2 # This file is part of the GROMACS molecular simulation package.
4 # Copyright (c) 2012,2013,2014,2015,2016, 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 # If the user did not set GMX_GPU we'll consider this option to be
36 # in "auto" mode meaning that we will:
37 # - search for CUDA and set GMX_GPU=ON we find it
38 # - check whether GPUs are present
39 # - if CUDA is not found but GPUs were detected issue a warning
40 if (NOT DEFINED GMX_GPU)
41 set(GMX_GPU_AUTO TRUE CACHE INTERNAL "GPU acceleration will be selected automatically")
43 option(GMX_GPU "Enable GPU acceleration" OFF)
45 if(GMX_GPU AND GMX_DOUBLE)
46 message(FATAL_ERROR "GPU acceleration is not available in double precision!")
48 if(GMX_GPU_AUTO AND GMX_DOUBLE)
49 message(WARNING "GPU acceleration is not available in double precision, disabled!")
50 set_property(CACHE GMX_GPU PROPERTY VALUE OFF)
51 set_property(CACHE GMX_GPU_AUTO PROPERTY VALUE OFF)
54 # detect GPUs in the build host machine
55 if ((GMX_GPU OR GMX_GPU_AUTO) AND NOT GMX_GPU_DETECTION_DONE)
60 # We need to call find_package even when we've already done the detection/setup
61 if(GMX_GPU OR GMX_GPU_AUTO)
62 if(NOT GMX_GPU AND NOT GMX_DETECT_GPU_AVAILABLE)
63 # Stay quiet when detection has occured and found no GPU.
64 # Noise is acceptable when there is a GPU or the user required one.
65 set(FIND_CUDA_QUIETLY QUIET)
68 # Cmake tries to use the static cuda runtime by default,
69 # but this leads to unusable GPU builds on OS X.
71 set(CUDA_USE_STATIC_CUDA_RUNTIME OFF CACHE STRING "Use the static version of the CUDA runtime library if available")
74 find_package(CUDA ${REQUIRED_CUDA_VERSION} ${FIND_CUDA_QUIETLY})
77 # Depending on the current vale of GMX_GPU and GMX_GPU_AUTO:
78 # - OFF, FALSE: Will skip this detection/setup.
79 # - OFF, TRUE : Will keep GMX_GPU=OFF if no CUDA is detected, but will assemble
80 # a warning message which will be issued at the end of the
81 # configuration if GPU(s) were found in the build system.
82 # - ON , FALSE: The user requested GPU build and this requires CUDA, so we will
83 # fail if it is not available.
84 # - ON , TRUE : Can't happen (GMX_GPU=ON can only be user-set at this point)
85 if((GMX_GPU OR GMX_GPU_AUTO) AND NOT GMX_GPU_DETECTION_DONE)
86 if (EXISTS ${CUDA_TOOLKIT_ROOT_DIR})
87 set(CUDA_FOUND TRUE CACHE INTERNAL "Whether the CUDA toolkit was found" FORCE)
89 set(CUDA_FOUND FALSE CACHE INTERNAL "Whether the CUDA toolkit was found" FORCE)
92 # assemble warning/error message
93 if (GMX_DETECT_GPU_AVAILABLE)
94 set(_msg "${GMX_DETECT_GPU_COUNT} NVIDIA GPU(s) found in the system")
97 if (NOT GMX_DETECT_GPU_INFO STREQUAL "")
99 foreach(gpu ${GMX_DETECT_GPU_INFO})
105 # TODO remove the second part of the message when we'll have compute
106 # capability information from the detection.
108 Compute capability information not available, consult the NVIDIA website:
109 https://developer.nvidia.com/cuda-gpus")
112 set(CUDA_NOTFOUND_MESSAGE "mdrun supports native GPU acceleration on NVIDIA hardware with compute capability >= ${REQUIRED_CUDA_COMPUTE_CAPABILITY} (Fermi or later). This requires the NVIDIA CUDA toolkit, which was not found. Its location can be hinted by setting the CUDA_TOOLKIT_ROOT_DIR CMake option (does not work as an environment variable). The typical location would be /usr/local/cuda[-version]. Note that CPU or GPU acceleration can be selected at runtime.
119 # Disable GPU acceleration in auto mode
120 message(STATUS "No compatible CUDA toolkit found (v5.0+), disabling native GPU acceleration")
121 set_property(CACHE GMX_GPU PROPERTY VALUE OFF)
122 set(CUDA_NOTFOUND_AUTO ON)
124 # the user requested CUDA, but it wasn't found
125 message(FATAL_ERROR "${CUDA_NOTFOUND_MESSAGE}")
129 message(STATUS "Enabling native GPU acceleration")
130 set_property(CACHE GMX_GPU PROPERTY VALUE ON)
132 endif() # NOT CUDA_FOUND
135 # Try to find NVML if a GPU accelerated binary should be build.
137 if (DEFINED NVML_LIBRARY)
138 set(NVML_FIND_QUIETLY TRUE)
141 option(GMX_USE_NVML "Use NVML support for better CUDA performance" ${NVML_FOUND})
142 mark_as_advanced(GMX_USE_NVML)
145 include_directories(SYSTEM ${NVML_INCLUDE_DIR})
147 list(APPEND GMX_EXTRA_LIBRARIES ${NVML_LIBRARY})
149 message(FATAL_ERROR "NVML support was required, but was not detected. Please consult the install guide.")
154 # Annoyingly enough, FindCUDA leaves a few variables behind as non-advanced.
155 # We need to mark these advanced outside the conditional, otherwise, if the
156 # user turns GMX_GPU=OFF after a failed cmake pass, these variables will be
157 # left behind in the cache.
158 mark_as_advanced(CUDA_SDK_ROOT_DIR
159 CUDA_USE_STATIC_CUDA_RUNTIME
160 CUDA_dl_LIBRARY CUDA_rt_LIBRARY
163 mark_as_advanced(CUDA_TOOLKIT_ROOT_DIR)
164 mark_as_advanced(CUDA_HOST_COMPILER)
167 # Try to execute ${CUDA_NVCC_EXECUTABLE} --version and set the output
168 # (or an error string) in the argument variable.
169 # Note that semicolon is used as separator for nvcc.
172 # COMPILER_INFO - [output variable] string with compiler path, ID and
173 # some compiler-provided information
174 # COMPILER_FLAGS - [output variable] flags for the compiler
176 macro(get_cuda_compiler_info COMPILER_INFO COMPILER_FLAGS)
177 if(CUDA_NVCC_EXECUTABLE)
179 # Get the nvcc version string. This is multi-line, but since it is only 4 lines
180 # and might change in the future it is better to store than trying to parse out
181 # the version from the current format.
182 execute_process(COMMAND ${CUDA_NVCC_EXECUTABLE} --version
183 RESULT_VARIABLE _nvcc_version_res
184 OUTPUT_VARIABLE _nvcc_version_out
185 ERROR_VARIABLE _nvcc_version_err
186 OUTPUT_STRIP_TRAILING_WHITESPACE)
187 if (${_nvcc_version_res} EQUAL 0)
188 # Fix multi-line mess: Replace newline with ";" so we can use it in a define
189 string(REPLACE "\n" ";" _nvcc_info_singleline ${_nvcc_version_out})
190 SET(${COMPILER_INFO} "${CUDA_NVCC_EXECUTABLE} ${_nvcc_info_singleline}")
191 string(TOUPPER ${CMAKE_BUILD_TYPE} _build_type)
192 SET(_compiler_flags "${CUDA_NVCC_FLAGS_${_build_type}}")
193 if(CUDA_PROPAGATE_HOST_FLAGS)
194 string(REGEX REPLACE "[ ]+" ";" _cxx_flags_nospace "${BUILD_CXXFLAGS}")
196 SET(${COMPILER_FLAGS} "${CUDA_NVCC_FLAGS}${CUDA_NVCC_FLAGS_${_build_type}}; ${_cxx_flags_nospace}")
198 SET(${COMPILER_INFO} "N/A")
199 SET(${COMPILER_FLAGS} "N/A")
204 include(CMakeDependentOption)
205 include(gmxOptionUtilities)
208 if(NOT CUDA_NVCC_EXECUTABLE)
209 message(FATAL_ERROR "nvcc is required for a CUDA build, please set CUDA_TOOLKIT_ROOT_DIR appropriately")
212 # set up nvcc options
213 include(gmxManageNvccConfig)
215 gmx_check_if_changed(_cuda_version_changed CUDA_VERSION)
217 # Generate CUDA RT API version string which will end up in config.h
218 # We do this because nvcc is silly enough to not define its own version
219 # (which should match the CUDA runtime API version AFAICT) and we want to
220 # avoid creating the fragile dependency on cuda_runtime_api.h.
222 # NOTE: CUDA v7.5 is expected to have nvcc define it own version, so in the
223 # future we should switch to using that version string instead of our own.
224 if (NOT GMX_CUDA_VERSION OR _cuda_version_changed)
225 MATH(EXPR GMX_CUDA_VERSION "${CUDA_VERSION_MAJOR}*1000 + ${CUDA_VERSION_MINOR}*10")
228 if (_cuda_version_changed)
229 # check the generated CUDA API version against the one present in cuda_runtime_api.h
230 try_compile(_get_cuda_version_compile_res
232 ${CMAKE_SOURCE_DIR}/cmake/TestCUDAVersion.c
233 COMPILE_DEFINITIONS "-DGMX_CUDA_VERSION=${GMX_CUDA_VERSION}"
234 CMAKE_FLAGS "-DINCLUDE_DIRECTORIES=${CUDA_TOOLKIT_INCLUDE}"
235 OUTPUT_VARIABLE _get_cuda_version_compile_out)
237 if (NOT _get_cuda_version_compile_res)
238 if (_get_cuda_version_compile_out MATCHES "CUDA version mismatch")
239 message(FATAL_ERROR "The CUDA API version generated internally from the compiler version does not match the version reported by cuda.h. This means either that the CUDA detection picked up mismatching nvcc and the CUDA headers (likely not part of the same toolkit installation) or that there is an error in the internal version generation. If you are sure that it is not the former causing the error (check the relevant cache variables), define the GMX_CUDA_VERSION cache variable to work around the error.")
241 message(FATAL_ERROR "Could not detect CUDA runtime API version")
246 # no OpenMP is no good!
248 message(WARNING "To use GPU acceleration efficiently, mdrun requires OpenMP multi-threading. Without OpenMP a single CPU core can be used with a GPU which is not optimal. Note that with MPI multiple processes can be forced to use a single GPU, but this is typically inefficient. You need to set both C and C++ compilers that support OpenMP (CC and CXX environment variables, respectively) when using GPUs.")
252 cmake_dependent_option(GMX_CUDA_NB_SINGLE_COMPILATION_UNIT
253 "Whether to compile the CUDA non-bonded module using a single compilation unit." ON
255 mark_as_advanced(GMX_CUDA_NB_SINGLE_COMPILATION_UNIT)
258 # We need to use single compilation unit for kernels:
259 # - when compiling for CC 2.x devices where buggy kernel code is generated
260 gmx_check_if_changed(_gmx_cuda_target_changed GMX_CUDA_TARGET_SM GMX_CUDA_TARGET_COMPUTE CUDA_NVCC_FLAGS)
261 if(_gmx_cuda_target_changed OR NOT GMX_GPU_DETECTION_DONE)
262 if((NOT GMX_CUDA_TARGET_SM AND NOT GMX_CUDA_TARGET_COMPUTE) OR
263 (GMX_CUDA_TARGET_SM MATCHES "2[01]" OR GMX_CUDA_TARGET_COMPUTE MATCHES "2[01]"))
264 message(STATUS "Enabling single compilation unit for the CUDA non-bonded module. Multiple compilation units are not compatible with CC 2.x devices, to enable the feature specify only CC >=3.0 target architectures in GMX_CUDA_TARGET_SM/GMX_CUDA_TARGET_COMPUTE.")
265 set_property(CACHE GMX_CUDA_NB_SINGLE_COMPILATION_UNIT PROPERTY VALUE ON)
267 message(STATUS "Enabling multiple compilation units for the CUDA non-bonded module.")
268 set_property(CACHE GMX_CUDA_NB_SINGLE_COMPILATION_UNIT PROPERTY VALUE OFF)