1 # If the user did not set GMX_GPU we'll consider this option to be
2 # in "auto" mode meaning that we will:
3 # - search for CUDA and set GMX_GPU=ON we find it
4 # - check whether GPUs are present
5 # - if CUDA is not found but GPUs were detected issue a warning
6 if (NOT DEFINED GMX_GPU)
7 set(GMX_GPU_AUTO TRUE CACHE INTERNAL "GPU acceleration will be selected automatically")
9 option(GMX_GPU "Enable GPU acceleration" OFF)
11 if(GMX_GPU AND GMX_DOUBLE)
12 message(FATAL_ERROR "GPU acceleration is not available in double precision!")
14 if(GMX_GPU_AUTO AND GMX_DOUBLE)
15 message(WARNING "GPU acceleration is not available in double precision, disabled!")
16 set_property(CACHE GMX_GPU PROPERTY VALUE OFF)
17 set_property(CACHE GMX_GPU_AUTO PROPERTY VALUE OFF)
20 # detect GPUs in the build host machine
21 if (GMX_GPU OR GMX_GPU_AUTO AND NOT GMX_GPU_DETECTION_DONE)
26 # Depending on the current vale of GMX_GPU and GMX_GPU_AUTO:
27 # - OFF, FALSE: Will skip this detection.
28 # - OFF, TRUE : Will keep GMX_GPU=OFF if no CUDA is detected, but will assemble
29 # a warning message which will be issued at the end of the
30 # configuration if GPU(s) were found in the build system.
31 # - ON , FALSE: The user requested GPU builds, will require CUDA and will fail
32 # if it is not available.
33 # - ON , TRUE : Can't happen (GMX_GPU=ON can only be user-set at this point)
34 if(GMX_GPU OR GMX_GPU_AUTO AND NOT GMX_GPU_DETECTION_DONE)
35 # We support CUDA >=v3.2 on *nix, but <= v4.1 doesn't work with MSVC
37 find_package(CUDA 4.1)
39 find_package(CUDA 3.2)
42 if (EXISTS ${CUDA_TOOLKIT_ROOT_DIR})
43 set(CUDA_FOUND TRUE CACHE INTERNAL "Whether the CUDA toolkit was found" FORCE)
45 set(CUDA_FOUND FALSE CACHE INTERNAL "Whether the CUDA toolkit was found" FORCE)
48 # assemble warning/error message
49 if (GMX_DETECT_GPU_AVAILABLE)
51 ${GMX_DETECT_GPU_COUNT} NVIDIA GPU(s) found in the system")
54 if (NOT GMX_DETECT_GPU_INFO STREQUAL "")
56 foreach(gpu ${GMX_DETECT_GPU_INFO})
62 # TODO remove the second part of the message when we'll have compute
63 # capability information from the detection.
65 Compute capability information not available, consult the NVIDIA website:
66 https://developer.nvidia.com/cuda-gpus
70 set(CUDA_NOTFOUND_MESSAGE "
71 mdrun supports native GPU acceleration on NVIDIA hardware with compute
72 capability >=2.0. This requires the NVIDIA CUDA library, which was not
73 found; the location can be hinted by setting CUDA_TOOLKIT_ROOT_DIR as
74 a CMake option (It does not work as an environment variable).
75 The typical location would be /usr/local/cuda[-version].
76 Note that CPU or GPU acceleration can be selected at runtime!
83 # Disable GPU acceleration in auto mode
84 message(STATUS "Disabling native GPU acceleration")
85 set_property(CACHE GMX_GPU PROPERTY VALUE OFF)
86 set(CUDA_NOTFOUND_AUTO ON)
88 # the user requested CUDA, but it wasn't found
89 message(FATAL_ERROR "${CUDA_NOTFOUND_MESSAGE}")
93 message(STATUS "Enabling native GPU acceleration")
94 set_property(CACHE GMX_GPU PROPERTY VALUE ON)
96 endif() # NOT CUDA_FOUND
98 # Annoyingly enough, FindCUDA leaves a few variables behind as non-advanced.
99 # We need to mark these advanced outside the conditional, otherwise, if the
100 # user turns GMX_GPU=OFF after a failed cmake pass, these variables will be
101 # left behind in the cache.
102 mark_as_advanced(CUDA_BUILD_CUBIN CUDA_BUILD_EMULATION CUDA_SDK_ROOT_DIR CUDA_VERBOSE_BUILD)
105 # set up nvcc options
106 include(gmxManageNvccConfig)
108 # Version info (semicolon used as line separator) for nvcc.
109 get_nvcc_version_info()
111 # Check whether we can use atomic operations needed for polling wait for GPU
112 # (to avoid the cudaStreamSynchronize + ECC bug).
113 # With thread-MPI testing atomics has already been carried out, but without
114 # thread-MPI we need to invoke the atomics test independently.
115 if (NOT GMX_THREAD_MPI)
116 set(TEST_TMPI_ATOMICS_ONLY ON CACHE INTERNAL
117 "Test only the atomic operations of thread-MPI.")
121 # no OpenMP is no good!
124 To use GPU acceleration efficiently, mdrun requires OpenMP multi-threading.
125 With no OpenMP a single CPU core can be used with a GPU which is not optimal.
126 Note that with MPI multiple processes can be forced to use a single GPU, but this
127 typically inefficient. Note that you need to set both C and C++ compilers that
128 support OpenMP (CC and CXX environment variables, respectively) when using GPUs.")