2 message(STATUS "CUDA compilation is disabled (due to only Visual Studio compiler supported on your platform).")
6 if(NOT UNIX AND CV_CLANG)
7 message(STATUS "CUDA compilation is disabled (due to Clang unsupported on your platform).")
12 if(((NOT CMAKE_VERSION VERSION_LESS "3.9.0") # requires https://gitlab.kitware.com/cmake/cmake/merge_requests/663
13 OR OPENCV_CUDA_FORCE_EXTERNAL_CMAKE_MODULE)
14 AND NOT OPENCV_CUDA_FORCE_BUILTIN_CMAKE_MODULE)
15 ocv_update(CUDA_LINK_LIBRARIES_KEYWORD "LINK_PRIVATE")
16 find_host_package(CUDA "${MIN_VER_CUDA}" QUIET)
18 # Use OpenCV's patched "FindCUDA" module
19 set(CMAKE_MODULE_PATH "${OpenCV_SOURCE_DIR}/cmake" ${CMAKE_MODULE_PATH})
22 set(CUDA_TARGET_OS_VARIANT "Android")
24 find_host_package(CUDA "${MIN_VER_CUDA}" QUIET)
26 list(REMOVE_AT CMAKE_MODULE_PATH 0)
41 set(CMAKE_MODULE_PATH "${OpenCV_SOURCE_DIR}/cmake" ${CMAKE_MODULE_PATH})
42 find_host_package(CUDNN "${MIN_VER_CUDNN}")
43 list(REMOVE_AT CMAKE_MODULE_PATH 0)
51 find_cuda_helper_libs(nvcuvid)
53 find_cuda_helper_libs(nvcuvenc)
55 if(CUDA_nvcuvid_LIBRARY)
58 if(CUDA_nvcuvenc_LIBRARY)
63 message(STATUS "CUDA detected: " ${CUDA_VERSION})
65 set(_generations "Fermi" "Kepler" "Maxwell" "Pascal" "Volta" "Turing")
66 if(NOT CMAKE_CROSSCOMPILING)
67 list(APPEND _generations "Auto")
69 set(CUDA_GENERATION "" CACHE STRING "Build CUDA device code only for specific GPU architecture. Leave empty to build for all architectures.")
70 if( CMAKE_VERSION VERSION_GREATER "2.8" )
71 set_property( CACHE CUDA_GENERATION PROPERTY STRINGS "" ${_generations} )
75 if(NOT ";${_generations};" MATCHES ";${CUDA_GENERATION};")
76 string(REPLACE ";" ", " _generations "${_generations}")
77 message(FATAL_ERROR "ERROR: ${_generations} Generations are suppered.")
79 unset(CUDA_ARCH_BIN CACHE)
80 unset(CUDA_ARCH_PTX CACHE)
83 SET(DETECT_ARCHS_COMMAND "${CUDA_NVCC_EXECUTABLE}" ${CUDA_NVCC_FLAGS} "${OpenCV_SOURCE_DIR}/cmake/checks/OpenCVDetectCudaArch.cu" "--run")
84 if(WIN32 AND CMAKE_LINKER) #Workaround for VS cl.exe not being in the env. path
85 get_filename_component(host_compiler_bindir ${CMAKE_LINKER} DIRECTORY)
86 SET(DETECT_ARCHS_COMMAND ${DETECT_ARCHS_COMMAND} "-ccbin" "${host_compiler_bindir}")
89 set(__cuda_arch_ptx "")
90 if(CUDA_GENERATION STREQUAL "Fermi")
91 set(__cuda_arch_bin "2.0")
92 elseif(CUDA_GENERATION STREQUAL "Kepler")
93 set(__cuda_arch_bin "3.0 3.5 3.7")
94 elseif(CUDA_GENERATION STREQUAL "Maxwell")
95 set(__cuda_arch_bin "5.0 5.2")
96 elseif(CUDA_GENERATION STREQUAL "Pascal")
97 set(__cuda_arch_bin "6.0 6.1")
98 elseif(CUDA_GENERATION STREQUAL "Volta")
99 set(__cuda_arch_bin "7.0")
100 elseif(CUDA_GENERATION STREQUAL "Turing")
101 set(__cuda_arch_bin "7.5")
102 elseif(CUDA_GENERATION STREQUAL "Auto")
103 execute_process( COMMAND ${DETECT_ARCHS_COMMAND}
104 WORKING_DIRECTORY "${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeTmp/"
105 RESULT_VARIABLE _nvcc_res OUTPUT_VARIABLE _nvcc_out
106 ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
107 string(REGEX REPLACE ".*\n" "" _nvcc_out "${_nvcc_out}") #Strip leading warning messages, if any
108 if(NOT _nvcc_res EQUAL 0)
109 message(STATUS "Automatic detection of CUDA generation failed. Going to build for all known architectures.")
111 set(__cuda_arch_bin "${_nvcc_out}")
112 string(REPLACE "2.1" "2.1(2.0)" __cuda_arch_bin "${__cuda_arch_bin}")
116 if(NOT DEFINED __cuda_arch_bin)
118 set(__cuda_arch_bin "3.2")
119 set(__cuda_arch_ptx "")
121 execute_process( COMMAND ${DETECT_ARCHS_COMMAND}
122 WORKING_DIRECTORY "${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeTmp/"
123 RESULT_VARIABLE _nvcc_res OUTPUT_VARIABLE _nvcc_out
124 ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
125 string(REGEX REPLACE ".*\n" "" _nvcc_out "${_nvcc_out}") #Strip leading warning messages, if any
126 if(NOT _nvcc_res EQUAL 0)
127 message(STATUS "Automatic detection of CUDA generation failed. Going to build for all known architectures.")
128 set(__cuda_arch_bin "5.3 6.2 7.2")
130 set(__cuda_arch_bin "${_nvcc_out}")
131 string(REPLACE "2.1" "2.1(2.0)" __cuda_arch_bin "${__cuda_arch_bin}")
133 set(__cuda_arch_ptx "")
135 if(CUDA_VERSION VERSION_LESS "9.0")
136 set(__cuda_arch_bin "2.0 3.0 3.5 3.7 5.0 5.2 6.0 6.1")
137 elseif(CUDA_VERSION VERSION_LESS "10.0")
138 set(__cuda_arch_bin "3.0 3.5 3.7 5.0 5.2 6.0 6.1 7.0")
140 set(__cuda_arch_bin "3.0 3.5 3.7 5.0 5.2 6.0 6.1 7.0 7.5")
145 set(CUDA_ARCH_BIN ${__cuda_arch_bin} CACHE STRING "Specify 'real' GPU architectures to build binaries for, BIN(PTX) format is supported")
146 set(CUDA_ARCH_PTX ${__cuda_arch_ptx} CACHE STRING "Specify 'virtual' PTX architectures to build PTX intermediate code for")
148 string(REGEX REPLACE "\\." "" ARCH_BIN_NO_POINTS "${CUDA_ARCH_BIN}")
149 string(REGEX REPLACE "\\." "" ARCH_PTX_NO_POINTS "${CUDA_ARCH_PTX}")
151 # Ckeck if user specified 1.0 compute capability: we don't support it
152 string(REGEX MATCH "1.0" HAS_ARCH_10 "${CUDA_ARCH_BIN} ${CUDA_ARCH_PTX}")
153 set(CUDA_ARCH_BIN_OR_PTX_10 0)
154 if(NOT ${HAS_ARCH_10} STREQUAL "")
155 set(CUDA_ARCH_BIN_OR_PTX_10 1)
158 # NVCC flags to be set
159 set(NVCC_FLAGS_EXTRA "")
161 # These vars will be passed into the templates
162 set(OPENCV_CUDA_ARCH_BIN "")
163 set(OPENCV_CUDA_ARCH_PTX "")
164 set(OPENCV_CUDA_ARCH_FEATURES "")
166 # Tell NVCC to add binaries for the specified GPUs
167 string(REGEX MATCHALL "[0-9()]+" ARCH_LIST "${ARCH_BIN_NO_POINTS}")
168 foreach(ARCH IN LISTS ARCH_LIST)
169 if(ARCH MATCHES "([0-9]+)\\(([0-9]+)\\)")
170 # User explicitly specified PTX for the concrete BIN
171 set(NVCC_FLAGS_EXTRA ${NVCC_FLAGS_EXTRA} -gencode arch=compute_${CMAKE_MATCH_2},code=sm_${CMAKE_MATCH_1})
172 set(OPENCV_CUDA_ARCH_BIN "${OPENCV_CUDA_ARCH_BIN} ${CMAKE_MATCH_1}")
173 set(OPENCV_CUDA_ARCH_FEATURES "${OPENCV_CUDA_ARCH_FEATURES} ${CMAKE_MATCH_2}")
175 # User didn't explicitly specify PTX for the concrete BIN, we assume PTX=BIN
176 set(NVCC_FLAGS_EXTRA ${NVCC_FLAGS_EXTRA} -gencode arch=compute_${ARCH},code=sm_${ARCH})
177 set(OPENCV_CUDA_ARCH_BIN "${OPENCV_CUDA_ARCH_BIN} ${ARCH}")
178 set(OPENCV_CUDA_ARCH_FEATURES "${OPENCV_CUDA_ARCH_FEATURES} ${ARCH}")
181 set(NVCC_FLAGS_EXTRA ${NVCC_FLAGS_EXTRA} -D_FORCE_INLINES)
183 # Tell NVCC to add PTX intermediate code for the specified architectures
184 string(REGEX MATCHALL "[0-9]+" ARCH_LIST "${ARCH_PTX_NO_POINTS}")
185 foreach(ARCH IN LISTS ARCH_LIST)
186 set(NVCC_FLAGS_EXTRA ${NVCC_FLAGS_EXTRA} -gencode arch=compute_${ARCH},code=compute_${ARCH})
187 set(OPENCV_CUDA_ARCH_PTX "${OPENCV_CUDA_ARCH_PTX} ${ARCH}")
188 set(OPENCV_CUDA_ARCH_FEATURES "${OPENCV_CUDA_ARCH_FEATURES} ${ARCH}")
191 # These vars will be processed in other scripts
192 set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} ${NVCC_FLAGS_EXTRA})
193 set(OpenCV_CUDA_CC "${NVCC_FLAGS_EXTRA}")
196 set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-Xptxas;-dlcm=ca")
199 message(STATUS "CUDA NVCC target flags: ${CUDA_NVCC_FLAGS}")
201 OCV_OPTION(CUDA_FAST_MATH "Enable --use_fast_math for CUDA compiler " OFF)
204 set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} --use_fast_math)
207 mark_as_advanced(CUDA_BUILD_CUBIN CUDA_BUILD_EMULATION CUDA_VERBOSE_BUILD CUDA_SDK_ROOT_DIR)
209 macro(ocv_cuda_filter_options)
210 foreach(var CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_RELEASE CMAKE_CXX_FLAGS_DEBUG)
211 set(${var}_backup_in_cuda_compile_ "${${var}}")
214 # we remove -Winconsistent-missing-override and -Qunused-arguments
215 # just in case we are compiling CUDA with gcc but OpenCV with clang
216 string(REPLACE "-Winconsistent-missing-override" "" ${var} "${${var}}")
217 string(REPLACE "-Qunused-arguments" "" ${var} "${${var}}")
220 # we remove /EHa as it generates warnings under windows
221 string(REPLACE "/EHa" "" ${var} "${${var}}")
223 # we remove -ggdb3 flag as it leads to preprocessor errors when compiling CUDA files (CUDA 4.1)
224 string(REPLACE "-ggdb3" "" ${var} "${${var}}")
226 # we remove -Wsign-promo as it generates warnings under linux
227 string(REPLACE "-Wsign-promo" "" ${var} "${${var}}")
229 # we remove -Wno-sign-promo as it generates warnings under linux
230 string(REPLACE "-Wno-sign-promo" "" ${var} "${${var}}")
232 # we remove -Wno-delete-non-virtual-dtor because it's used for C++ compiler
233 # but NVCC uses C compiler by default
234 string(REPLACE "-Wno-delete-non-virtual-dtor" "" ${var} "${${var}}")
236 # we remove -frtti because it's used for C++ compiler
237 # but NVCC uses C compiler by default
238 string(REPLACE "-frtti" "" ${var} "${${var}}")
240 string(REPLACE "-fvisibility-inlines-hidden" "" ${var} "${${var}}")
242 # cc1: warning: command line option '-Wsuggest-override' is valid for C++/ObjC++ but not for C
243 string(REPLACE "-Wsuggest-override" "" ${var} "${${var}}")
245 # issue: #11552 (from OpenCVCompilerOptions.cmake)
246 string(REGEX REPLACE "-Wimplicit-fallthrough(=[0-9]+)? " "" ${var} "${${var}}")
248 # removal of custom specified options
249 if(OPENCV_CUDA_NVCC_FILTEROUT_OPTIONS)
250 foreach(__flag ${OPENCV_CUDA_NVCC_FILTEROUT_OPTIONS})
251 string(REPLACE "${__flag}" "" ${var} "${${var}}")
257 macro(ocv_cuda_compile VAR)
258 ocv_cuda_filter_options()
260 if(BUILD_SHARED_LIBS)
261 set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xcompiler -DCVAPI_EXPORTS)
265 set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xcompiler -fPIC --std=c++11)
268 set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xcompiler -fno-finite-math-only)
271 if(CMAKE_CROSSCOMPILING AND (ARM OR AARCH64))
272 set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xlinker --unresolved-symbols=ignore-in-shared-libs)
275 # disabled because of multiple warnings during building nvcc auto generated files
276 if(CV_GCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER "4.6.0")
277 ocv_warnings_disable(CMAKE_CXX_FLAGS -Wunused-but-set-variable)
280 CUDA_COMPILE(${VAR} ${ARGN})
282 foreach(var CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_RELEASE CMAKE_CXX_FLAGS_DEBUG)
283 set(${var} "${${var}_backup_in_cuda_compile_}")
284 unset(${var}_backup_in_cuda_compile_)
288 unset(CUDA_ARCH_BIN CACHE)
289 unset(CUDA_ARCH_PTX CACHE)
293 set(CUDA_LIBS_PATH "")
294 foreach(p ${CUDA_LIBRARIES} ${CUDA_npp_LIBRARY})
295 get_filename_component(_tmp ${p} PATH)
296 list(APPEND CUDA_LIBS_PATH ${_tmp})
300 foreach(p ${CUDA_cublas_LIBRARY})
301 get_filename_component(_tmp ${p} PATH)
302 list(APPEND CUDA_LIBS_PATH ${_tmp})
307 foreach(p ${CUDNN_LIBRARIES})
308 get_filename_component(_tmp ${p} PATH)
309 list(APPEND CUDA_LIBS_PATH ${_tmp})
314 foreach(p ${CUDA_cufft_LIBRARY})
315 get_filename_component(_tmp ${p} PATH)
316 list(APPEND CUDA_LIBS_PATH ${_tmp})
320 list(REMOVE_DUPLICATES CUDA_LIBS_PATH)
321 link_directories(${CUDA_LIBS_PATH})
323 set(CUDA_LIBRARIES_ABS ${CUDA_LIBRARIES})
324 ocv_convert_to_lib_name(CUDA_LIBRARIES ${CUDA_LIBRARIES})
325 set(CUDA_npp_LIBRARY_ABS ${CUDA_npp_LIBRARY})
326 ocv_convert_to_lib_name(CUDA_npp_LIBRARY ${CUDA_npp_LIBRARY})
328 set(CUDA_cublas_LIBRARY_ABS ${CUDA_cublas_LIBRARY})
329 ocv_convert_to_lib_name(CUDA_cublas_LIBRARY ${CUDA_cublas_LIBRARY})
332 set(CUDNN_LIBRARIES_ABS ${CUDNN_LIBRARIES})
333 ocv_convert_to_lib_name(CUDNN_LIBRARIES ${CUDNN_LIBRARIES})
336 set(CUDA_cufft_LIBRARY_ABS ${CUDA_cufft_LIBRARY})
337 ocv_convert_to_lib_name(CUDA_cufft_LIBRARY ${CUDA_cufft_LIBRARY})