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 /* kernel does #include "gromacs/math/utilities.h" */
64 /* Move the actual useful stuff here: */
67 #define M_FLOAT_1_SQRTPI 0.564189583547756f
70 /*! \brief Constants for platform-dependent defaults for the prune kernel's j4 processing concurrency.
72 * Initialized using macros that can be overridden at compile-time (using #GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY).
75 const int c_oclPruneKernelJ4ConcurrencyDEFAULT
= GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY_DEFAULT
;
78 /*! \brief Returns the j4 processing concurrency parameter for the vendor \p vendorId
79 * \param vendorId takes values from #ocl_vendor_id_t.
81 static inline int getOclPruneKernelJ4Concurrency(int vendorId
)
85 default: return c_oclPruneKernelJ4ConcurrencyDEFAULT
;
90 /*! \brief Electrostatic OpenCL kernel flavors.
92 * Types of electrostatics implementations available in the OpenCL non-bonded
93 * force kernels. These represent both the electrostatics types implemented
94 * by the kernels (cut-off, RF, and Ewald - a subset of what's defined in
95 * enums.h) as well as encode implementation details analytical/tabulated
96 * and single or twin cut-off (for Ewald kernels).
97 * Note that the cut-off and RF kernels have only analytical flavor and unlike
98 * in the CPU kernels, the tabulated kernels are ATM Ewald-only.
100 * The row-order of pointers to different electrostatic kernels defined in
101 * nbnxn_cuda.cu by the nb_*_kfunc_ptr function pointer table
102 * should match the order of enumerated types below.
109 eelOclEWALD_TAB_TWIN
,
111 eelOclEWALD_ANA_TWIN
,
115 /*! \brief VdW OpenCL kernel flavors.
117 * The enumerates values correspond to the LJ implementations in the OpenCL non-bonded
120 * The column-order of pointers to different electrostatic kernels defined in
121 * nbnxn_cuda.cu by the nb_*_kfunc_ptr function pointer table
122 * should match the order of enumerated types below.
136 /*! \brief Pruning kernel flavors.
138 * The values correspond to the first call of the pruning post-list generation
139 * and the rolling pruning, respectively.
149 * \brief Staging area for temporary data downloaded from the GPU.
151 * The energies/shift forces get downloaded here first, before getting added
152 * to the CPU-side aggregate values.
154 typedef struct cl_nb_staging
156 float* e_lj
; /**< LJ energy */
157 float* e_el
; /**< electrostatic energy */
158 float (*fshift
)[3]; /**< float3 buffer with shift forces */
162 * \brief Nonbonded atom data - both inputs and outputs.
164 typedef struct cl_atomdata
166 int natoms
; /**< number of atoms */
167 int natoms_local
; /**< number of local atoms */
168 int nalloc
; /**< allocation size for the atom data (xq, f) */
170 cl_mem xq
; /**< float4 buffer with atom coordinates + charges, size natoms */
172 cl_mem f
; /**< float3 buffer with force output array, size natoms */
173 size_t f_elem_size
; /**< Size in bytes for one element of f buffer */
175 cl_mem e_lj
; /**< LJ energy output, size 1 */
176 cl_mem e_el
; /**< Electrostatics energy input, size 1 */
178 cl_mem fshift
; /**< float3 buffer with shift forces */
179 size_t fshift_elem_size
; /**< Size in bytes for one element of fshift buffer */
181 int ntypes
; /**< number of atom types */
182 cl_mem atom_types
; /**< int buffer with atom type indices, size natoms */
183 cl_mem lj_comb
; /**< float2 buffer with sqrt(c6),sqrt(c12), size natoms */
185 cl_mem shift_vec
; /**< float3 buffer with shifts values */
186 size_t shift_vec_elem_size
; /**< Size in bytes for one element of shift_vec buffer */
188 cl_bool bShiftVecUploaded
; /**< true if the shift vector has been uploaded */
192 * \brief Parameters required for the OpenCL nonbonded calculations.
194 typedef struct cl_nbparam
197 int eeltype
; /**< type of electrostatics, takes values from #eelOcl */
198 int vdwtype
; /**< type of VdW impl., takes values from #evdwOcl */
200 float epsfac
; /**< charge multiplication factor */
201 float c_rf
; /**< Reaction-field/plain cutoff electrostatics const. */
202 float two_k_rf
; /**< Reaction-field electrostatics constant */
203 float ewald_beta
; /**< Ewald/PME parameter */
204 float sh_ewald
; /**< Ewald/PME correction term substracted from the direct-space potential */
205 float sh_lj_ewald
; /**< LJ-Ewald/PME correction term added to the correction potential */
206 float ewaldcoeff_lj
; /**< LJ-Ewald/PME coefficient */
208 float rcoulomb_sq
; /**< Coulomb cut-off squared */
210 float rvdw_sq
; /**< VdW cut-off squared */
211 float rvdw_switch
; /**< VdW switched cut-off */
212 float rlistOuter_sq
; /**< Full, outer pair-list cut-off squared */
213 float rlistInner_sq
; /**< Inner, dynamic pruned pair-list cut-off squared */
214 bool useDynamicPruning
; /**< True if we use dynamic pair-list pruning */
216 shift_consts_t dispersion_shift
; /**< VdW shift dispersion constants */
217 shift_consts_t repulsion_shift
; /**< VdW shift repulsion constants */
218 switch_consts_t vdw_switch
; /**< VdW switch constants */
220 /* LJ non-bonded parameters - accessed through texture memory */
221 cl_mem nbfp_climg2d
; /**< nonbonded parameter table with C6/C12 pairs per atom type-pair, 2*ntype^2 elements */
222 cl_mem nbfp_comb_climg2d
; /**< nonbonded parameter table per atom type, 2*ntype elements */
224 /* Ewald Coulomb force table data - accessed through texture memory */
225 float coulomb_tab_scale
; /**< table scale/spacing */
226 cl_mem coulomb_tab_climg2d
; /**< pointer to the table in the device memory */
230 * \brief Data structure shared between the OpenCL device code and OpenCL host code
232 * Must not contain OpenCL objects (buffers)
233 * TODO: review, improve */
234 typedef struct cl_nbparam_params
237 int eeltype
; /**< type of electrostatics, takes values from #eelCu */
238 int vdwtype
; /**< type of VdW impl., takes values from #evdwCu */
240 float epsfac
; /**< charge multiplication factor */
241 float c_rf
; /**< Reaction-field/plain cutoff electrostatics const. */
242 float two_k_rf
; /**< Reaction-field electrostatics constant */
243 float ewald_beta
; /**< Ewald/PME parameter */
244 float sh_ewald
; /**< Ewald/PME correction term substracted from the direct-space potential */
245 float sh_lj_ewald
; /**< LJ-Ewald/PME correction term added to the correction potential */
246 float ewaldcoeff_lj
; /**< LJ-Ewald/PME coefficient */
248 float rcoulomb_sq
; /**< Coulomb cut-off squared */
250 float rvdw_sq
; /**< VdW cut-off squared */
251 float rvdw_switch
; /**< VdW switched cut-off */
252 float rlistOuter_sq
; /**< Full, outer pair-list cut-off squared */
253 float rlistInner_sq
; /**< Inner, dynamic pruned pair-list cut-off squared */
255 shift_consts_t dispersion_shift
; /**< VdW shift dispersion constants */
256 shift_consts_t repulsion_shift
; /**< VdW shift repulsion constants */
257 switch_consts_t vdw_switch
; /**< VdW switch constants */
259 /* Ewald Coulomb force table data - accessed through texture memory */
260 float coulomb_tab_scale
; /**< table scale/spacing */
261 } cl_nbparam_params_t
;
265 * \brief Pair list data.
267 using cl_plist_t
= Nbnxm::gpu_plist
;
270 * \brief Typedef of actual timer type.
272 typedef struct Nbnxm::gpu_timers_t cl_timers_t
;
275 * \brief Main data structure for OpenCL nonbonded force calculations.
277 struct gmx_nbnxn_ocl_t
279 const gmx_device_info_t
* dev_info
; /**< OpenCL device information */
280 struct gmx_device_runtime_data_t
* dev_rundata
; /**< OpenCL runtime data (context, kernels) */
282 /**< Pointers to non-bonded kernel functions
283 * organized similar with nb_kfunc_xxx arrays in nbnxn_ocl.cpp */
285 cl_kernel kernel_noener_noprune_ptr
[eelOclNR
][evdwOclNR
];
286 cl_kernel kernel_ener_noprune_ptr
[eelOclNR
][evdwOclNR
];
287 cl_kernel kernel_noener_prune_ptr
[eelOclNR
][evdwOclNR
];
288 cl_kernel kernel_ener_prune_ptr
[eelOclNR
][evdwOclNR
];
290 cl_kernel kernel_pruneonly
[ePruneNR
]; /**< prune kernels, ePruneKind defined the kernel kinds */
292 bool bPrefetchLjParam
; /**< true if prefetching fg i-atom LJ parameters should be used in the kernels */
294 /**< auxiliary kernels implementing memset-like functions */
296 cl_kernel kernel_memset_f
;
297 cl_kernel kernel_memset_f2
;
298 cl_kernel kernel_memset_f3
;
299 cl_kernel kernel_zero_e_fshift
;
302 cl_bool bUseTwoStreams
; /**< true if doing both local/non-local NB work on GPU */
303 cl_bool bNonLocalStreamActive
; /**< true indicates that the nonlocal_done event was enqueued */
305 cl_atomdata_t
* atdat
; /**< atom data */
306 cl_nbparam_t
* nbparam
; /**< parameters required for the non-bonded calc. */
307 gmx::EnumerationArray
<Nbnxm::InteractionLocality
, cl_plist_t
*> plist
; /**< pair-list data structures (local and non-local) */
308 cl_nb_staging_t nbst
; /**< staging area where fshift/energies get downloaded */
310 gmx::EnumerationArray
<Nbnxm::InteractionLocality
, cl_command_queue
> stream
; /**< local and non-local GPU queues */
312 /** events used for synchronization */
313 cl_event nonlocal_done
; /**< event triggered when the non-local non-bonded kernel
314 is done (and the local transfer can proceed) */
315 cl_event misc_ops_and_local_H2D_done
; /**< event triggered when the tasks issued in
316 the local stream that need to precede the
317 non-local force calculations are done
318 (e.g. f buffer 0-ing, local x/q H2D) */
320 //! True if there has been local/nonlocal GPU work, either bonded or nonbonded, scheduled
321 // to be executed in the current domain. As long as bonded work is not split up into
322 // local/nonlocal, if there is bonded GPU work, both flags will be true.
323 gmx::EnumerationArray
<Nbnxm::InteractionLocality
, bool> haveWork
;
326 cl_bool bDoTime
; /**< True if event-based timing is enabled. */
327 cl_timers_t
* timers
; /**< OpenCL event-based timers. */
328 struct gmx_wallclock_gpu_nbnxn_t
* timings
; /**< Timing data. TODO: deprecate this and query timers for accumulated data instead */
331 #endif /* NBNXN_OPENCL_TYPES_H */