2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2014,2015,2016,2017,2018, 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.
35 #ifndef CUDA_ARCH_UTILS_CUH_
36 #define CUDA_ARCH_UTILS_CUH_
40 #include "gromacs/utility/basedefinitions.h"
43 * \brief CUDA arch dependent definitions.
45 * \author Szilard Pall <pall.szilard@gmail.com>
48 /* GMX_PTX_ARCH is set to the virtual arch (PTX) version targeted by
49 * the current compiler pass or zero for the host pass and it is
50 * intended to be used instead of __CUDA_ARCH__.
53 #define GMX_PTX_ARCH 0
55 #define GMX_PTX_ARCH __CUDA_ARCH__
58 /* Until CC 5.2 and likely for the near future all NVIDIA architectures
59 have a warp size of 32, but this could change later. If it does, the
60 following constants should depend on the value of GMX_PTX_ARCH.
62 static const int warp_size = 32;
63 static const int warp_size_log2 = 5;
64 /*! \brief Bitmask corresponding to all threads active in a warp.
65 * NOTE that here too we assume 32-wide warps.
67 static const unsigned int c_fullWarpMask = 0xffffffff;
69 /* Below are backward-compatibility wrappers for CUDA 9 warp-wide intrinsics. */
71 /*! \brief Compatibility wrapper around the CUDA __syncwarp() instrinsic. */
72 static __forceinline__ __device__
73 void gmx_syncwarp(const unsigned int activeMask = c_fullWarpMask)
75 #if GMX_CUDA_VERSION < 9000
76 /* no sync needed on pre-Volta. */
77 GMX_UNUSED_VALUE(activeMask);
79 __syncwarp(activeMask);
83 /*! \brief Compatibility wrapper around the CUDA __ballot()/__ballot_sync() instrinsic. */
84 static __forceinline__ __device__
85 unsigned int gmx_ballot_sync(const unsigned int activeMask,
88 #if GMX_CUDA_VERSION < 9000
89 GMX_UNUSED_VALUE(activeMask);
90 return __ballot(pred);
92 return __ballot_sync(activeMask, pred);
96 /*! \brief Compatibility wrapper around the CUDA __any()/__any_sync() instrinsic. */
97 static __forceinline__ __device__
98 int gmx_any_sync(const unsigned int activeMask,
101 #if GMX_CUDA_VERSION < 9000
102 GMX_UNUSED_VALUE(activeMask);
105 return __any_sync(activeMask, pred);
109 /*! \brief Compatibility wrapper around the CUDA __shfl_up()/__shfl_up_sync() instrinsic. */
110 template <typename T>
111 static __forceinline__ __device__
112 T gmx_shfl_up_sync(const unsigned int activeMask,
115 int width = warp_size)
117 #if GMX_CUDA_VERSION < 9000
118 GMX_UNUSED_VALUE(activeMask);
119 return __shfl_up(var, offset, width);
121 return __shfl_up_sync(activeMask, var, offset, width);
125 /*! \brief Compatibility wrapper around the CUDA __shfl_down()/__shfl_down_sync() instrinsic. */
126 template <typename T>
127 static __forceinline__ __device__
128 T gmx_shfl_down_sync(const unsigned int activeMask,
131 int width = warp_size)
133 #if GMX_CUDA_VERSION < 9000
134 GMX_UNUSED_VALUE(activeMask);
135 return __shfl_down(var, offset, width);
137 return __shfl_down_sync(activeMask, var, offset, width);
141 /*! \brief Allow disabling CUDA textures using the GMX_DISABLE_CUDA_TEXTURES macro.
143 * Only texture objects supported.
144 * Disable texture support missing in clang (all versions up to <=5.0-dev as of writing).
146 * This option will not influence functionality. All features using textures ought
147 * to have fallback for texture-less reads (direct/LDG loads), all new code needs
148 * to provide fallback code.
150 #if defined(GMX_DISABLE_CUDA_TEXTURES) || (defined(__clang__) && defined(__CUDA__))
151 #define DISABLE_CUDA_TEXTURES 1
153 #define DISABLE_CUDA_TEXTURES 0
156 /*! \brief True if the use of texture fetch in the CUDA kernels is disabled. */
157 static const bool c_disableCudaTextures = DISABLE_CUDA_TEXTURES;
160 /* CUDA architecture technical characteristics. Needs macros because it is used
161 * in the __launch_bounds__ function qualifiers and might need it in preprocessor
166 #if GMX_PTX_ARCH <= 370 // CC 3.x
167 #define GMX_CUDA_MAX_BLOCKS_PER_MP 16
168 #define GMX_CUDA_MAX_THREADS_PER_MP 2048
170 /* Note that this final branch covers all future architectures (current gen
171 * is 6.x as of writing), hence assuming that these *currently defined* upper
172 * limits will not be lowered.
174 #define GMX_CUDA_MAX_BLOCKS_PER_MP 32
175 #define GMX_CUDA_MAX_THREADS_PER_MP 2048
178 #define GMX_CUDA_MAX_BLOCKS_PER_MP 0
179 #define GMX_CUDA_MAX_THREADS_PER_MP 0
182 #endif /* CUDA_ARCH_UTILS_CUH_ */