2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016 by the GROMACS development team.
5 * Copyright (c) 2017,2018,2019,2020, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
39 * Data types used internally in the nbnxm_ocl module.
41 * \author Anca Hamuraru <anca@streamcomputing.eu>
42 * \author Szilárd Páll <pszilard@kth.se>
43 * \ingroup module_nbnxm
46 #ifndef GMX_NBNXM_NBNXM_OPENCL_TYPES_H
47 #define GMX_NBNXM_NBNXM_OPENCL_TYPES_H
49 #include "gromacs/gpu_utils/devicebuffer.h"
50 #include "gromacs/gpu_utils/gmxopencl.h"
51 #include "gromacs/gpu_utils/gputraits_ocl.h"
52 #include "gromacs/gpu_utils/oclutils.h"
53 #include "gromacs/mdtypes/interaction_const.h"
54 #include "gromacs/nbnxm/gpu_types_common.h"
55 #include "gromacs/nbnxm/nbnxm.h"
56 #include "gromacs/nbnxm/pairlist.h"
57 #include "gromacs/utility/enumerationhelpers.h"
58 #include "gromacs/utility/fatalerror.h"
59 #include "gromacs/utility/real.h"
61 #include "nbnxm_ocl_consts.h"
63 struct gmx_wallclock_gpu_nbnxn_t;
65 /* kernel does #include "gromacs/math/utilities.h" */
66 /* Move the actual useful stuff here: */
69 #define M_FLOAT_1_SQRTPI 0.564189583547756f
72 /*! \brief Constants for platform-dependent defaults for the prune kernel's j4 processing concurrency.
74 * Initialized using macros that can be overridden at compile-time (using #GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY).
77 const int c_oclPruneKernelJ4ConcurrencyDEFAULT = GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY_DEFAULT;
80 /*! \brief Returns the j4 processing concurrency parameter for the vendor \p vendorId
81 * \param vendorId takes values from #ocl_vendor_id_t.
83 static inline int getOclPruneKernelJ4Concurrency(int vendorId)
87 default: return c_oclPruneKernelJ4ConcurrencyDEFAULT;
92 /*! \brief Electrostatic OpenCL kernel flavors.
94 * Types of electrostatics implementations available in the OpenCL non-bonded
95 * force kernels. These represent both the electrostatics types implemented
96 * by the kernels (cut-off, RF, and Ewald - a subset of what's defined in
97 * enums.h) as well as encode implementation details analytical/tabulated
98 * and single or twin cut-off (for Ewald kernels).
99 * Note that the cut-off and RF kernels have only analytical flavor and unlike
100 * in the CPU kernels, the tabulated kernels are ATM Ewald-only.
102 * The row-order of pointers to different electrostatic kernels defined in
103 * nbnxn_cuda.cu by the nb_*_kfunc_ptr function pointer table
104 * should match the order of enumerated types below.
111 eelOclEWALD_TAB_TWIN,
113 eelOclEWALD_ANA_TWIN,
117 /*! \brief VdW OpenCL kernel flavors.
119 * The enumerates values correspond to the LJ implementations in the OpenCL non-bonded
122 * The column-order of pointers to different electrostatic kernels defined in
123 * nbnxn_cuda.cu by the nb_*_kfunc_ptr function pointer table
124 * should match the order of enumerated types below.
138 /*! \brief Pruning kernel flavors.
140 * The values correspond to the first call of the pruning post-list generation
141 * and the rolling pruning, respectively.
151 * \brief Staging area for temporary data downloaded from the GPU.
153 * The energies/shift forces get downloaded here first, before getting added
154 * to the CPU-side aggregate values.
159 float* e_lj = nullptr;
160 //! electrostatic energy
161 float* e_el = nullptr;
162 //! float3 buffer with shift forces
163 float (*fshift)[3] = nullptr;
167 * \brief Nonbonded atom data - both inputs and outputs.
169 typedef struct cl_atomdata
173 //! number of local atoms
175 //! allocation size for the atom data (xq, f)
178 //! float4 buffer with atom coordinates + charges, size natoms
181 //! float3 buffer with force output array, size natoms
184 //! LJ energy output, size 1
186 //! Electrostatics energy input, size 1
189 //! float3 buffer with shift forces
192 //! number of atom types
194 //! int buffer with atom type indices, size natoms
196 //! float2 buffer with sqrt(c6),sqrt(c12), size natoms
199 //! float3 buffer with shifts values
202 //! true if the shift vector has been uploaded
203 bool bShiftVecUploaded;
207 * \brief Parameters required for the OpenCL nonbonded calculations.
209 typedef struct cl_nbparam
212 //! type of electrostatics, takes values from #eelOcl
214 //! type of VdW impl., takes values from #evdwOcl
217 //! charge multiplication factor
219 //! Reaction-field/plain cutoff electrostatics const.
221 //! Reaction-field electrostatics constant
223 //! Ewald/PME parameter
225 //! Ewald/PME correction term substracted from the direct-space potential
227 //! LJ-Ewald/PME correction term added to the correction potential
229 //! LJ-Ewald/PME coefficient
232 //! Coulomb cut-off squared
235 //! VdW cut-off squared
237 //! VdW switched cut-off
239 //! Full, outer pair-list cut-off squared
241 //! Inner, dynamic pruned pair-list cut-off squared
243 //! True if we use dynamic pair-list pruning
244 bool useDynamicPruning;
246 //! VdW shift dispersion constants
247 shift_consts_t dispersion_shift;
248 //! VdW shift repulsion constants
249 shift_consts_t repulsion_shift;
250 //! VdW switch constants
251 switch_consts_t vdw_switch;
253 /* LJ non-bonded parameters - accessed through texture memory */
254 //! nonbonded parameter table with C6/C12 pairs per atom type-pair, 2*ntype^2 elements
256 //! nonbonded parameter table per atom type, 2*ntype elements
257 cl_mem nbfp_comb_climg2d;
259 /* Ewald Coulomb force table data - accessed through texture memory */
260 //! table scale/spacing
261 float coulomb_tab_scale;
262 //! pointer to the table in the device memory
263 cl_mem coulomb_tab_climg2d;
267 * \brief Data structure shared between the OpenCL device code and OpenCL host code
269 * Must not contain OpenCL objects (buffers)
270 * TODO: review, improve */
271 typedef struct cl_nbparam_params
274 //! type of electrostatics, takes values from #eelCu
276 //! type of VdW impl., takes values from #evdwCu
279 //! charge multiplication factor
281 //! Reaction-field/plain cutoff electrostatics const.
283 //! Reaction-field electrostatics constant
285 //! Ewald/PME parameter
287 //! Ewald/PME correction term substracted from the direct-space potential
289 //! LJ-Ewald/PME correction term added to the correction potential
291 //! LJ-Ewald/PME coefficient
294 //! Coulomb cut-off squared
297 //! VdW cut-off squared
299 //! VdW switched cut-off
301 //! Full, outer pair-list cut-off squared
303 //! Inner, dynamic pruned pair-list cut-off squared
306 //! VdW shift dispersion constants
307 shift_consts_t dispersion_shift;
308 //! VdW shift repulsion constants
309 shift_consts_t repulsion_shift;
310 //! VdW switch constants
311 switch_consts_t vdw_switch;
313 /* Ewald Coulomb force table data - accessed through texture memory */
314 //! table scale/spacing
315 float coulomb_tab_scale;
316 } cl_nbparam_params_t;
320 * \brief Pair list data.
322 using cl_plist_t = Nbnxm::gpu_plist;
325 * \brief Typedef of actual timer type.
327 typedef struct Nbnxm::gpu_timers_t cl_timers_t;
330 * \brief Main data structure for OpenCL nonbonded force calculations.
334 //! OpenCL device information
335 const gmx_device_info_t* dev_info = nullptr;
336 //! OpenCL runtime data (context, kernels)
337 struct gmx_device_runtime_data_t* dev_rundata = nullptr;
339 /**< Pointers to non-bonded kernel functions
340 * organized similar with nb_kfunc_xxx arrays in nbnxn_ocl.cpp */
342 cl_kernel kernel_noener_noprune_ptr[eelOclNR][evdwOclNR] = { { nullptr } };
343 cl_kernel kernel_ener_noprune_ptr[eelOclNR][evdwOclNR] = { { nullptr } };
344 cl_kernel kernel_noener_prune_ptr[eelOclNR][evdwOclNR] = { { nullptr } };
345 cl_kernel kernel_ener_prune_ptr[eelOclNR][evdwOclNR] = { { nullptr } };
347 //! prune kernels, ePruneKind defined the kernel kinds
348 cl_kernel kernel_pruneonly[ePruneNR] = { nullptr };
350 //! true if prefetching fg i-atom LJ parameters should be used in the kernels
351 bool bPrefetchLjParam = false;
353 /**< auxiliary kernels implementing memset-like functions */
355 cl_kernel kernel_memset_f = nullptr;
356 cl_kernel kernel_memset_f2 = nullptr;
357 cl_kernel kernel_memset_f3 = nullptr;
358 cl_kernel kernel_zero_e_fshift = nullptr;
361 //! true if doing both local/non-local NB work on GPU
362 bool bUseTwoStreams = false;
363 //! true indicates that the nonlocal_done event was enqueued
364 bool bNonLocalStreamActive = false;
367 cl_atomdata_t* atdat = nullptr;
368 //! parameters required for the non-bonded calc.
369 cl_nbparam_t* nbparam = nullptr;
370 //! pair-list data structures (local and non-local)
371 gmx::EnumerationArray<Nbnxm::InteractionLocality, cl_plist_t*> plist = { nullptr };
372 //! staging area where fshift/energies get downloaded
375 //! local and non-local GPU queues
376 gmx::EnumerationArray<Nbnxm::InteractionLocality, cl_command_queue> stream;
378 /*! \brief Events used for synchronization */
380 /*! \brief Event triggered when the non-local non-bonded
381 * kernel is done (and the local transfer can proceed) */
382 cl_event nonlocal_done = nullptr;
383 /*! \brief Event triggered when the tasks issued in the local
384 * stream that need to precede the non-local force or buffer
385 * operation calculations are done (e.g. f buffer 0-ing, local
386 * x/q H2D, buffer op initialization in local stream that is
387 * required also by nonlocal stream ) */
388 cl_event misc_ops_and_local_H2D_done = nullptr;
391 //! True if there has been local/nonlocal GPU work, either bonded or nonbonded, scheduled
392 // to be executed in the current domain. As long as bonded work is not split up into
393 // local/nonlocal, if there is bonded GPU work, both flags will be true.
394 gmx::EnumerationArray<Nbnxm::InteractionLocality, bool> haveWork;
397 //! True if event-based timing is enabled.
398 bool bDoTime = false;
399 //! OpenCL event-based timers.
400 cl_timers_t* timers = nullptr;
401 //! Timing data. TODO: deprecate this and query timers for accumulated data instead
402 gmx_wallclock_gpu_nbnxn_t* timings = nullptr;
405 #endif /* NBNXN_OPENCL_TYPES_H */