Imported Upstream version 3.25.0
[platform/upstream/cmake.git] / Modules / CMakeDetermineCUDACompiler.cmake
1 # Distributed under the OSI-approved BSD 3-Clause License.  See accompanying
2 # file Copyright.txt or https://cmake.org/licensing for details.
3
4 include(${CMAKE_ROOT}/Modules/CMakeDetermineCompiler.cmake)
5 include(${CMAKE_ROOT}/Modules/CMakeParseImplicitLinkInfo.cmake)
6
7 if( NOT ( ("${CMAKE_GENERATOR}" MATCHES "Make") OR
8           ("${CMAKE_GENERATOR}" MATCHES "Ninja") OR
9           ("${CMAKE_GENERATOR}" MATCHES "Visual Studio (1|[9][0-9])") ) )
10   message(FATAL_ERROR "CUDA language not currently supported by \"${CMAKE_GENERATOR}\" generator")
11 endif()
12
13 if(${CMAKE_GENERATOR} MATCHES "Visual Studio")
14   if(DEFINED ENV{CUDAHOSTCXX} OR DEFINED CMAKE_CUDA_HOST_COMPILER)
15     message(WARNING "Visual Studio does not support specifying CUDAHOSTCXX or CMAKE_CUDA_HOST_COMPILER. Using the C++ compiler provided by Visual Studio.")
16   endif()
17 else()
18   if(NOT CMAKE_CUDA_COMPILER)
19     set(CMAKE_CUDA_COMPILER_INIT NOTFOUND)
20
21     # prefer the environment variable CUDACXX
22     if(NOT $ENV{CUDACXX} STREQUAL "")
23       get_filename_component(CMAKE_CUDA_COMPILER_INIT $ENV{CUDACXX} PROGRAM PROGRAM_ARGS CMAKE_CUDA_FLAGS_ENV_INIT)
24       if(CMAKE_CUDA_FLAGS_ENV_INIT)
25         set(CMAKE_CUDA_COMPILER_ARG1 "${CMAKE_CUDA_FLAGS_ENV_INIT}" CACHE STRING "Arguments to CUDA compiler")
26       endif()
27       if(NOT EXISTS ${CMAKE_CUDA_COMPILER_INIT})
28         message(FATAL_ERROR "Could not find compiler set in environment variable CUDACXX:\n$ENV{CUDACXX}.\n${CMAKE_CUDA_COMPILER_INIT}")
29       endif()
30     endif()
31
32     # finally list compilers to try
33     if(NOT CMAKE_CUDA_COMPILER_INIT)
34       set(CMAKE_CUDA_COMPILER_LIST nvcc)
35     endif()
36
37     set(_CMAKE_CUDA_COMPILER_PATHS "$ENV{CUDA_PATH}/bin")
38     _cmake_find_compiler(CUDA)
39     unset(_CMAKE_CUDA_COMPILER_PATHS)
40   else()
41     _cmake_find_compiler_path(CUDA)
42   endif()
43
44   mark_as_advanced(CMAKE_CUDA_COMPILER)
45
46   #Allow the user to specify a host compiler except for Visual Studio
47   if(NOT $ENV{CUDAHOSTCXX} STREQUAL "")
48     get_filename_component(CMAKE_CUDA_HOST_COMPILER $ENV{CUDAHOSTCXX} PROGRAM)
49     if(NOT EXISTS ${CMAKE_CUDA_HOST_COMPILER})
50       message(FATAL_ERROR "Could not find compiler set in environment variable CUDAHOSTCXX:\n$ENV{CUDAHOSTCXX}.\n${CMAKE_CUDA_HOST_COMPILER}")
51     endif()
52   endif()
53 endif()
54
55 if(NOT "$ENV{CUDAARCHS}" STREQUAL "")
56   set(CMAKE_CUDA_ARCHITECTURES "$ENV{CUDAARCHS}" CACHE STRING "CUDA architectures")
57 endif()
58
59 # Build a small source file to identify the compiler.
60 if(NOT CMAKE_CUDA_COMPILER_ID_RUN)
61   set(CMAKE_CUDA_COMPILER_ID_RUN 1)
62
63   include(${CMAKE_ROOT}/Modules/CMakeDetermineCompilerId.cmake)
64
65   if(${CMAKE_GENERATOR} MATCHES "Visual Studio")
66     # We will not know CMAKE_CUDA_COMPILER until the main compiler id step
67     # below extracts it, but we do know that the compiler id will be NVIDIA.
68     set(CMAKE_CUDA_COMPILER_ID "NVIDIA")
69   else()
70     # We determine the vendor to help with find the toolkit and use the right flags for detection right away.
71     # The main compiler identification is still needed below to extract other information.
72     list(APPEND CMAKE_CUDA_COMPILER_ID_VENDORS NVIDIA Clang)
73     set(CMAKE_CUDA_COMPILER_ID_VENDOR_REGEX_NVIDIA "nvcc: NVIDIA \\(R\\) Cuda compiler driver")
74     set(CMAKE_CUDA_COMPILER_ID_VENDOR_REGEX_Clang "(clang version)")
75     CMAKE_DETERMINE_COMPILER_ID_VENDOR(CUDA "--version")
76
77     if(CMAKE_CUDA_COMPILER_ID STREQUAL "Clang" AND WIN32)
78       message(FATAL_ERROR "Clang with CUDA is not yet supported on Windows. See CMake issue #20776.")
79     endif()
80
81     # Find the CUDA toolkit. We store the CMAKE_CUDA_COMPILER_TOOLKIT_ROOT, CMAKE_CUDA_COMPILER_TOOLKIT_VERSION and
82     # CMAKE_CUDA_COMPILER_LIBRARY_ROOT in CMakeCUDACompiler.cmake so FindCUDAToolkit can avoid searching on future
83     # runs and the toolkit is the same.
84     # This is very similar to FindCUDAToolkit, but somewhat simplified since we can issue fatal errors
85     # if we fail and we don't need to account for searching the libraries.
86
87     # For NVCC we can easily deduce the SDK binary directory from the compiler path.
88     if(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
89       set(_CUDA_NVCC_EXECUTABLE "${CMAKE_CUDA_COMPILER}")
90     else()
91       # Search using CUDAToolkit_ROOT and then CUDA_PATH for equivalence with FindCUDAToolkit.
92       # In FindCUDAToolkit CUDAToolkit_ROOT is searched automatically due to being in a find_package().
93       # First we search candidate non-default paths to give them priority.
94       find_program(_CUDA_NVCC_EXECUTABLE
95         NAMES nvcc nvcc.exe
96         PATHS ${CUDAToolkit_ROOT}
97         ENV CUDAToolkit_ROOT
98         ENV CUDA_PATH
99         PATH_SUFFIXES bin
100         NO_DEFAULT_PATH
101       )
102
103       # If we didn't find NVCC, then try the default paths.
104       find_program(_CUDA_NVCC_EXECUTABLE
105         NAMES nvcc nvcc.exe
106         PATH_SUFFIXES bin
107       )
108
109       # If the user specified CUDAToolkit_ROOT but nvcc could not be found, this is an error.
110       if(NOT _CUDA_NVCC_EXECUTABLE AND (DEFINED CUDAToolkit_ROOT OR DEFINED ENV{CUDAToolkit_ROOT}))
111         set(fail_base "Could not find nvcc executable in path specified by")
112
113         if(DEFINED CUDAToolkit_ROOT)
114           message(FATAL_ERROR "${fail_base} CUDAToolkit_ROOT=${CUDAToolkit_ROOT}")
115         elseif(DEFINED ENV{CUDAToolkit_ROOT})
116           message(FATAL_ERROR "${fail_base} environment variable CUDAToolkit_ROOT=$ENV{CUDAToolkit_ROOT}")
117         endif()
118       endif()
119
120       # CUDAToolkit_ROOT cmake/env variable not specified, try platform defaults.
121       #
122       # - Linux: /usr/local/cuda-X.Y
123       # - macOS: /Developer/NVIDIA/CUDA-X.Y
124       # - Windows: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\vX.Y
125       #
126       # We will also search the default symlink location /usr/local/cuda first since
127       # if CUDAToolkit_ROOT is not specified, it is assumed that the symlinked
128       # directory is the desired location.
129       if(NOT _CUDA_NVCC_EXECUTABLE)
130         if(UNIX)
131           if(NOT APPLE)
132             set(platform_base "/usr/local/cuda-")
133           else()
134             set(platform_base "/Developer/NVIDIA/CUDA-")
135           endif()
136         else()
137           set(platform_base "C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v")
138         endif()
139
140         # Build out a descending list of possible cuda installations, e.g.
141         file(GLOB possible_paths "${platform_base}*")
142         # Iterate the glob results and create a descending list.
143         set(versions)
144         foreach(p ${possible_paths})
145           # Extract version number from end of string
146           string(REGEX MATCH "[0-9][0-9]?\\.[0-9]$" p_version ${p})
147           if(IS_DIRECTORY ${p} AND p_version)
148             list(APPEND versions ${p_version})
149           endif()
150         endforeach()
151
152         # Sort numerically in descending order, so we try the newest versions first.
153         list(SORT versions COMPARE NATURAL ORDER DESCENDING)
154
155         # With a descending list of versions, populate possible paths to search.
156         set(search_paths)
157         foreach(v ${versions})
158           list(APPEND search_paths "${platform_base}${v}")
159         endforeach()
160
161         # Force the global default /usr/local/cuda to the front on Unix.
162         if(UNIX)
163           list(INSERT search_paths 0 "/usr/local/cuda")
164         endif()
165
166         # Now search for nvcc again using the platform default search paths.
167         find_program(_CUDA_NVCC_EXECUTABLE
168           NAMES nvcc nvcc.exe
169           PATHS ${search_paths}
170           PATH_SUFFIXES bin
171         )
172
173         # We are done with these variables now, cleanup.
174         unset(platform_base)
175         unset(possible_paths)
176         unset(versions)
177         unset(search_paths)
178
179         if(NOT _CUDA_NVCC_EXECUTABLE)
180           message(FATAL_ERROR "Failed to find nvcc.\nCompiler ${CMAKE_CUDA_COMPILER_ID} requires the CUDA toolkit. Please set the CUDAToolkit_ROOT variable.")
181         endif()
182       endif()
183     endif()
184
185     # Given that NVCC can be provided by multiple different sources (NVIDIA HPC SDK, CUDA Toolkit, distro)
186     # each of which has a different layout, we need to extract the CUDA toolkit root from the compiler
187     # itself, allowing us to support numerous different scattered toolkit layouts
188     execute_process(COMMAND ${_CUDA_NVCC_EXECUTABLE} "-v" "__cmake_determine_cuda"
189       OUTPUT_VARIABLE _CUDA_NVCC_OUT ERROR_VARIABLE _CUDA_NVCC_OUT)
190     if(_CUDA_NVCC_OUT MATCHES "\\#\\$ TOP=([^\r\n]*)")
191       get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${CMAKE_MATCH_1}" ABSOLUTE)
192     else()
193       get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${_CUDA_NVCC_EXECUTABLE}" DIRECTORY)
194       get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}" DIRECTORY)
195     endif()
196
197     if(_CUDA_NVCC_OUT MATCHES "\\#\\$ NVVMIR_LIBRARY_DIR=([^\r\n]*)")
198       get_filename_component(_CUDA_NVVMIR_LIBRARY_DIR "${CMAKE_MATCH_1}" ABSOLUTE)
199
200       #We require the path to end in `/nvvm/libdevice'
201       if(_CUDA_NVVMIR_LIBRARY_DIR MATCHES "nvvm/libdevice$")
202         get_filename_component(_CUDA_NVVMIR_LIBRARY_DIR "${_CUDA_NVVMIR_LIBRARY_DIR}/../.." ABSOLUTE)
203         set(CMAKE_CUDA_COMPILER_LIBRARY_ROOT_FROM_NVVMIR_LIBRARY_DIR "${_CUDA_NVVMIR_LIBRARY_DIR}")
204       endif()
205
206       unset(_CUDA_NVVMIR_LIBRARY_DIR)
207       unset(_cuda_nvvmir_dir_name)
208     endif()
209     unset(_CUDA_NVCC_OUT)
210
211     set(CMAKE_CUDA_DEVICE_LINKER "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}/bin/nvlink${CMAKE_EXECUTABLE_SUFFIX}")
212     set(CMAKE_CUDA_FATBINARY "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}/bin/fatbinary${CMAKE_EXECUTABLE_SUFFIX}")
213
214     # In a non-scattered installation the following are equivalent to CMAKE_CUDA_COMPILER_TOOLKIT_ROOT.
215     # We first check for a non-scattered installation to prefer it over a scattered installation.
216
217     # CMAKE_CUDA_COMPILER_LIBRARY_ROOT contains the device library.
218     if(DEFINED CMAKE_CUDA_COMPILER_LIBRARY_ROOT_FROM_NVVMIR_LIBRARY_DIR)
219       set(CMAKE_CUDA_COMPILER_LIBRARY_ROOT "${CMAKE_CUDA_COMPILER_LIBRARY_ROOT_FROM_NVVMIR_LIBRARY_DIR}")
220     elseif(EXISTS "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}/nvvm/libdevice")
221       set(CMAKE_CUDA_COMPILER_LIBRARY_ROOT "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}")
222     elseif(CMAKE_SYSROOT_LINK AND EXISTS "${CMAKE_SYSROOT_LINK}/usr/lib/cuda/nvvm/libdevice")
223       set(CMAKE_CUDA_COMPILER_LIBRARY_ROOT "${CMAKE_SYSROOT_LINK}/usr/lib/cuda")
224     elseif(EXISTS "${CMAKE_SYSROOT}/usr/lib/cuda/nvvm/libdevice")
225       set(CMAKE_CUDA_COMPILER_LIBRARY_ROOT "${CMAKE_SYSROOT}/usr/lib/cuda")
226     else()
227       message(FATAL_ERROR "Couldn't find CUDA library root.")
228     endif()
229     unset(CMAKE_CUDA_COMPILER_LIBRARY_ROOT_FROM_NVVMIR_LIBRARY_DIR)
230
231     # CMAKE_CUDA_COMPILER_TOOLKIT_LIBRARY_ROOT contains the linking stubs necessary for device linking and other low-level library files.
232     if(CMAKE_SYSROOT_LINK AND EXISTS "${CMAKE_SYSROOT_LINK}/usr/lib/nvidia-cuda-toolkit/bin/crt/link.stub")
233       set(CMAKE_CUDA_COMPILER_TOOLKIT_LIBRARY_ROOT "${CMAKE_SYSROOT_LINK}/usr/lib/nvidia-cuda-toolkit")
234     elseif(EXISTS "${CMAKE_SYSROOT}/usr/lib/nvidia-cuda-toolkit/bin/crt/link.stub")
235       set(CMAKE_CUDA_COMPILER_TOOLKIT_LIBRARY_ROOT "${CMAKE_SYSROOT}/usr/lib/nvidia-cuda-toolkit")
236     else()
237       set(CMAKE_CUDA_COMPILER_TOOLKIT_LIBRARY_ROOT "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}")
238     endif()
239   endif()
240
241   # For regular nvcc we the toolkit version is the same as the compiler version and we can parse it from the vendor test output.
242   # For Clang we need to invoke nvcc to get version output.
243   if(NOT CMAKE_GENERATOR MATCHES "Visual Studio")
244     if(CMAKE_CUDA_COMPILER_ID STREQUAL "Clang")
245       execute_process(COMMAND ${_CUDA_NVCC_EXECUTABLE} "--version" OUTPUT_VARIABLE CMAKE_CUDA_COMPILER_ID_OUTPUT)
246     endif()
247
248     if(CMAKE_CUDA_COMPILER_ID_OUTPUT MATCHES [=[V([0-9]+\.[0-9]+\.[0-9]+)]=])
249       set(CMAKE_CUDA_COMPILER_TOOLKIT_VERSION "${CMAKE_MATCH_1}")
250     endif()
251   endif()
252
253   set(CMAKE_CUDA_COMPILER_ID_FLAGS_ALWAYS "-v")
254
255   if(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
256     set(nvcc_test_flags "--keep --keep-dir tmp")
257     if(CMAKE_CUDA_HOST_COMPILER)
258       string(APPEND nvcc_test_flags " -ccbin=\"${CMAKE_CUDA_HOST_COMPILER}\"")
259     endif()
260     # If we have extracted the vendor as NVIDIA we should require detection to
261     # work. If we don't, users will get confusing errors later about failure
262     # to detect a default value for CMAKE_CUDA_ARCHITECTURES
263     set(CMAKE_CUDA_COMPILER_ID_REQUIRE_SUCCESS ON)
264   elseif(CMAKE_CUDA_COMPILER_ID STREQUAL "Clang")
265     set(clang_test_flags "--cuda-path=\"${CMAKE_CUDA_COMPILER_LIBRARY_ROOT}\"")
266     if(CMAKE_CROSSCOMPILING)
267       # Need to pass the host target and include directories if we're crosscompiling.
268       string(APPEND clang_test_flags " --sysroot=\"${CMAKE_SYSROOT}\" --target=${CMAKE_CUDA_COMPILER_TARGET}")
269     endif()
270   endif()
271
272   # Rest of the code treats an empty value as equivalent to "use the defaults".
273   # Error out early to prevent confusing errors as a result of this.
274   # Note that this also catches invalid non-numerical values such as "a".
275   if(DEFINED CMAKE_CUDA_ARCHITECTURES)
276     if(CMAKE_CUDA_ARCHITECTURES STREQUAL "")
277       message(FATAL_ERROR "CMAKE_CUDA_ARCHITECTURES must be non-empty if set.")
278     elseif(CMAKE_CUDA_ARCHITECTURES AND NOT CMAKE_CUDA_ARCHITECTURES MATCHES "^([0-9]+a?(-real|-virtual)?(;[0-9]+a?(-real|-virtual)?|;)*|all|all-major|native)$")
279       message(FATAL_ERROR
280         "CMAKE_CUDA_ARCHITECTURES:\n"
281         "  ${CMAKE_CUDA_ARCHITECTURES}\n"
282         "is not one of the following:\n"
283         "  * a semicolon-separated list of integers, each optionally\n"
284         "    followed by '-real' or '-virtual'\n"
285         "  * a special value: all, all-major, native\n"
286         )
287     endif()
288   endif()
289
290   if(CMAKE_CUDA_COMPILER_ID STREQUAL "Clang")
291     # Clang doesn't automatically select an architecture supported by the SDK.
292     # Try in reverse order of deprecation with the most recent at front (i.e. the most likely to work for new setups).
293     foreach(arch "52" "30" "20")
294       list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${clang_test_flags} --cuda-gpu-arch=sm_${arch}")
295     endforeach()
296   elseif(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
297     list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${nvcc_test_flags}")
298   endif()
299
300   # We perform compiler identification for a second time to extract implicit linking info and host compiler for NVCC.
301   # We need to unset the compiler ID otherwise CMAKE_DETERMINE_COMPILER_ID() doesn't work.
302   set(CMAKE_CUDA_COMPILER_ID)
303   set(CMAKE_CUDA_PLATFORM_ID)
304   file(READ ${CMAKE_ROOT}/Modules/CMakePlatformId.h.in
305     CMAKE_CUDA_COMPILER_ID_PLATFORM_CONTENT)
306
307   CMAKE_DETERMINE_COMPILER_ID(CUDA CUDAFLAGS CMakeCUDACompilerId.cu)
308
309   if(${CMAKE_GENERATOR} MATCHES "Visual Studio")
310     # Now that we have the path to nvcc, we can compute the toolkit root.
311     get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${CMAKE_CUDA_COMPILER}" DIRECTORY)
312     get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}" DIRECTORY)
313     set(CMAKE_CUDA_COMPILER_LIBRARY_ROOT "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}")
314
315     # The compiler comes with the toolkit, so the versions are the same.
316     set(CMAKE_CUDA_COMPILER_TOOLKIT_VERSION ${CMAKE_CUDA_COMPILER_VERSION})
317   endif()
318
319   include(${CMAKE_ROOT}/Modules/CUDA/architectures.cmake)
320
321   _cmake_find_compiler_sysroot(CUDA)
322 endif()
323
324 set(_CMAKE_PROCESSING_LANGUAGE "CUDA")
325 include(CMakeFindBinUtils)
326 include(Compiler/${CMAKE_CUDA_COMPILER_ID}-FindBinUtils OPTIONAL)
327 unset(_CMAKE_PROCESSING_LANGUAGE)
328
329 if(MSVC_CUDA_ARCHITECTURE_ID)
330   set(SET_MSVC_CUDA_ARCHITECTURE_ID
331     "set(MSVC_CUDA_ARCHITECTURE_ID ${MSVC_CUDA_ARCHITECTURE_ID})")
332 endif()
333
334 if(${CMAKE_GENERATOR} MATCHES "Visual Studio")
335   set(CMAKE_CUDA_HOST_LINK_LAUNCHER "${CMAKE_LINKER}")
336   set(CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES "")
337   set(CMAKE_CUDA_HOST_IMPLICIT_LINK_DIRECTORIES "")
338   set(CMAKE_CUDA_HOST_IMPLICIT_LINK_FRAMEWORK_DIRECTORIES "")
339
340   # We do not currently detect CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES but we
341   # do need to detect CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT from the compiler by
342   # looking at which cudart library exists in the implicit link libraries passed
343   # to the host linker.
344   if(CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT MATCHES "link\\.exe [^\n]*cudart_static\\.lib")
345     set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT "STATIC")
346   elseif(CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT MATCHES "link\\.exe [^\n]*cudart\\.lib")
347     set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT "SHARED")
348   else()
349     set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT "NONE")
350   endif()
351   set(_SET_CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT
352     "set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT \"${CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT}\")")
353 elseif(CMAKE_CUDA_COMPILER_ID STREQUAL "Clang")
354   string(REGEX MATCHALL "-target-cpu sm_([0-9]+)" target_cpus "${CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT}")
355
356   foreach(cpu ${target_cpus})
357     string(REGEX MATCH "-target-cpu sm_([0-9]+)" dont_care "${cpu}")
358     list(APPEND architectures_detected "${CMAKE_MATCH_1}")
359   endforeach()
360
361   # Find target directory when crosscompiling.
362   if(CMAKE_CROSSCOMPILING)
363     if(CMAKE_SYSTEM_PROCESSOR STREQUAL "armv7-a")
364       # Support for NVPACK
365       set(_CUDA_TARGET_NAME "armv7-linux-androideabi")
366     elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "arm")
367       set(_CUDA_TARGET_NAME "armv7-linux-gnueabihf")
368     elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64")
369       if(ANDROID_ARCH_NAME STREQUAL "arm64")
370         set(_CUDA_TARGET_NAME "aarch64-linux-androideabi")
371       else()
372         set(_CUDA_TARGET_NAME "aarch64-linux")
373       endif()
374     elseif(CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
375       set(_CUDA_TARGET_NAME "x86_64-linux")
376     endif()
377
378     if(EXISTS "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}/targets/${_CUDA_TARGET_NAME}")
379       set(_CUDA_TARGET_DIR "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}/targets/${_CUDA_TARGET_NAME}")
380     endif()
381   endif()
382
383   # If not already set we can simply use the toolkit root or it's a scattered installation.
384   if(NOT _CUDA_TARGET_DIR)
385     set(_CUDA_TARGET_DIR "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}")
386   endif()
387
388   # We can't use find_library() yet at this point, so try a few guesses.
389   if(EXISTS "${_CUDA_TARGET_DIR}/lib64")
390     set(_CUDA_LIBRARY_DIR "${_CUDA_TARGET_DIR}/lib64")
391   elseif(EXISTS "${_CUDA_TARGET_DIR}/lib/x64")
392     set(_CUDA_LIBRARY_DIR "${_CUDA_TARGET_DIR}/lib/x64")
393   elseif(EXISTS "${_CUDA_TARGET_DIR}/lib")
394     set(_CUDA_LIBRARY_DIR "${_CUDA_TARGET_DIR}/lib")
395   else()
396     message(FATAL_ERROR "Unable to find _CUDA_LIBRARY_DIR based on _CUDA_TARGET_DIR=${_CUDA_TARGET_DIR}")
397   endif()
398
399   # _CUDA_TARGET_DIR always points to the directory containing the include directory.
400   # On a scattered installation /usr, on a non-scattered something like /usr/local/cuda or /usr/local/cuda-10.2/targets/aarch64-linux.
401   if(EXISTS "${_CUDA_TARGET_DIR}/include/cuda_runtime.h")
402     set(_CUDA_INCLUDE_DIR "${_CUDA_TARGET_DIR}/include")
403   else()
404     message(FATAL_ERROR "Unable to find cuda_runtime.h in \"${_CUDA_TARGET_DIR}/include\" for _CUDA_INCLUDE_DIR.")
405   endif()
406
407   # Clang does not add any CUDA SDK libraries or directories when invoking the host linker.
408   # Add the CUDA toolkit library directory ourselves so that linking works.
409   # The CUDA runtime libraries are handled elsewhere by CMAKE_CUDA_RUNTIME_LIBRARY.
410   set(CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES "${_CUDA_INCLUDE_DIR}")
411   set(CMAKE_CUDA_HOST_IMPLICIT_LINK_DIRECTORIES "${_CUDA_LIBRARY_DIR}")
412   set(CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES "")
413   set(CMAKE_CUDA_HOST_IMPLICIT_LINK_FRAMEWORK_DIRECTORIES "")
414 elseif(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
415   set(_nvcc_log "")
416   string(REPLACE "\r" "" _nvcc_output_orig "${CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT}")
417   if(_nvcc_output_orig MATCHES "#\\\$ +PATH= *([^\n]*)\n")
418     set(_nvcc_path "${CMAKE_MATCH_1}")
419     string(APPEND _nvcc_log "  found 'PATH=' string: [${_nvcc_path}]\n")
420     string(REPLACE ":" ";" _nvcc_path "${_nvcc_path}")
421   else()
422     set(_nvcc_path "")
423     string(REPLACE "\n" "\n    " _nvcc_output_log "\n${_nvcc_output_orig}")
424     string(APPEND _nvcc_log "  no 'PATH=' string found in nvcc output:${_nvcc_output_log}\n")
425   endif()
426   if(_nvcc_output_orig MATCHES "#\\\$ +LIBRARIES= *([^\n]*)\n")
427     set(_nvcc_libraries "${CMAKE_MATCH_1}")
428     string(APPEND _nvcc_log "  found 'LIBRARIES=' string: [${_nvcc_libraries}]\n")
429   else()
430     set(_nvcc_libraries "")
431     string(REPLACE "\n" "\n    " _nvcc_output_log "\n${_nvcc_output_orig}")
432     string(APPEND _nvcc_log "  no 'LIBRARIES=' string found in nvcc output:${_nvcc_output_log}\n")
433   endif()
434
435   set(_nvcc_link_line "")
436   if(_nvcc_libraries)
437     # Remove variable assignments.
438     string(REGEX REPLACE "#\\\$ *[^= ]+=[^\n]*\n" "" _nvcc_output "${_nvcc_output_orig}")
439     # Encode [] characters that break list expansion.
440     string(REPLACE "[" "{==={" _nvcc_output "${_nvcc_output}")
441     string(REPLACE "]" "}===}" _nvcc_output "${_nvcc_output}")
442     # Split lines.
443     string(REGEX REPLACE "\n+(#\\\$ )?" ";" _nvcc_output "${_nvcc_output}")
444     foreach(line IN LISTS _nvcc_output)
445       set(_nvcc_output_line "${line}")
446       string(REPLACE "{==={" "[" _nvcc_output_line "${_nvcc_output_line}")
447       string(REPLACE "}===}" "]" _nvcc_output_line "${_nvcc_output_line}")
448       string(APPEND _nvcc_log "  considering line: [${_nvcc_output_line}]\n")
449       if("${_nvcc_output_line}" MATCHES "^ *nvlink")
450         string(APPEND _nvcc_log "    ignoring nvlink line\n")
451       elseif(_nvcc_libraries)
452         if("${_nvcc_output_line}" MATCHES "(@\"?((tmp/)?a\\.exe\\.res)\"?)")
453           set(_nvcc_link_res_arg "${CMAKE_MATCH_1}")
454           set(_nvcc_link_res_file "${CMAKE_MATCH_2}")
455           set(_nvcc_link_res "${CMAKE_PLATFORM_INFO_DIR}/CompilerIdCUDA/${_nvcc_link_res_file}")
456           if(EXISTS "${_nvcc_link_res}")
457             file(READ "${_nvcc_link_res}" _nvcc_link_res_content)
458             string(REPLACE "${_nvcc_link_res_arg}" "${_nvcc_link_res_content}" _nvcc_output_line "${_nvcc_output_line}")
459           endif()
460         endif()
461         string(FIND "${_nvcc_output_line}" "${_nvcc_libraries}" _nvcc_libraries_pos)
462         if(NOT _nvcc_libraries_pos EQUAL -1)
463           set(_nvcc_link_line "${_nvcc_output_line}")
464           string(APPEND _nvcc_log "    extracted link line: [${_nvcc_link_line}]\n")
465         endif()
466       endif()
467     endforeach()
468   endif()
469
470   if(_nvcc_link_line)
471     if("x${CMAKE_CUDA_SIMULATE_ID}" STREQUAL "xMSVC")
472       set(CMAKE_CUDA_HOST_LINK_LAUNCHER "${CMAKE_LINKER}")
473     else()
474       #extract the compiler that is being used for linking
475       separate_arguments(_nvcc_link_line_args UNIX_COMMAND "${_nvcc_link_line}")
476       list(GET _nvcc_link_line_args 0 _nvcc_host_link_launcher)
477       if(IS_ABSOLUTE "${_nvcc_host_link_launcher}")
478         string(APPEND _nvcc_log "  extracted link launcher absolute path: [${_nvcc_host_link_launcher}]\n")
479         set(CMAKE_CUDA_HOST_LINK_LAUNCHER "${_nvcc_host_link_launcher}")
480       else()
481         string(APPEND _nvcc_log "  extracted link launcher name: [${_nvcc_host_link_launcher}]\n")
482         find_program(_nvcc_find_host_link_launcher
483           NAMES ${_nvcc_host_link_launcher}
484           PATHS ${_nvcc_path} NO_DEFAULT_PATH)
485         find_program(_nvcc_find_host_link_launcher
486           NAMES ${_nvcc_host_link_launcher})
487         if(_nvcc_find_host_link_launcher)
488           string(APPEND _nvcc_log "  found link launcher absolute path: [${_nvcc_find_host_link_launcher}]\n")
489           set(CMAKE_CUDA_HOST_LINK_LAUNCHER "${_nvcc_find_host_link_launcher}")
490         else()
491           string(APPEND _nvcc_log "  could not find link launcher absolute path\n")
492           set(CMAKE_CUDA_HOST_LINK_LAUNCHER "${_nvcc_host_link_launcher}")
493         endif()
494         unset(_nvcc_find_host_link_launcher CACHE)
495       endif()
496     endif()
497
498     #prefix the line with cuda-fake-ld so that implicit link info believes it is
499     #a link line
500     set(_nvcc_link_line "cuda-fake-ld ${_nvcc_link_line}")
501     CMAKE_PARSE_IMPLICIT_LINK_INFO("${_nvcc_link_line}"
502                                    CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES
503                                    CMAKE_CUDA_HOST_IMPLICIT_LINK_DIRECTORIES
504                                    CMAKE_CUDA_HOST_IMPLICIT_LINK_FRAMEWORK_DIRECTORIES
505                                    log
506                                    "${CMAKE_CUDA_IMPLICIT_OBJECT_REGEX}"
507                                    LANGUAGE CUDA)
508
509     # Detect CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT from the compiler by looking at which
510     # cudart library exists in the implicit link libraries passed to the host linker.
511     # This is required when a project sets the cuda runtime library as part of the
512     # initial flags.
513     if(";${CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES};" MATCHES [[;cudart_static(\.lib)?;]])
514       set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT "STATIC")
515     elseif(";${CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES};" MATCHES [[;cudart(\.lib)?;]])
516       set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT "SHARED")
517     else()
518       set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT "NONE")
519     endif()
520     set(_SET_CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT
521       "set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT \"${CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT}\")")
522
523     file(APPEND ${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeOutput.log
524       "Parsed CUDA nvcc implicit link information from above output:\n${_nvcc_log}\n${log}\n\n")
525   else()
526     file(APPEND ${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeError.log
527       "Failed to parse CUDA nvcc implicit link information:\n${_nvcc_log}\n\n")
528     message(FATAL_ERROR "Failed to extract nvcc implicit link line.")
529   endif()
530 endif()
531
532 # CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES is detected above as the list of
533 # libraries that the CUDA compiler implicitly passes to the host linker.
534 # CMake invokes the host linker directly and so needs to pass these libraries.
535 # We filter out those that should not be passed unconditionally both here
536 # and from CMAKE_CUDA_IMPLICIT_LINK_LIBRARIES in CMakeTestCUDACompiler.
537 set(CMAKE_CUDA_IMPLICIT_LINK_LIBRARIES_EXCLUDE
538   # The CUDA runtime libraries are controlled by CMAKE_CUDA_RUNTIME_LIBRARY.
539   cudart        cudart.lib
540   cudart_static cudart_static.lib
541   cudadevrt     cudadevrt.lib
542
543   # Dependencies of the CUDA static runtime library on Linux hosts.
544   rt
545   pthread
546   dl
547   )
548 list(REMOVE_ITEM CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES ${CMAKE_CUDA_IMPLICIT_LINK_LIBRARIES_EXCLUDE})
549
550 if(CMAKE_CUDA_COMPILER_SYSROOT)
551   string(CONCAT _SET_CMAKE_CUDA_COMPILER_SYSROOT
552     "set(CMAKE_CUDA_COMPILER_SYSROOT \"${CMAKE_CUDA_COMPILER_SYSROOT}\")\n"
553     "set(CMAKE_COMPILER_SYSROOT \"${CMAKE_CUDA_COMPILER_SYSROOT}\")")
554 else()
555   set(_SET_CMAKE_CUDA_COMPILER_SYSROOT "")
556 endif()
557
558 # Determine CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES
559 if(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
560   set(CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES)
561   string(REPLACE "\r" "" _nvcc_output_orig "${CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT}")
562   if(_nvcc_output_orig MATCHES "#\\\$ +INCLUDES= *([^\n]*)\n")
563     set(_nvcc_includes "${CMAKE_MATCH_1}")
564     string(APPEND _nvcc_log "  found 'INCLUDES=' string: [${_nvcc_includes}]\n")
565   else()
566     set(_nvcc_includes "")
567     string(REPLACE "\n" "\n    " _nvcc_output_log "\n${_nvcc_output_orig}")
568     string(APPEND _nvcc_log "  no 'INCLUDES=' string found in nvcc output:${_nvcc_output_log}\n")
569   endif()
570   if(_nvcc_includes)
571     # across all operating system each include directory is prefixed with -I
572     separate_arguments(_nvcc_output NATIVE_COMMAND "${_nvcc_includes}")
573     foreach(line IN LISTS _nvcc_output)
574       string(REGEX REPLACE "^-I" "" line "${line}")
575       get_filename_component(line "${line}" ABSOLUTE)
576       list(APPEND CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES "${line}")
577     endforeach()
578
579     file(APPEND ${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeOutput.log
580       "Parsed CUDA nvcc include information from above output:\n${_nvcc_log}\n${log}\n\n")
581   else()
582     file(APPEND ${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeOutput.log
583       "Failed to detect CUDA nvcc include information:\n${_nvcc_log}\n\n")
584   endif()
585
586   string(REGEX MATCHALL "-arch compute_([0-9]+)" target_cpus "${CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT}")
587
588   foreach(cpu ${target_cpus})
589     string(REGEX MATCH "-arch compute_([0-9]+)" dont_care "${cpu}")
590     list(APPEND architectures_detected "${CMAKE_MATCH_1}")
591   endforeach()
592 endif()
593
594 # If the user didn't set the architectures, then set them to a default.
595 # If the user did, then make sure those architectures worked.
596 if("${CMAKE_CUDA_ARCHITECTURES}" STREQUAL "")
597   cmake_policy(GET CMP0104 _CUDA_CMP0104)
598
599   if(NOT CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA" OR _CUDA_CMP0104 STREQUAL "NEW")
600     set(CMAKE_CUDA_ARCHITECTURES "${architectures_detected}" CACHE STRING "CUDA architectures")
601
602     if(NOT CMAKE_CUDA_ARCHITECTURES)
603       message(FATAL_ERROR "Failed to detect a default CUDA architecture.\n\nCompiler output:\n${CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT}")
604     endif()
605   endif()
606 endif()
607
608 # configure all variables set in this file
609 configure_file(${CMAKE_ROOT}/Modules/CMakeCUDACompiler.cmake.in
610   ${CMAKE_PLATFORM_INFO_DIR}/CMakeCUDACompiler.cmake
611   @ONLY
612 )
613
614 # Don't leak variables unnecessarily to user code.
615 unset(_CUDA_INCLUDE_DIR CACHE)
616 unset(_CUDA_NVCC_EXECUTABLE CACHE)
617 unset(_CUDA_LIBRARY_DIR)
618 unset(_CUDA_TARGET_DIR)
619 unset(_CUDA_TARGET_NAME)
620
621 unset(architectures_detected)
622
623 set(CMAKE_CUDA_COMPILER_ENV_VAR "CUDACXX")
624 set(CMAKE_CUDA_HOST_COMPILER_ENV_VAR "CUDAHOSTCXX")