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.
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 option(GMX_CLANG_CUDA "Use clang for CUDA" OFF)
47 # CUDA 7.0 or later required, override req. version
48 set(REQUIRED_CUDA_VERSION 7.0)
51 if(GMX_GPU AND GMX_DOUBLE)
52 message(FATAL_ERROR "GPU acceleration is not available in double precision!")
54 if(GMX_GPU_AUTO AND GMX_DOUBLE)
55 message(WARNING "GPU acceleration is not available in double precision, disabled!")
56 set_property(CACHE GMX_GPU PROPERTY VALUE OFF)
57 set_property(CACHE GMX_GPU_AUTO PROPERTY VALUE OFF)
60 # detect GPUs in the build host machine
61 if ((GMX_GPU OR GMX_GPU_AUTO) AND NOT GMX_GPU_DETECTION_DONE)
66 # We need to call find_package even when we've already done the detection/setup
67 if(GMX_GPU OR GMX_GPU_AUTO)
68 if(NOT GMX_GPU AND NOT GMX_DETECT_GPU_AVAILABLE)
69 # Stay quiet when detection has occured and found no GPU.
70 # Noise is acceptable when there is a GPU or the user required one.
71 set(FIND_CUDA_QUIETLY QUIET)
74 # Cmake tries to use the static cuda runtime by default,
75 # but this leads to unusable GPU builds on OS X.
77 set(CUDA_USE_STATIC_CUDA_RUNTIME OFF CACHE STRING "Use the static version of the CUDA runtime library if available")
80 find_package(CUDA ${REQUIRED_CUDA_VERSION} ${FIND_CUDA_QUIETLY})
83 # Depending on the current vale of GMX_GPU and GMX_GPU_AUTO:
84 # - OFF, FALSE: Will skip this detection/setup.
85 # - OFF, TRUE : Will keep GMX_GPU=OFF if no CUDA is detected, but will assemble
86 # a warning message which will be issued at the end of the
87 # configuration if GPU(s) were found in the build system.
88 # - ON , FALSE: The user requested GPU build and this requires CUDA, so we will
89 # fail if it is not available.
90 # - ON , TRUE : Can't happen (GMX_GPU=ON can only be user-set at this point)
91 if((GMX_GPU OR GMX_GPU_AUTO) AND NOT GMX_GPU_DETECTION_DONE)
92 if (EXISTS ${CUDA_TOOLKIT_ROOT_DIR})
93 set(CUDA_FOUND TRUE CACHE INTERNAL "Whether the CUDA toolkit was found" FORCE)
95 set(CUDA_FOUND FALSE CACHE INTERNAL "Whether the CUDA toolkit was found" FORCE)
98 # assemble warning/error message
99 if (GMX_DETECT_GPU_AVAILABLE)
100 set(_msg "${GMX_DETECT_GPU_COUNT} NVIDIA GPU(s) found in the system")
103 if (NOT GMX_DETECT_GPU_INFO STREQUAL "")
105 foreach(gpu ${GMX_DETECT_GPU_INFO})
111 # TODO remove the second part of the message when we'll have compute
112 # capability information from the detection.
114 Compute capability information not available, consult the NVIDIA website:
115 https://developer.nvidia.com/cuda-gpus")
118 set(CUDA_NOTFOUND_MESSAGE "mdrun supports native GPU acceleration on NVIDIA hardware with compute capability >= ${REQUIRED_CUDA_COMPUTE_CAPABILITY} (Fermi or later). This requires the NVIDIA CUDA toolkit, which was not found. Its location can be hinted by setting the CUDA_TOOLKIT_ROOT_DIR CMake option (does not work as an environment variable). The typical location would be /usr/local/cuda[-version]. Note that CPU or GPU acceleration can be selected at runtime.
125 # Disable GPU acceleration in auto mode
126 message(STATUS "No compatible CUDA toolkit found (v5.0+), disabling native GPU acceleration")
127 set_property(CACHE GMX_GPU PROPERTY VALUE OFF)
128 set(CUDA_NOTFOUND_AUTO ON)
130 # the user requested CUDA, but it wasn't found
131 message(FATAL_ERROR "${CUDA_NOTFOUND_MESSAGE}")
135 message(STATUS "Enabling native GPU acceleration")
136 set_property(CACHE GMX_GPU PROPERTY VALUE ON)
138 endif() # NOT CUDA_FOUND
141 # Try to find NVML if a GPU accelerated binary should be build.
143 if (DEFINED NVML_LIBRARY)
144 set(NVML_FIND_QUIETLY TRUE)
147 option(GMX_USE_NVML "Use NVML support for better CUDA performance" ${NVML_FOUND})
148 mark_as_advanced(GMX_USE_NVML)
151 include_directories(SYSTEM ${NVML_INCLUDE_DIR})
153 list(APPEND GMX_EXTRA_LIBRARIES ${NVML_LIBRARY})
155 message(FATAL_ERROR "NVML support was required, but was not detected. Please consult the install guide.")
160 # Annoyingly enough, FindCUDA leaves a few variables behind as non-advanced.
161 # We need to mark these advanced outside the conditional, otherwise, if the
162 # user turns GMX_GPU=OFF after a failed cmake pass, these variables will be
163 # left behind in the cache.
164 mark_as_advanced(CUDA_SDK_ROOT_DIR
165 CUDA_USE_STATIC_CUDA_RUNTIME
166 CUDA_dl_LIBRARY CUDA_rt_LIBRARY
169 mark_as_advanced(CUDA_TOOLKIT_ROOT_DIR)
170 mark_as_advanced(CUDA_HOST_COMPILER)
173 # Try to execute ${CUDA_NVCC_EXECUTABLE} --version and set the output
174 # (or an error string) in the argument variable.
175 # Note that semicolon is used as separator for nvcc.
178 # COMPILER_INFO - [output variable] string with compiler path, ID and
179 # some compiler-provided information
180 # COMPILER_FLAGS - [output variable] flags for the compiler
182 macro(get_cuda_compiler_info COMPILER_INFO COMPILER_FLAGS)
183 if(NOT GMX_CLANG_CUDA)
184 if(CUDA_NVCC_EXECUTABLE)
186 # Get the nvcc version string. This is multi-line, but since it is only 4 lines
187 # and might change in the future it is better to store than trying to parse out
188 # the version from the current format.
189 execute_process(COMMAND ${CUDA_NVCC_EXECUTABLE} --version
190 RESULT_VARIABLE _nvcc_version_res
191 OUTPUT_VARIABLE _nvcc_version_out
192 ERROR_VARIABLE _nvcc_version_err
193 OUTPUT_STRIP_TRAILING_WHITESPACE)
194 if (${_nvcc_version_res} EQUAL 0)
195 # Fix multi-line mess: Replace newline with ";" so we can use it in a define
196 string(REPLACE "\n" ";" _nvcc_info_singleline ${_nvcc_version_out})
197 SET(${COMPILER_INFO} "${CUDA_NVCC_EXECUTABLE} ${_nvcc_info_singleline}")
198 string(TOUPPER ${CMAKE_BUILD_TYPE} _build_type)
199 SET(_compiler_flags "${CUDA_NVCC_FLAGS_${_build_type}}")
200 if(CUDA_PROPAGATE_HOST_FLAGS)
201 string(REGEX REPLACE "[ ]+" ";" _cxx_flags_nospace "${BUILD_CXXFLAGS}")
203 SET(${COMPILER_FLAGS} "${CUDA_NVCC_FLAGS}${CUDA_NVCC_FLAGS_${_build_type}}; ${_cxx_flags_nospace}")
205 SET(${COMPILER_INFO} "N/A")
206 SET(${COMPILER_FLAGS} "N/A")
210 # CXX compiler is the CUDA compiler
211 set(${COMPILER_INFO} "${CMAKE_CXX_COMPILER} ${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION}")
212 # there are some extra flags
213 set(${COMPILER_FLAGS} "${CMAKE_CXX_FLAGS} ${CMAKE_CXX_FLAGS_${_build_type}} ${GMX_CUDA_CLANG_FLAGS}")
217 macro(enable_multiple_cuda_compilation_units)
218 message(STATUS "Enabling multiple compilation units for the CUDA non-bonded module.")
219 set_property(CACHE GMX_CUDA_NB_SINGLE_COMPILATION_UNIT PROPERTY VALUE OFF)
222 include(CMakeDependentOption)
223 include(gmxOptionUtilities)
226 if(NOT GMX_CLANG_CUDA)
227 if(NOT CUDA_NVCC_EXECUTABLE)
228 message(FATAL_ERROR "nvcc is required for a CUDA build, please set CUDA_TOOLKIT_ROOT_DIR appropriately")
230 # set up nvcc options
231 include(gmxManageNvccConfig)
233 include(gmxManageClangCudaConfig)
236 gmx_check_if_changed(_cuda_version_changed CUDA_VERSION)
238 # Generate CUDA RT API version string which will end up in config.h
239 # We do this because nvcc is silly enough to not define its own version
240 # (which should match the CUDA runtime API version AFAICT) and we want to
241 # avoid creating the fragile dependency on cuda_runtime_api.h.
243 # NOTE: CUDA v7.5 is expected to have nvcc define it own version, so in the
244 # future we should switch to using that version string instead of our own.
245 if (NOT GMX_CUDA_VERSION OR _cuda_version_changed)
246 MATH(EXPR GMX_CUDA_VERSION "${CUDA_VERSION_MAJOR}*1000 + ${CUDA_VERSION_MINOR}*10")
249 if (_cuda_version_changed)
250 # check the generated CUDA API version against the one present in cuda_runtime_api.h
251 try_compile(_get_cuda_version_compile_res
253 ${CMAKE_SOURCE_DIR}/cmake/TestCUDAVersion.c
254 COMPILE_DEFINITIONS "-DGMX_CUDA_VERSION=${GMX_CUDA_VERSION}"
255 CMAKE_FLAGS "-DINCLUDE_DIRECTORIES=${CUDA_TOOLKIT_INCLUDE}"
256 OUTPUT_VARIABLE _get_cuda_version_compile_out)
258 if (NOT _get_cuda_version_compile_res)
259 if (_get_cuda_version_compile_out MATCHES "CUDA version mismatch")
260 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.")
262 message(FATAL_ERROR "Could not detect CUDA runtime API version")
267 # no OpenMP is no good!
269 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.")
274 set (_GMX_CUDA_NB_SINGLE_COMPILATION_UNIT_DEFAULT FALSE)
276 set (_GMX_CUDA_NB_SINGLE_COMPILATION_UNIT_DEFAULT TRUE)
278 cmake_dependent_option(GMX_CUDA_NB_SINGLE_COMPILATION_UNIT
279 "Whether to compile the CUDA non-bonded module using a single compilation unit." ${_GMX_CUDA_NB_SINGLE_COMPILATION_UNIT_DEFAULT}
281 mark_as_advanced(GMX_CUDA_NB_SINGLE_COMPILATION_UNIT)
283 if (GMX_GPU AND NOT GMX_CLANG_CUDA)
284 # We need to use single compilation unit for kernels:
285 # when compiling with nvcc for CC 2.x devices where buggy kernel code is generated
286 gmx_check_if_changed(_gmx_cuda_target_changed GMX_CUDA_TARGET_SM GMX_CUDA_TARGET_COMPUTE CUDA_NVCC_FLAGS)
288 if(_gmx_cuda_target_changed OR NOT GMX_GPU_DETECTION_DONE)
289 if((NOT GMX_CUDA_TARGET_SM AND NOT GMX_CUDA_TARGET_COMPUTE) OR
290 (GMX_CUDA_TARGET_SM MATCHES "2[01]" OR GMX_CUDA_TARGET_COMPUTE MATCHES "2[01]"))
291 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.")
292 set_property(CACHE GMX_CUDA_NB_SINGLE_COMPILATION_UNIT PROPERTY VALUE ON)
294 message(STATUS "Enabling multiple compilation units for the CUDA non-bonded module.")
295 set_property(CACHE GMX_CUDA_NB_SINGLE_COMPILATION_UNIT PROPERTY VALUE OFF)