per reinit from the CPU structure. */
317 std::shared_ptr<PmeShared> common; // TODO: make the CPU structure use the same type
319 /*! \brief The settings. */
320 PmeGpuSettings settings;
322 /*! \brief The host-side buffers.
323 * The device-side buffers are buried in kernelParams, but that will have to change.
325 PmeGpuStaging staging;
327 /*! \brief Number of local atoms, padded to be divisible by PME_ATOM_DATA_ALIGNMENT.
328 * Used for kernel scheduling.
329 * kernelParams.atoms.nAtoms is the actual atom count to be used for data copying.
330 * TODO: this and the next member represent a memory allocation/padding properties -
331 * what a container type should do ideally.
334 /*! \brief Number of local atoms, padded to be divisible by PME_ATOM_DATA_ALIGNMENT
335 * if c_usePadding is true.
336 * Used only as a basic size for almost all the atom data allocations
337 * (spline parameter data is also aligned by PME_SPREADGATHER_PARTICLES_PER_WARP).
338 * This should be the same as (c_usePadding ? nAtomsPadded : kernelParams.atoms.nAtoms).
339 * kernelParams.atoms.nAtoms is the actual atom count to be used for most data copying.
343 /*! \brief A pointer to the device used during the execution. */
344 gmx_device_info_t *deviceInfo;
346 /*! \brief A single structure encompassing all the PME data used on GPU.
347 * Its value is the only argument to all the PME GPU kernels.
348 * \todo Test whether this should be copied to the constant GPU memory once for each computation
349 * (or even less often with no box updates) instead of being an argument.
351 std::shared_ptr<PmeGpuKernelParams> kernelParams;
353 /*! \brief The pointer to GPU-framework specific host-side data, such as CUDA streams and events. */
354 std::shared_ptr<PmeGpuSpecific> archSpecific; /* FIXME: make it an unique_ptr */