2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017, 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 * \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_data_mgmt.h"
61 #include "gromacs/mdlib/nbnxn_pairlist.h"
62 #include "gromacs/timing/gpu_timing.h"
63 #include "gromacs/utility/cstringutil.h"
64 #include "gromacs/utility/gmxassert.h"
66 #include "nbnxn_cuda_types.h"
69 * Texture references are created at compile-time and need to be declared
70 * at file scope as global variables (see http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#texture-reference-api).
71 * The texture references below are used in two translation units;
72 * we declare them here along the kernels that use them (when compiling legacy Fermi kernels),
73 * and provide getters (see below) used by the data_mgmt module where the
74 * textures are bound/unbound.
75 * (In principle we could do it the other way arond, but that would likely require
76 * device linking and we'd rather avoid technical hurdles.)
78 /*! Texture reference for LJ C6/C12 parameters; bound to cu_nbparam_t.nbfp */
79 texture<float, 1, cudaReadModeElementType> nbfp_texref;
81 /*! Texture reference for LJ-PME parameters; bound to cu_nbparam_t.nbfp_comb */
82 texture<float, 1, cudaReadModeElementType> nbfp_comb_texref;
84 /*! Texture reference for Ewald coulomb force table; bound to cu_nbparam_t.coulomb_tab */
85 texture<float, 1, cudaReadModeElementType> coulomb_tab_texref;
88 /***** The kernel declarations/definitions come here *****/
90 /* Top-level kernel declaration generation: will generate through multiple
91 * inclusion the following flavors for all kernel declarations:
92 * - force-only output;
93 * - force and energy output;
94 * - force-only with pair list pruning;
95 * - force and energy output with pair list pruning.
97 #define FUNCTION_DECLARATION_ONLY
99 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernels.cuh"
100 /** Force & energy **/
101 #define CALC_ENERGIES
102 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernels.cuh"
105 /*** Pair-list pruning kernels ***/
108 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernels.cuh"
109 /** Force & energy **/
110 #define CALC_ENERGIES
111 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernels.cuh"
115 /* Prune-only kernels */
116 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_pruneonly.cuh"
117 #undef FUNCTION_DECLARATION_ONLY
119 /* Now generate the function definitions if we are using a single compilation unit. */
120 #if GMX_CUDA_NB_SINGLE_COMPILATION_UNIT
121 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_F_noprune.cu"
122 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_F_prune.cu"
123 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_VF_noprune.cu"
124 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_VF_prune.cu"
125 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_pruneonly.cu"
127 /* Prevent compilation in multiple compilation unit mode for CC 2.x. Although we have
128 * build-time checks to prevent this, the user could manually tweaks nvcc flags
129 * which would lead to buggy kernels getting compiled.
131 #if GMX_PTX_ARCH > 0 && GMX_PTX_ARCH <= 210 && !defined(__clang__)
132 #error Due to an CUDA nvcc compiler bug, the CUDA non-bonded module can not be compiled with multiple compilation units for CC 2.x devices. If you have changed the nvcc flags manually, either use the GMX_CUDA_TARGET_* variables instead or set GMX_CUDA_NB_SINGLE_COMPILATION_UNIT=ON CMake option.
134 #endif /* GMX_CUDA_NB_SINGLE_COMPILATION_UNIT */
137 /*! Nonbonded kernel function pointer type */
138 typedef void (*nbnxn_cu_kfunc_ptr_t)(const cu_atomdata_t,
143 /*********************************/
145 /* XXX switch between chevron and cudaLaunch (supported only in CUDA >=7.0)
146 -- only for benchmarking purposes */
147 static const bool bUseCudaLaunchKernel =
148 (GMX_CUDA_VERSION >= 7000) && (getenv("GMX_DISABLE_CUDALAUNCH") == NULL);
150 /* XXX always/never run the energy/pruning kernels -- only for benchmarking purposes */
151 static bool always_ener = (getenv("GMX_GPU_ALWAYS_ENER") != NULL);
152 static bool never_ener = (getenv("GMX_GPU_NEVER_ENER") != NULL);
153 static bool always_prune = (getenv("GMX_GPU_ALWAYS_PRUNE") != NULL);
156 /*! Returns the number of blocks to be used for the nonbonded GPU kernel. */
157 static inline int calc_nb_kernel_nblock(int nwork_units, const gmx_device_info_t *dinfo)
162 /* CUDA does not accept grid dimension of 0 (which can happen e.g. with an
163 empty domain) and that case should be handled before this point. */
164 assert(nwork_units > 0);
166 max_grid_x_size = dinfo->prop.maxGridSize[0];
168 /* do we exceed the grid x dimension limit? */
169 if (nwork_units > max_grid_x_size)
171 gmx_fatal(FARGS, "Watch out, the input system is too large to simulate!\n"
172 "The number of nonbonded work units (=number of super-clusters) exceeds the"
173 "maximum grid size in x dimension (%d > %d)!", nwork_units, max_grid_x_size);
180 /* Constant arrays listing all kernel function pointers and enabling selection
181 of a kernel in an elegant manner. */
183 /*! Pointers to the non-bonded kernels organized in 2-dim arrays by:
184 * electrostatics and VDW type.
186 * Note that the row- and column-order of function pointers has to match the
187 * order of corresponding enumerated electrostatics and vdw types, resp.,
188 * defined in nbnxn_cuda_types.h.
191 /*! Force-only kernel function pointers. */
192 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_noener_noprune_ptr[eelCuNR][evdwCuNR] =
194 { 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 },
195 { 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 },
196 { 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 },
197 { 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 },
198 { 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 },
199 { 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 }
202 /*! Force + energy kernel function pointers. */
203 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_ener_noprune_ptr[eelCuNR][evdwCuNR] =
205 { 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 },
206 { 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 },
207 { 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 },
208 { 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 },
209 { 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 },
210 { 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 }
213 /*! Force + pruning kernel function pointers. */
214 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_noener_prune_ptr[eelCuNR][evdwCuNR] =
216 { 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 },
217 { 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 },
218 { 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 },
219 { 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 },
220 { 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 },
221 { 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 }
224 /*! Force + energy + pruning kernel function pointers. */
225 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_ener_prune_ptr[eelCuNR][evdwCuNR] =
227 { 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 },
228 { 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 },
229 { 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 },
230 { 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 },
231 { 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 },
232 { 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 }
235 /*! Return a pointer to the kernel version to be executed at the current step. */
236 static inline nbnxn_cu_kfunc_ptr_t select_nbnxn_kernel(int eeltype,
240 const gmx_device_info_t gmx_unused *devInfo)
242 nbnxn_cu_kfunc_ptr_t res;
244 GMX_ASSERT(eeltype < eelCuNR,
245 "The electrostatics type requested is not implemented in the CUDA kernels.");
246 GMX_ASSERT(evdwtype < evdwCuNR,
247 "The VdW type requested is not implemented in the CUDA kernels.");
249 /* assert assumptions made by the kernels */
250 GMX_ASSERT(c_nbnxnGpuClusterSize*c_nbnxnGpuClusterSize/c_nbnxnGpuClusterpairSplit == devInfo->prop.warpSize,
251 "The CUDA kernels require the cluster_size_i*cluster_size_j/nbnxn_gpu_clusterpair_split to match the warp size of the architecture targeted.");
257 res = nb_kfunc_ener_prune_ptr[eeltype][evdwtype];
261 res = nb_kfunc_ener_noprune_ptr[eeltype][evdwtype];
268 res = nb_kfunc_noener_prune_ptr[eeltype][evdwtype];
272 res = nb_kfunc_noener_noprune_ptr[eeltype][evdwtype];
279 /*! \brief Calculates the amount of shared memory required by the nonbonded kernel in use. */
280 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)
286 /* size of shmem (force-buffers/xq/atom type preloading) */
287 /* NOTE: with the default kernel on sm3.0 we need shmem only for pre-loading */
288 /* i-atom x+q in shared memory */
289 shmem = c_numClPerSupercl * c_clSize * sizeof(float4);
290 /* cj in shared memory, for each warp separately */
291 shmem += num_threads_z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(int);
292 if (dinfo->prop.major >= 3)
294 if (nbp->vdwtype == evdwCuCUTCOMBGEOM ||
295 nbp->vdwtype == evdwCuCUTCOMBLB)
297 /* i-atom LJ combination parameters in shared memory */
298 shmem += c_numClPerSupercl * c_clSize * sizeof(float2);
302 /* i-atom types in shared memory */
303 shmem += c_numClPerSupercl * c_clSize * sizeof(int);
306 if (dinfo->prop.major < 3)
308 /* force reduction buffers in shared memory */
309 shmem += c_clSize * c_clSize * 3 * sizeof(float);
314 /*! As we execute nonbonded workload in separate streams, before launching
315 the kernel we need to make sure that he following operations have completed:
316 - atomdata allocation and related H2D transfers (every nstlist step);
317 - pair list H2D transfer (every nstlist step);
318 - shift vector H2D transfer (every nstlist step);
319 - force (+shift force and energy) output clearing (every step).
321 These operations are issued in the local stream at the beginning of the step
322 and therefore always complete before the local kernel launch. The non-local
323 kernel is launched after the local on the same device/context hence it is
324 inherently scheduled after the operations in the local stream (including the
325 above "misc_ops") on pre-GK110 devices with single hardware queue, but on later
326 devices with multiple hardware queues the dependency needs to be enforced.
327 We use the misc_ops_and_local_H2D_done event to record the point where
328 the local x+q H2D (and all preceding) tasks are complete and synchronize
329 with this event in the non-local stream before launching the non-bonded kernel.
331 void nbnxn_gpu_launch_kernel(gmx_nbnxn_cuda_t *nb,
332 const nbnxn_atomdata_t *nbatom,
337 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
338 /* CUDA kernel launch-related stuff */
340 dim3 dim_block, dim_grid;
341 nbnxn_cu_kfunc_ptr_t nb_kernel = NULL; /* fn pointer to the nonbonded kernel */
343 cu_atomdata_t *adat = nb->atdat;
344 cu_nbparam_t *nbp = nb->nbparam;
345 cu_plist_t *plist = nb->plist[iloc];
346 cu_timers_t *t = nb->timers;
347 cudaStream_t stream = nb->stream[iloc];
349 bool bCalcEner = flags & GMX_FORCE_ENERGY;
350 bool bCalcFshift = flags & GMX_FORCE_VIRIAL;
351 bool bDoTime = nb->bDoTime;
353 /* turn energy calculation always on/off (for debugging/testing only) */
354 bCalcEner = (bCalcEner || always_ener) && !never_ener;
356 /* Don't launch the non-local kernel if there is no work to do.
357 Doing the same for the local kernel is more complicated, since the
358 local part of the force array also depends on the non-local kernel.
359 So to avoid complicating the code and to reduce the risk of bugs,
360 we always call the local kernel, the local x+q copy and later (not in
361 this function) the stream wait, local f copyback and the f buffer
362 clearing. All these operations, except for the local interaction kernel,
363 are needed for the non-local interactions. The skip of the local kernel
364 call is taken care of later in this function. */
365 if (canSkipWork(nb, iloc))
367 plist->haveFreshList = false;
372 /* calculate the atom data index range based on locality */
376 adat_len = adat->natoms_local;
380 adat_begin = adat->natoms_local;
381 adat_len = adat->natoms - adat->natoms_local;
384 /* beginning of timed HtoD section */
387 t->nb_h2d[iloc].openTimingRegion(stream);
391 cu_copy_H2D_async(adat->xq + adat_begin, nbatom->x + adat_begin * 4,
392 adat_len * sizeof(*adat->xq), stream);
396 t->nb_h2d[iloc].closeTimingRegion(stream);
399 /* When we get here all misc operations issues in the local stream as well as
400 the local xq H2D are done,
401 so we record that in the local stream and wait for it in the nonlocal one. */
402 if (nb->bUseTwoStreams)
404 if (iloc == eintLocal)
406 stat = cudaEventRecord(nb->misc_ops_and_local_H2D_done, stream);
407 CU_RET_ERR(stat, "cudaEventRecord on misc_ops_and_local_H2D_done failed");
411 stat = cudaStreamWaitEvent(stream, nb->misc_ops_and_local_H2D_done, 0);
412 CU_RET_ERR(stat, "cudaStreamWaitEvent on misc_ops_and_local_H2D_done failed");
416 if (nbp->useDynamicPruning && plist->haveFreshList)
418 /* Prunes for rlistOuter and rlistInner, sets plist->haveFreshList=false
419 (TODO: ATM that's the way the timing accounting can distinguish between
420 separate prune kernel and combined force+prune, maybe we need a better way?).
422 nbnxn_gpu_launch_kernel_pruneonly(nb, iloc, 1);
425 if (plist->nsci == 0)
427 /* Don't launch an empty local kernel (not allowed with CUDA) */
431 /* beginning of timed nonbonded calculation section */
434 t->nb_k[iloc].openTimingRegion(stream);
437 /* get the pointer to the kernel flavor we need to use */
438 nb_kernel = select_nbnxn_kernel(nbp->eeltype,
441 (plist->haveFreshList && !nb->timers->didPrune[iloc]) || always_prune,
444 /* Kernel launch config:
445 * - The thread block dimensions match the size of i-clusters, j-clusters,
446 * and j-cluster concurrency, in x, y, and z, respectively.
447 * - The 1D block-grid contains as many blocks as super-clusters.
449 int num_threads_z = 1;
450 if (nb->dev_info->prop.major == 3 && nb->dev_info->prop.minor == 7)
454 nblock = calc_nb_kernel_nblock(plist->nsci, nb->dev_info);
455 dim_block = dim3(c_clSize, c_clSize, num_threads_z);
456 dim_grid = dim3(nblock, 1, 1);
457 shmem = calc_shmem_required_nonbonded(num_threads_z, nb->dev_info, nbp);
461 fprintf(debug, "Non-bonded GPU launch configuration:\n\tThread block: %ux%ux%u\n\t"
462 "\tGrid: %ux%u\n\t#Super-clusters/clusters: %d/%d (%d)\n"
464 dim_block.x, dim_block.y, dim_block.z,
465 dim_grid.x, dim_grid.y, plist->nsci*c_numClPerSupercl,
466 c_numClPerSupercl, plist->na_c,
470 if (bUseCudaLaunchKernel)
472 gmx_unused void* kernel_args[4];
473 kernel_args[0] = adat;
474 kernel_args[1] = nbp;
475 kernel_args[2] = plist;
476 kernel_args[3] = &bCalcFshift;
478 #if GMX_CUDA_VERSION >= 7000
479 cudaLaunchKernel((void *)nb_kernel, dim_grid, dim_block, kernel_args, shmem, stream);
484 nb_kernel<<< dim_grid, dim_block, shmem, stream>>> (*adat, *nbp, *plist, bCalcFshift);
486 CU_LAUNCH_ERR("k_calc_nb");
490 t->nb_k[iloc].closeTimingRegion(stream);
493 #if (defined(WIN32) || defined( _WIN32 ))
494 /* Windows: force flushing WDDM queue */
495 stat = cudaStreamQuery(stream);
499 /*! Calculates the amount of shared memory required by the CUDA kernel in use. */
500 static inline int calc_shmem_required_prune(const int num_threads_z)
504 /* i-atom x in shared memory */
505 shmem = c_numClPerSupercl * c_clSize * sizeof(float4);
506 /* cj in shared memory, for each warp separately */
507 shmem += num_threads_z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(int);
512 void nbnxn_gpu_launch_kernel_pruneonly(gmx_nbnxn_cuda_t *nb,
516 cu_atomdata_t *adat = nb->atdat;
517 cu_nbparam_t *nbp = nb->nbparam;
518 cu_plist_t *plist = nb->plist[iloc];
519 cu_timers_t *t = nb->timers;
520 cudaStream_t stream = nb->stream[iloc];
522 bool bDoTime = nb->bDoTime;
524 if (plist->haveFreshList)
526 GMX_ASSERT(numParts == 1, "With first pruning we expect 1 part");
528 /* Set rollingPruningNumParts to signal that it is not set */
529 plist->rollingPruningNumParts = 0;
530 plist->rollingPruningPart = 0;
534 if (plist->rollingPruningNumParts == 0)
536 plist->rollingPruningNumParts = numParts;
540 GMX_ASSERT(numParts == plist->rollingPruningNumParts, "It is not allowed to change numParts in between list generation steps");
544 /* Use a local variable for part and update in plist, so we can return here
545 * without duplicating the part increment code.
547 int part = plist->rollingPruningPart;
549 plist->rollingPruningPart++;
550 if (plist->rollingPruningPart >= plist->rollingPruningNumParts)
552 plist->rollingPruningPart = 0;
555 /* Compute the number of list entries to prune in this pass */
556 int numSciInPart = (plist->nsci - part)/numParts;
558 /* Don't launch the kernel if there is no work to do (not allowed with CUDA) */
559 if (numSciInPart <= 0)
561 plist->haveFreshList = false;
566 GpuRegionTimer *timer = nullptr;
569 timer = &(plist->haveFreshList ? t->prune_k[iloc] : t->rollingPrune_k[iloc]);
572 /* beginning of timed prune calculation section */
575 timer->openTimingRegion(stream);
578 /* Kernel launch config:
579 * - The thread block dimensions match the size of i-clusters, j-clusters,
580 * and j-cluster concurrency, in x, y, and z, respectively.
581 * - The 1D block-grid contains as many blocks as super-clusters.
583 int num_threads_z = c_cudaPruneKernelJ4Concurrency;
584 int nblock = calc_nb_kernel_nblock(numSciInPart, nb->dev_info);
585 dim3 dim_block = dim3(c_clSize, c_clSize, num_threads_z);
586 dim3 dim_grid = dim3(nblock, 1, 1);
587 int shmem = calc_shmem_required_prune(num_threads_z);
591 fprintf(debug, "Pruning GPU kernel launch configuration:\n\tThread block: %dx%dx%d\n\t"
592 "\tGrid: %dx%d\n\t#Super-clusters/clusters: %d/%d (%d)\n"
594 dim_block.x, dim_block.y, dim_block.z,
595 dim_grid.x, dim_grid.y, numSciInPart*c_numClPerSupercl,
596 c_numClPerSupercl, plist->na_c,
600 if (bUseCudaLaunchKernel)
602 gmx_unused void* kernel_args[5];
603 kernel_args[0] = adat;
604 kernel_args[1] = nbp;
605 kernel_args[2] = plist;
606 kernel_args[3] = &numParts;
607 kernel_args[4] = ∂
609 #if GMX_CUDA_VERSION >= 7000
610 if (plist->haveFreshList)
612 cudaLaunchKernel((void *)nbnxn_kernel_prune_cuda<true>, dim_grid, dim_block, kernel_args, shmem, stream);
616 cudaLaunchKernel((void *)nbnxn_kernel_prune_cuda<false>, dim_grid, dim_block, kernel_args, shmem, stream);
622 if (plist->haveFreshList)
624 nbnxn_kernel_prune_cuda<true><<< dim_grid, dim_block, shmem, stream>>> (*adat, *nbp, *plist, numParts, part);
628 nbnxn_kernel_prune_cuda<false><<< dim_grid, dim_block, shmem, stream>>> (*adat, *nbp, *plist, numParts, part);
631 CU_LAUNCH_ERR("k_pruneonly");
633 /* TODO: consider a more elegant way to track which kernel has been called
634 (combined or separate 1st pass prune, rolling prune). */
635 if (plist->haveFreshList)
637 plist->haveFreshList = false;
638 /* Mark that pruning has been done */
639 nb->timers->didPrune[iloc] = true;
643 /* Mark that rolling pruning has been done */
644 nb->timers->didRollingPrune[iloc] = true;
649 timer->closeTimingRegion(stream);
652 #if (defined(WIN32) || defined( _WIN32 ))
653 /* Windows: force flushing WDDM queue */
654 stat = cudaStreamQuery(stream);
658 void nbnxn_gpu_launch_cpyback(gmx_nbnxn_cuda_t *nb,
659 const nbnxn_atomdata_t *nbatom,
664 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
667 /* determine interaction locality from atom locality */
672 else if (NONLOCAL_A(aloc))
679 sprintf(stmp, "Invalid atom locality passed (%d); valid here is only "
680 "local (%d) or nonlocal (%d)", aloc, eatLocal, eatNonlocal);
684 cu_atomdata_t *adat = nb->atdat;
685 cu_timers_t *t = nb->timers;
686 bool bDoTime = nb->bDoTime;
687 cudaStream_t stream = nb->stream[iloc];
689 bool bCalcEner = flags & GMX_FORCE_ENERGY;
690 bool bCalcFshift = flags & GMX_FORCE_VIRIAL;
692 /* don't launch non-local copy-back if there was no non-local work to do */
693 if (canSkipWork(nb, iloc))
698 /* calculate the atom data index range based on locality */
702 adat_len = adat->natoms_local;
706 adat_begin = adat->natoms_local;
707 adat_len = adat->natoms - adat->natoms_local;
710 /* beginning of timed D2H section */
713 t->nb_d2h[iloc].openTimingRegion(stream);
716 /* With DD the local D2H transfer can only start after the non-local
717 kernel has finished. */
718 if (iloc == eintLocal && nb->bUseTwoStreams)
720 stat = cudaStreamWaitEvent(stream, nb->nonlocal_done, 0);
721 CU_RET_ERR(stat, "cudaStreamWaitEvent on nonlocal_done failed");
725 cu_copy_D2H_async(nbatom->out[0].f + adat_begin * 3, adat->f + adat_begin,
726 (adat_len)*sizeof(*adat->f), stream);
728 /* After the non-local D2H is launched the nonlocal_done event can be
729 recorded which signals that the local D2H can proceed. This event is not
730 placed after the non-local kernel because we want the non-local data
732 if (iloc == eintNonlocal)
734 stat = cudaEventRecord(nb->nonlocal_done, stream);
735 CU_RET_ERR(stat, "cudaEventRecord on nonlocal_done failed");
738 /* only transfer energies in the local stream */
744 cu_copy_D2H_async(nb->nbst.fshift, adat->fshift,
745 SHIFTS * sizeof(*nb->nbst.fshift), stream);
751 cu_copy_D2H_async(nb->nbst.e_lj, adat->e_lj,
752 sizeof(*nb->nbst.e_lj), stream);
753 cu_copy_D2H_async(nb->nbst.e_el, adat->e_el,
754 sizeof(*nb->nbst.e_el), stream);
760 t->nb_d2h[iloc].closeTimingRegion(stream);
764 /*! \brief Count pruning kernel time if either kernel has been triggered
766 * We do the accounting for either of the two pruning kernel flavors:
767 * - 1st pass prune: ran during the current step (prior to the force kernel);
768 * - rolling prune: ran at the end of the previous step (prior to the current step H2D xq);
770 * Note that the resetting of cu_timers_t::didPrune and cu_timers_t::didRollingPrune should happen
771 * after calling this function.
773 * \param[in] timers structs with CUDA timer objects
774 * \param[inout] timings GPU task timing data
775 * \param[in] iloc interaction locality
777 static void countPruneKernelTime(cu_timers_t *timers,
778 gmx_wallclock_gpu_t *timings,
781 // We might have not done any pruning (e.g. if we skipped with empty domains).
782 if (!timers->didPrune[iloc] && !timers->didRollingPrune[iloc])
787 if (timers->didPrune[iloc])
789 timings->pruneTime.c++;
790 timings->pruneTime.t += timers->prune_k[iloc].getLastRangeTime();
792 if (timers->didRollingPrune[iloc])
794 timings->dynamicPruneTime.c++;
795 timings->dynamicPruneTime.t += timers->rollingPrune_k[iloc].getLastRangeTime();
799 void nbnxn_gpu_wait_for_gpu(gmx_nbnxn_cuda_t *nb,
801 real *e_lj, real *e_el, rvec *fshift)
803 /* NOTE: only implemented for single-precision at this time */
807 /* determine interaction locality from atom locality */
812 else if (NONLOCAL_A(aloc))
819 sprintf(stmp, "Invalid atom locality passed (%d); valid here is only "
820 "local (%d) or nonlocal (%d)", aloc, eatLocal, eatNonlocal);
824 cu_plist_t *plist = nb->plist[iloc];
825 cu_timers_t *timers = nb->timers;
826 struct gmx_wallclock_gpu_t *timings = nb->timings;
827 nb_staging nbst = nb->nbst;
829 bool bCalcEner = flags & GMX_FORCE_ENERGY;
830 bool bCalcFshift = flags & GMX_FORCE_VIRIAL;
832 /* turn energy calculation always on/off (for debugging/testing only) */
833 bCalcEner = (bCalcEner || always_ener) && !never_ener;
835 /* Launch wait/update timers & counters and do reduction into staging buffers
836 BUT skip it when during the non-local phase there was actually no work to do.
837 This is consistent with nbnxn_gpu_launch_kernel.
839 NOTE: if timing with multiple GPUs (streams) becomes possible, the
840 counters could end up being inconsistent due to not being incremented
841 on some of the nodes! */
842 if (!canSkipWork(nb, iloc))
844 stat = cudaStreamSynchronize(nb->stream[iloc]);
845 CU_RET_ERR(stat, "cudaStreamSynchronize failed in cu_blockwait_nb");
847 /* timing data accumulation */
850 /* only increase counter once (at local F wait) */
854 timings->ktime[plist->haveFreshList ? 1 : 0][bCalcEner ? 1 : 0].c += 1;
858 timings->ktime[plist->haveFreshList ? 1 : 0][bCalcEner ? 1 : 0].t += timers->nb_k[iloc].getLastRangeTime();
860 /* X/q H2D and F D2H timings */
861 timings->nb_h2d_t += timers->nb_h2d[iloc].getLastRangeTime();
862 timings->nb_d2h_t += timers->nb_d2h[iloc].getLastRangeTime();
864 /* Count the pruning kernel times for both cases:1st pass (at search step)
865 and rolling pruning (if called at the previous step).
866 We do the accounting here as this is the only sync point where we
867 know (without checking or additional sync-ing) that prune tasks in
868 in the current stream have completed (having just blocking-waited
869 for the force D2H). */
870 countPruneKernelTime(timers, timings, iloc);
872 /* only count atdat and pair-list H2D at pair-search step */
873 if (timers->didPairlistH2D[iloc])
875 /* atdat transfer timing (add only once, at local F wait) */
879 timings->pl_h2d_t += timers->atdat.getLastRangeTime();
882 timings->pl_h2d_t += timers->pl_h2d[iloc].getLastRangeTime();
884 /* Clear the timing flag for the next step */
885 timers->didPairlistH2D[iloc] = false;
889 /* add up energies and shift forces (only once at local F wait) */
900 for (int i = 0; i < SHIFTS; i++)
902 fshift[i][0] += nbst.fshift[i].x;
903 fshift[i][1] += nbst.fshift[i].y;
904 fshift[i][2] += nbst.fshift[i].z;
910 /* Always reset both pruning flags (doesn't hurt doing it even when timing is off). */
911 timers->didPrune[iloc] = timers->didRollingPrune[iloc] = false;
913 /* Turn off initial list pruning (doesn't hurt if this is not pair-search step). */
914 plist->haveFreshList = false;
917 const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_nbfp_texref()
922 const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_nbfp_comb_texref()
924 return nbfp_comb_texref;
927 const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_coulomb_tab_texref()
929 return coulomb_tab_texref;
932 void nbnxn_cuda_set_cacheconfig(const gmx_device_info_t *devinfo)
936 for (int i = 0; i < eelCuNR; i++)
938 for (int j = 0; j < evdwCuNR; j++)
940 if (devinfo->prop.major >= 3)
942 /* Default kernel on sm 3.x and later 32/32 kB Shared/L1 */
943 cudaFuncSetCacheConfig(nb_kfunc_ener_prune_ptr[i][j], cudaFuncCachePreferEqual);
944 cudaFuncSetCacheConfig(nb_kfunc_ener_noprune_ptr[i][j], cudaFuncCachePreferEqual);
945 cudaFuncSetCacheConfig(nb_kfunc_noener_prune_ptr[i][j], cudaFuncCachePreferEqual);
946 stat = cudaFuncSetCacheConfig(nb_kfunc_noener_noprune_ptr[i][j], cudaFuncCachePreferEqual);
950 /* On Fermi prefer L1 gives 2% higher performance */
951 /* Default kernel on sm_2.x 16/48 kB Shared/L1 */
952 cudaFuncSetCacheConfig(nb_kfunc_ener_prune_ptr[i][j], cudaFuncCachePreferL1);
953 cudaFuncSetCacheConfig(nb_kfunc_ener_noprune_ptr[i][j], cudaFuncCachePreferL1);
954 cudaFuncSetCacheConfig(nb_kfunc_noener_prune_ptr[i][j], cudaFuncCachePreferL1);
955 stat = cudaFuncSetCacheConfig(nb_kfunc_noener_noprune_ptr[i][j], cudaFuncCachePreferL1);
957 CU_RET_ERR(stat, "cudaFuncSetCacheConfig failed");