2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018, 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 * \brief Define CUDA implementation of nbnxn_gpu.h
38 * \author Szilard Pall <pall.szilard@gmail.com>
47 #include "gromacs/mdlib/nbnxn_gpu.h"
54 #include "nbnxn_cuda.h"
56 #include "gromacs/gpu_utils/cudautils.cuh"
57 #include "gromacs/mdlib/force_flags.h"
58 #include "gromacs/mdlib/nb_verlet.h"
59 #include "gromacs/mdlib/nbnxn_gpu_common.h"
60 #include "gromacs/mdlib/nbnxn_gpu_common_utils.h"
61 #include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
62 #include "gromacs/mdlib/nbnxn_pairlist.h"
63 #include "gromacs/timing/gpu_timing.h"
64 #include "gromacs/utility/cstringutil.h"
65 #include "gromacs/utility/gmxassert.h"
67 #include "nbnxn_cuda_types.h"
70 /***** The kernel declarations/definitions come here *****/
72 /* Top-level kernel declaration generation: will generate through multiple
73 * inclusion the following flavors for all kernel declarations:
74 * - force-only output;
75 * - force and energy output;
76 * - force-only with pair list pruning;
77 * - force and energy output with pair list pruning.
79 #define FUNCTION_DECLARATION_ONLY
81 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernels.cuh"
82 /** Force & energy **/
84 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernels.cuh"
87 /*** Pair-list pruning kernels ***/
90 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernels.cuh"
91 /** Force & energy **/
93 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernels.cuh"
97 /* Prune-only kernels */
98 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_pruneonly.cuh"
99 #undef FUNCTION_DECLARATION_ONLY
101 /* Now generate the function definitions if we are using a single compilation unit. */
102 #if GMX_CUDA_NB_SINGLE_COMPILATION_UNIT
103 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_F_noprune.cu"
104 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_F_prune.cu"
105 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_VF_noprune.cu"
106 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_VF_prune.cu"
107 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_pruneonly.cu"
108 #endif /* GMX_CUDA_NB_SINGLE_COMPILATION_UNIT */
111 /*! Nonbonded kernel function pointer type */
112 typedef void (*nbnxn_cu_kfunc_ptr_t)(const cu_atomdata_t,
117 /*********************************/
119 /*! Returns the number of blocks to be used for the nonbonded GPU kernel. */
120 static inline int calc_nb_kernel_nblock(int nwork_units, const gmx_device_info_t *dinfo)
125 /* CUDA does not accept grid dimension of 0 (which can happen e.g. with an
126 empty domain) and that case should be handled before this point. */
127 assert(nwork_units > 0);
129 max_grid_x_size = dinfo->prop.maxGridSize[0];
131 /* do we exceed the grid x dimension limit? */
132 if (nwork_units > max_grid_x_size)
134 gmx_fatal(FARGS, "Watch out, the input system is too large to simulate!\n"
135 "The number of nonbonded work units (=number of super-clusters) exceeds the"
136 "maximum grid size in x dimension (%d > %d)!", nwork_units, max_grid_x_size);
143 /* Constant arrays listing all kernel function pointers and enabling selection
144 of a kernel in an elegant manner. */
146 /*! Pointers to the non-bonded kernels organized in 2-dim arrays by:
147 * electrostatics and VDW type.
149 * Note that the row- and column-order of function pointers has to match the
150 * order of corresponding enumerated electrostatics and vdw types, resp.,
151 * defined in nbnxn_cuda_types.h.
154 /*! Force-only kernel function pointers. */
155 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_noener_noprune_ptr[eelCuNR][evdwCuNR] =
157 { nbnxn_kernel_ElecCut_VdwLJ_F_cuda, nbnxn_kernel_ElecCut_VdwLJCombGeom_F_cuda, nbnxn_kernel_ElecCut_VdwLJCombLB_F_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_F_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_F_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_cuda },
158 { nbnxn_kernel_ElecRF_VdwLJ_F_cuda, nbnxn_kernel_ElecRF_VdwLJCombGeom_F_cuda, nbnxn_kernel_ElecRF_VdwLJCombLB_F_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_F_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_F_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_cuda },
159 { nbnxn_kernel_ElecEwQSTab_VdwLJ_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_cuda },
160 { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_cuda },
161 { nbnxn_kernel_ElecEw_VdwLJ_F_cuda, nbnxn_kernel_ElecEw_VdwLJCombGeom_F_cuda, nbnxn_kernel_ElecEw_VdwLJCombLB_F_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_F_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_cuda },
162 { nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_cuda }
165 /*! Force + energy kernel function pointers. */
166 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_ener_noprune_ptr[eelCuNR][evdwCuNR] =
168 { nbnxn_kernel_ElecCut_VdwLJ_VF_cuda, nbnxn_kernel_ElecCut_VdwLJCombGeom_VF_cuda, nbnxn_kernel_ElecCut_VdwLJCombLB_VF_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_cuda },
169 { nbnxn_kernel_ElecRF_VdwLJ_VF_cuda, nbnxn_kernel_ElecRF_VdwLJCombGeom_VF_cuda, nbnxn_kernel_ElecRF_VdwLJCombLB_VF_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_cuda },
170 { nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_cuda },
171 { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_cuda },
172 { nbnxn_kernel_ElecEw_VdwLJ_VF_cuda, nbnxn_kernel_ElecEw_VdwLJCombGeom_VF_cuda, nbnxn_kernel_ElecEw_VdwLJCombLB_VF_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_cuda },
173 { nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_cuda }
176 /*! Force + pruning kernel function pointers. */
177 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_noener_prune_ptr[eelCuNR][evdwCuNR] =
179 { nbnxn_kernel_ElecCut_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJCombGeom_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJCombLB_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_prune_cuda },
180 { nbnxn_kernel_ElecRF_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJCombGeom_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJCombLB_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_prune_cuda },
181 { nbnxn_kernel_ElecEwQSTab_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_prune_cuda },
182 { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_prune_cuda },
183 { nbnxn_kernel_ElecEw_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJCombGeom_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJCombLB_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_prune_cuda },
184 { nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_prune_cuda }
187 /*! Force + energy + pruning kernel function pointers. */
188 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_ener_prune_ptr[eelCuNR][evdwCuNR] =
190 { nbnxn_kernel_ElecCut_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJCombGeom_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJCombLB_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_prune_cuda },
191 { nbnxn_kernel_ElecRF_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJCombGeom_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJCombLB_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_prune_cuda },
192 { nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_prune_cuda },
193 { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_prune_cuda },
194 { nbnxn_kernel_ElecEw_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJCombLB_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_prune_cuda },
195 { nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_prune_cuda }
198 /*! Return a pointer to the kernel version to be executed at the current step. */
199 static inline nbnxn_cu_kfunc_ptr_t select_nbnxn_kernel(int eeltype,
203 const gmx_device_info_t gmx_unused *devInfo)
205 nbnxn_cu_kfunc_ptr_t res;
207 GMX_ASSERT(eeltype < eelCuNR,
208 "The electrostatics type requested is not implemented in the CUDA kernels.");
209 GMX_ASSERT(evdwtype < evdwCuNR,
210 "The VdW type requested is not implemented in the CUDA kernels.");
212 /* assert assumptions made by the kernels */
213 GMX_ASSERT(c_nbnxnGpuClusterSize*c_nbnxnGpuClusterSize/c_nbnxnGpuClusterpairSplit == devInfo->prop.warpSize,
214 "The CUDA kernels require the cluster_size_i*cluster_size_j/nbnxn_gpu_clusterpair_split to match the warp size of the architecture targeted.");
220 res = nb_kfunc_ener_prune_ptr[eeltype][evdwtype];
224 res = nb_kfunc_ener_noprune_ptr[eeltype][evdwtype];
231 res = nb_kfunc_noener_prune_ptr[eeltype][evdwtype];
235 res = nb_kfunc_noener_noprune_ptr[eeltype][evdwtype];
242 /*! \brief Calculates the amount of shared memory required by the nonbonded kernel in use. */
243 static inline int calc_shmem_required_nonbonded(const int num_threads_z, const gmx_device_info_t gmx_unused *dinfo, const cu_nbparam_t *nbp)
249 /* size of shmem (force-buffers/xq/atom type preloading) */
250 /* NOTE: with the default kernel on sm3.0 we need shmem only for pre-loading */
251 /* i-atom x+q in shared memory */
252 shmem = c_numClPerSupercl * c_clSize * sizeof(float4);
253 /* cj in shared memory, for each warp separately */
254 shmem += num_threads_z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(int);
255 if (dinfo->prop.major >= 3)
257 if (nbp->vdwtype == evdwCuCUTCOMBGEOM ||
258 nbp->vdwtype == evdwCuCUTCOMBLB)
260 /* i-atom LJ combination parameters in shared memory */
261 shmem += c_numClPerSupercl * c_clSize * sizeof(float2);
265 /* i-atom types in shared memory */
266 shmem += c_numClPerSupercl * c_clSize * sizeof(int);
269 if (dinfo->prop.major < 3)
271 /* force reduction buffers in shared memory */
272 shmem += c_clSize * c_clSize * 3 * sizeof(float);
277 /*! As we execute nonbonded workload in separate streams, before launching
278 the kernel we need to make sure that he following operations have completed:
279 - atomdata allocation and related H2D transfers (every nstlist step);
280 - pair list H2D transfer (every nstlist step);
281 - shift vector H2D transfer (every nstlist step);
282 - force (+shift force and energy) output clearing (every step).
284 These operations are issued in the local stream at the beginning of the step
285 and therefore always complete before the local kernel launch. The non-local
286 kernel is launched after the local on the same device/context hence it is
287 inherently scheduled after the operations in the local stream (including the
288 above "misc_ops") on pre-GK110 devices with single hardware queue, but on later
289 devices with multiple hardware queues the dependency needs to be enforced.
290 We use the misc_ops_and_local_H2D_done event to record the point where
291 the local x+q H2D (and all preceding) tasks are complete and synchronize
292 with this event in the non-local stream before launching the non-bonded kernel.
294 void nbnxn_gpu_launch_kernel(gmx_nbnxn_cuda_t *nb,
295 const nbnxn_atomdata_t *nbatom,
300 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
301 /* CUDA kernel launch-related stuff */
303 dim3 dim_block, dim_grid;
304 nbnxn_cu_kfunc_ptr_t nb_kernel = NULL; /* fn pointer to the nonbonded kernel */
306 cu_atomdata_t *adat = nb->atdat;
307 cu_nbparam_t *nbp = nb->nbparam;
308 cu_plist_t *plist = nb->plist[iloc];
309 cu_timers_t *t = nb->timers;
310 cudaStream_t stream = nb->stream[iloc];
312 bool bCalcEner = flags & GMX_FORCE_ENERGY;
313 bool bCalcFshift = flags & GMX_FORCE_VIRIAL;
314 bool bDoTime = nb->bDoTime;
316 /* Don't launch the non-local kernel if there is no work to do.
317 Doing the same for the local kernel is more complicated, since the
318 local part of the force array also depends on the non-local kernel.
319 So to avoid complicating the code and to reduce the risk of bugs,
320 we always call the local kernel, the local x+q copy and later (not in
321 this function) the stream wait, local f copyback and the f buffer
322 clearing. All these operations, except for the local interaction kernel,
323 are needed for the non-local interactions. The skip of the local kernel
324 call is taken care of later in this function. */
325 if (canSkipWork(nb, iloc))
327 plist->haveFreshList = false;
332 /* calculate the atom data index range based on locality */
336 adat_len = adat->natoms_local;
340 adat_begin = adat->natoms_local;
341 adat_len = adat->natoms - adat->natoms_local;
344 /* beginning of timed HtoD section */
347 t->nb_h2d[iloc].openTimingRegion(stream);
351 cu_copy_H2D_async(adat->xq + adat_begin, nbatom->x + adat_begin * 4,
352 adat_len * sizeof(*adat->xq), stream);
356 t->nb_h2d[iloc].closeTimingRegion(stream);
359 /* When we get here all misc operations issues in the local stream as well as
360 the local xq H2D are done,
361 so we record that in the local stream and wait for it in the nonlocal one. */
362 if (nb->bUseTwoStreams)
364 if (iloc == eintLocal)
366 stat = cudaEventRecord(nb->misc_ops_and_local_H2D_done, stream);
367 CU_RET_ERR(stat, "cudaEventRecord on misc_ops_and_local_H2D_done failed");
371 stat = cudaStreamWaitEvent(stream, nb->misc_ops_and_local_H2D_done, 0);
372 CU_RET_ERR(stat, "cudaStreamWaitEvent on misc_ops_and_local_H2D_done failed");
376 if (nbp->useDynamicPruning && plist->haveFreshList)
378 /* Prunes for rlistOuter and rlistInner, sets plist->haveFreshList=false
379 (TODO: ATM that's the way the timing accounting can distinguish between
380 separate prune kernel and combined force+prune, maybe we need a better way?).
382 nbnxn_gpu_launch_kernel_pruneonly(nb, iloc, 1);
385 if (plist->nsci == 0)
387 /* Don't launch an empty local kernel (not allowed with CUDA) */
391 /* beginning of timed nonbonded calculation section */
394 t->nb_k[iloc].openTimingRegion(stream);
397 /* get the pointer to the kernel flavor we need to use */
398 nb_kernel = select_nbnxn_kernel(nbp->eeltype,
401 (plist->haveFreshList && !nb->timers->didPrune[iloc]),
404 /* Kernel launch config:
405 * - The thread block dimensions match the size of i-clusters, j-clusters,
406 * and j-cluster concurrency, in x, y, and z, respectively.
407 * - The 1D block-grid contains as many blocks as super-clusters.
409 int num_threads_z = 1;
410 if (nb->dev_info->prop.major == 3 && nb->dev_info->prop.minor == 7)
414 nblock = calc_nb_kernel_nblock(plist->nsci, nb->dev_info);
416 KernelLaunchConfig config;
417 config.blockSize[0] = c_clSize;
418 config.blockSize[1] = c_clSize;
419 config.blockSize[2] = num_threads_z;
420 config.gridSize[0] = nblock;
421 config.sharedMemorySize = calc_shmem_required_nonbonded(num_threads_z, nb->dev_info, nbp);
422 config.stream = stream;
426 fprintf(debug, "Non-bonded GPU launch configuration:\n\tThread block: %zux%zux%zu\n\t"
427 "\tGrid: %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n"
429 config.blockSize[0], config.blockSize[1], config.blockSize[2],
430 config.gridSize[0], config.gridSize[1], plist->nsci*c_numClPerSupercl,
431 c_numClPerSupercl, plist->na_c,
432 config.sharedMemorySize);
435 auto *timingEvent = bDoTime ? t->nb_k[iloc].fetchNextEvent() : nullptr;
436 const auto kernelArgs = prepareGpuKernelArguments(nb_kernel, config, adat, nbp, plist, &bCalcFshift);
437 launchGpuKernel(nb_kernel, config, timingEvent, "k_calc_nb", kernelArgs);
441 t->nb_k[iloc].closeTimingRegion(stream);
444 if (GMX_NATIVE_WINDOWS)
446 /* Windows: force flushing WDDM queue */
447 cudaStreamQuery(stream);
451 /*! Calculates the amount of shared memory required by the CUDA kernel in use. */
452 static inline int calc_shmem_required_prune(const int num_threads_z)
456 /* i-atom x in shared memory */
457 shmem = c_numClPerSupercl * c_clSize * sizeof(float4);
458 /* cj in shared memory, for each warp separately */
459 shmem += num_threads_z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(int);
464 void nbnxn_gpu_launch_kernel_pruneonly(gmx_nbnxn_cuda_t *nb,
468 cu_atomdata_t *adat = nb->atdat;
469 cu_nbparam_t *nbp = nb->nbparam;
470 cu_plist_t *plist = nb->plist[iloc];
471 cu_timers_t *t = nb->timers;
472 cudaStream_t stream = nb->stream[iloc];
474 bool bDoTime = nb->bDoTime;
476 if (plist->haveFreshList)
478 GMX_ASSERT(numParts == 1, "With first pruning we expect 1 part");
480 /* Set rollingPruningNumParts to signal that it is not set */
481 plist->rollingPruningNumParts = 0;
482 plist->rollingPruningPart = 0;
486 if (plist->rollingPruningNumParts == 0)
488 plist->rollingPruningNumParts = numParts;
492 GMX_ASSERT(numParts == plist->rollingPruningNumParts, "It is not allowed to change numParts in between list generation steps");
496 /* Use a local variable for part and update in plist, so we can return here
497 * without duplicating the part increment code.
499 int part = plist->rollingPruningPart;
501 plist->rollingPruningPart++;
502 if (plist->rollingPruningPart >= plist->rollingPruningNumParts)
504 plist->rollingPruningPart = 0;
507 /* Compute the number of list entries to prune in this pass */
508 int numSciInPart = (plist->nsci - part)/numParts;
510 /* Don't launch the kernel if there is no work to do (not allowed with CUDA) */
511 if (numSciInPart <= 0)
513 plist->haveFreshList = false;
518 GpuRegionTimer *timer = nullptr;
521 timer = &(plist->haveFreshList ? t->prune_k[iloc] : t->rollingPrune_k[iloc]);
524 /* beginning of timed prune calculation section */
527 timer->openTimingRegion(stream);
530 /* Kernel launch config:
531 * - The thread block dimensions match the size of i-clusters, j-clusters,
532 * and j-cluster concurrency, in x, y, and z, respectively.
533 * - The 1D block-grid contains as many blocks as super-clusters.
535 int num_threads_z = c_cudaPruneKernelJ4Concurrency;
536 int nblock = calc_nb_kernel_nblock(numSciInPart, nb->dev_info);
537 KernelLaunchConfig config;
538 config.blockSize[0] = c_clSize;
539 config.blockSize[1] = c_clSize;
540 config.blockSize[2] = num_threads_z;
541 config.gridSize[0] = nblock;
542 config.sharedMemorySize = calc_shmem_required_prune(num_threads_z);
543 config.stream = stream;
547 fprintf(debug, "Pruning GPU kernel launch configuration:\n\tThread block: %zux%zux%zu\n\t"
548 "\tGrid: %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n"
550 config.blockSize[0], config.blockSize[1], config.blockSize[2],
551 config.gridSize[0], config.gridSize[1], numSciInPart*c_numClPerSupercl,
552 c_numClPerSupercl, plist->na_c,
553 config.sharedMemorySize);
556 auto *timingEvent = bDoTime ? timer->fetchNextEvent() : nullptr;
557 constexpr char kernelName[] = "k_pruneonly";
558 const auto &kernel = plist->haveFreshList ? nbnxn_kernel_prune_cuda<true> : nbnxn_kernel_prune_cuda<false>;
559 const auto kernelArgs = prepareGpuKernelArguments(kernel, config, adat, nbp, plist, &numParts, &part);
560 launchGpuKernel(kernel, config, timingEvent, kernelName, kernelArgs);
562 /* TODO: consider a more elegant way to track which kernel has been called
563 (combined or separate 1st pass prune, rolling prune). */
564 if (plist->haveFreshList)
566 plist->haveFreshList = false;
567 /* Mark that pruning has been done */
568 nb->timers->didPrune[iloc] = true;
572 /* Mark that rolling pruning has been done */
573 nb->timers->didRollingPrune[iloc] = true;
578 timer->closeTimingRegion(stream);
581 if (GMX_NATIVE_WINDOWS)
583 /* Windows: force flushing WDDM queue */
584 cudaStreamQuery(stream);
588 void nbnxn_gpu_launch_cpyback(gmx_nbnxn_cuda_t *nb,
589 const nbnxn_atomdata_t *nbatom,
594 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
596 /* determine interaction locality from atom locality */
597 int iloc = gpuAtomToInteractionLocality(aloc);
599 cu_atomdata_t *adat = nb->atdat;
600 cu_timers_t *t = nb->timers;
601 bool bDoTime = nb->bDoTime;
602 cudaStream_t stream = nb->stream[iloc];
604 bool bCalcEner = flags & GMX_FORCE_ENERGY;
605 bool bCalcFshift = flags & GMX_FORCE_VIRIAL;
607 /* don't launch non-local copy-back if there was no non-local work to do */
608 if (canSkipWork(nb, iloc))
613 getGpuAtomRange(adat, aloc, adat_begin, adat_len);
615 /* beginning of timed D2H section */
618 t->nb_d2h[iloc].openTimingRegion(stream);
621 /* With DD the local D2H transfer can only start after the non-local
622 kernel has finished. */
623 if (iloc == eintLocal && nb->bUseTwoStreams)
625 stat = cudaStreamWaitEvent(stream, nb->nonlocal_done, 0);
626 CU_RET_ERR(stat, "cudaStreamWaitEvent on nonlocal_done failed");
630 cu_copy_D2H_async(nbatom->out[0].f + adat_begin * 3, adat->f + adat_begin,
631 (adat_len)*sizeof(*adat->f), stream);
633 /* After the non-local D2H is launched the nonlocal_done event can be
634 recorded which signals that the local D2H can proceed. This event is not
635 placed after the non-local kernel because we want the non-local data
637 if (iloc == eintNonlocal)
639 stat = cudaEventRecord(nb->nonlocal_done, stream);
640 CU_RET_ERR(stat, "cudaEventRecord on nonlocal_done failed");
643 /* only transfer energies in the local stream */
649 cu_copy_D2H_async(nb->nbst.fshift, adat->fshift,
650 SHIFTS * sizeof(*nb->nbst.fshift), stream);
656 cu_copy_D2H_async(nb->nbst.e_lj, adat->e_lj,
657 sizeof(*nb->nbst.e_lj), stream);
658 cu_copy_D2H_async(nb->nbst.e_el, adat->e_el,
659 sizeof(*nb->nbst.e_el), stream);
665 t->nb_d2h[iloc].closeTimingRegion(stream);
669 void nbnxn_cuda_set_cacheconfig(const gmx_device_info_t *devinfo)
673 for (int i = 0; i < eelCuNR; i++)
675 for (int j = 0; j < evdwCuNR; j++)
677 if (devinfo->prop.major >= 3)
679 /* Default kernel on sm 3.x and later 32/32 kB Shared/L1 */
680 cudaFuncSetCacheConfig(nb_kfunc_ener_prune_ptr[i][j], cudaFuncCachePreferEqual);
681 cudaFuncSetCacheConfig(nb_kfunc_ener_noprune_ptr[i][j], cudaFuncCachePreferEqual);
682 cudaFuncSetCacheConfig(nb_kfunc_noener_prune_ptr[i][j], cudaFuncCachePreferEqual);
683 stat = cudaFuncSetCacheConfig(nb_kfunc_noener_noprune_ptr[i][j], cudaFuncCachePreferEqual);
687 /* On Fermi prefer L1 gives 2% higher performance */
688 /* Default kernel on sm_2.x 16/48 kB Shared/L1 */
689 cudaFuncSetCacheConfig(nb_kfunc_ener_prune_ptr[i][j], cudaFuncCachePreferL1);
690 cudaFuncSetCacheConfig(nb_kfunc_ener_noprune_ptr[i][j], cudaFuncCachePreferL1);
691 cudaFuncSetCacheConfig(nb_kfunc_noener_prune_ptr[i][j], cudaFuncCachePreferL1);
692 stat = cudaFuncSetCacheConfig(nb_kfunc_noener_noprune_ptr[i][j], cudaFuncCachePreferL1);
694 CU_RET_ERR(stat, "cudaFuncSetCacheConfig failed");