summaryrefslogtreecommitdiffstats
path: root/Modules/CMakeDetermineCUDACompiler.cmake
blob: daca382fa297d5f9c25c66fbbfae82738dd386f4 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
# Distributed under the OSI-approved BSD 3-Clause License.  See accompanying
# file Copyright.txt or https://cmake.org/licensing for details.

include(${CMAKE_ROOT}/Modules/CMakeDetermineCompiler.cmake)
include(${CMAKE_ROOT}/Modules/CMakeParseImplicitLinkInfo.cmake)

if( NOT ( ("${CMAKE_GENERATOR}" MATCHES "Make") OR
          ("${CMAKE_GENERATOR}" MATCHES "Ninja") OR
          ("${CMAKE_GENERATOR}" MATCHES "Visual Studio (1|[9][0-9])") ) )
  message(FATAL_ERROR "CUDA language not currently supported by \"${CMAKE_GENERATOR}\" generator")
endif()

if(${CMAKE_GENERATOR} MATCHES "Visual Studio")
else()
  if(NOT CMAKE_CUDA_COMPILER)
    set(CMAKE_CUDA_COMPILER_INIT NOTFOUND)

      # prefer the environment variable CUDACXX
      if(NOT $ENV{CUDACXX} STREQUAL "")
        get_filename_component(CMAKE_CUDA_COMPILER_INIT $ENV{CUDACXX} PROGRAM PROGRAM_ARGS CMAKE_CUDA_FLAGS_ENV_INIT)
        if(CMAKE_CUDA_FLAGS_ENV_INIT)
          set(CMAKE_CUDA_COMPILER_ARG1 "${CMAKE_CUDA_FLAGS_ENV_INIT}" CACHE STRING "Arguments to CXX compiler")
        endif()
        if(NOT EXISTS ${CMAKE_CUDA_COMPILER_INIT})
          message(FATAL_ERROR "Could not find compiler set in environment variable CUDACXX:\n$ENV{CUDACXX}.\n${CMAKE_CUDA_COMPILER_INIT}")
        endif()
      endif()

    # finally list compilers to try
    if(NOT CMAKE_CUDA_COMPILER_INIT)
      set(CMAKE_CUDA_COMPILER_LIST nvcc)
    endif()

    _cmake_find_compiler(CUDA)
  else()
    _cmake_find_compiler_path(CUDA)
  endif()

  mark_as_advanced(CMAKE_CUDA_COMPILER)
endif()

#Allow the user to specify a host compiler
if(NOT $ENV{CUDAHOSTCXX} STREQUAL "")
  get_filename_component(CMAKE_CUDA_HOST_COMPILER $ENV{CUDAHOSTCXX} PROGRAM)
  if(NOT EXISTS ${CMAKE_CUDA_HOST_COMPILER})
    message(FATAL_ERROR "Could not find compiler set in environment variable CUDAHOSTCXX:\n$ENV{CUDAHOSTCXX}.\n${CMAKE_CUDA_HOST_COMPILER}")
  endif()
endif()

# Build a small source file to identify the compiler.
if(NOT CMAKE_CUDA_COMPILER_ID_RUN)
  set(CMAKE_CUDA_COMPILER_ID_RUN 1)

  include(${CMAKE_ROOT}/Modules/CMakeDetermineCompilerId.cmake)

  if(${CMAKE_GENERATOR} MATCHES "Visual Studio")
    # We will not know CMAKE_CUDA_COMPILER until the main compiler id step
    # below extracts it, but we do know that the compiler id will be NVIDIA.
    set(CMAKE_CUDA_COMPILER_ID "NVIDIA")
  else()
    # We determine the vendor to help with find the toolkit and use the right flags for detection right away.
    # The main compiler identification is still needed below to extract other information.
    list(APPEND CMAKE_CUDA_COMPILER_ID_VENDORS NVIDIA Clang)
    set(CMAKE_CUDA_COMPILER_ID_VENDOR_REGEX_NVIDIA "nvcc: NVIDIA \\(R\\) Cuda compiler driver")
    set(CMAKE_CUDA_COMPILER_ID_VENDOR_REGEX_Clang "(clang version)")
    CMAKE_DETERMINE_COMPILER_ID_VENDOR(CUDA "--version")

    # Find the CUDA toolkit. We store the CMAKE_CUDA_COMPILER_TOOLKIT_ROOT and CMAKE_CUDA_COMPILER_LIBRARY_ROOT
    # in CMakeCUDACompiler.cmake, so FindCUDAToolkit can avoid searching on future runs and the toolkit stays the same.
    # This is very similar to FindCUDAToolkit, but somewhat simplified since we can issue fatal errors
    # if we fail to find things we need and we don't need to account for searching the libraries.

    # For NVCC we can easily deduce the SDK binary directory from the compiler path.
    if(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
      set(_CUDA_NVCC_EXECUTABLE "${CMAKE_CUDA_COMPILER}")
    else()
      # Search using CUDAToolkit_ROOT and then CUDA_PATH for equivalence with FindCUDAToolkit.
      # In FindCUDAToolkit CUDAToolkit_ROOT is searched automatically due to being in a find_package().
      # First we search candidate non-default paths to give them priority.
      find_program(_CUDA_NVCC_EXECUTABLE
        NAMES nvcc nvcc.exe
        PATHS ${CUDAToolkit_ROOT}
        ENV CUDAToolkit_ROOT
        ENV CUDA_PATH
        PATH_SUFFIXES bin
        NO_DEFAULT_PATH
      )

      # If we didn't find NVCC, then try the default paths.
      find_program(_CUDA_NVCC_EXECUTABLE
        NAMES nvcc nvcc.exe
        PATH_SUFFIXES bin
      )

      # If the user specified CUDAToolkit_ROOT but nvcc could not be found, this is an error.
      if(NOT _CUDA_NVCC_EXECUTABLE AND (DEFINED CUDAToolkit_ROOT OR DEFINED ENV{CUDAToolkit_ROOT}))
        set(fail_base "Could not find nvcc executable in path specified by")

        if(DEFINED CUDAToolkit_ROOT)
          message(FATAL_ERROR "${fail_base} CUDAToolkit_ROOT=${CUDAToolkit_ROOT}")
        elseif(DEFINED ENV{CUDAToolkit_ROOT})
          message(FATAL_ERROR "${fail_base} environment variable CUDAToolkit_ROOT=$ENV{CUDAToolkit_ROOT}")
        endif()
      endif()

      # CUDAToolkit_ROOT cmake/env variable not specified, try platform defaults.
      #
      # - Linux: /usr/local/cuda-X.Y
      # - macOS: /Developer/NVIDIA/CUDA-X.Y
      # - Windows: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\vX.Y
      #
      # We will also search the default symlink location /usr/local/cuda first since
      # if CUDAToolkit_ROOT is not specified, it is assumed that the symlinked
      # directory is the desired location.
      if(NOT _CUDA_NVCC_EXECUTABLE)
        if(UNIX)
          if(NOT APPLE)
            set(platform_base "/usr/local/cuda-")
          else()
            set(platform_base "/Developer/NVIDIA/CUDA-")
          endif()
        else()
          set(platform_base "C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v")
        endif()

        # Build out a descending list of possible cuda installations, e.g.
        file(GLOB possible_paths "${platform_base}*")
        # Iterate the glob results and create a descending list.
        set(versions)
        foreach(p ${possible_paths})
          # Extract version number from end of string
          string(REGEX MATCH "[0-9][0-9]?\\.[0-9]$" p_version ${p})
          if(IS_DIRECTORY ${p} AND p_version)
            list(APPEND versions ${p_version})
          endif()
        endforeach()

        # Sort numerically in descending order, so we try the newest versions first.
        list(SORT versions COMPARE NATURAL ORDER DESCENDING)

        # With a descending list of versions, populate possible paths to search.
        set(search_paths)
        foreach(v ${versions})
          list(APPEND search_paths "${platform_base}${v}")
        endforeach()

        # Force the global default /usr/local/cuda to the front on Unix.
        if(UNIX)
          list(INSERT search_paths 0 "/usr/local/cuda")
        endif()

        # Now search for nvcc again using the platform default search paths.
        find_program(_CUDA_NVCC_EXECUTABLE
          NAMES nvcc nvcc.exe
          PATHS ${search_paths}
          PATH_SUFFIXES bin
        )

        # We are done with these variables now, cleanup.
        unset(platform_base)
        unset(possible_paths)
        unset(versions)
        unset(search_paths)

        if(NOT _CUDA_NVCC_EXECUTABLE)
          message(FATAL_ERROR "Could not find nvcc, please set CUDAToolkit_ROOT.")
        endif()
      endif()
    endif()

    get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${_CUDA_NVCC_EXECUTABLE}" DIRECTORY)
    get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}" DIRECTORY)

    # CMAKE_CUDA_COMPILER_LIBRARY_ROOT contains the device library and version file.
    # In a non-scattered installation this is equivalent to CMAKE_CUDA_COMPILER_TOOLKIT_ROOT.
    # We first check for a non-scattered installation to prefer it over a scattered installation.
    if(EXISTS "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}/version.txt")
      set(CMAKE_CUDA_COMPILER_LIBRARY_ROOT "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}")
    elseif(CMAKE_SYSROOT_LINK AND EXISTS "${CMAKE_SYSROOT_LINK}/usr/lib/cuda/version.txt")
      set(CMAKE_CUDA_COMPILER_LIBRARY_ROOT "${CMAKE_SYSROOT_LINK}/usr/lib/cuda")
    elseif(EXISTS "${CMAKE_SYSROOT}/usr/lib/cuda/version.txt")
      set(CMAKE_CUDA_COMPILER_LIBRARY_ROOT "${CMAKE_SYSROOT}/usr/lib/cuda")
    endif()
  endif()

  set(CMAKE_CUDA_COMPILER_ID_FLAGS_ALWAYS "-v")

  if(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
    set(nvcc_test_flags "--keep --keep-dir tmp")
    if(CMAKE_CUDA_HOST_COMPILER)
      string(APPEND nvcc_test_flags " -ccbin=\"${CMAKE_CUDA_HOST_COMPILER}\"")
    endif()
  elseif(CMAKE_CUDA_COMPILER_ID STREQUAL "Clang")
    if(WIN32)
      message(FATAL_ERROR "Clang with CUDA is not yet supported on Windows. See CMake issue #20776.")
    endif()

    set(clang_test_flags "--cuda-path=\"${CMAKE_CUDA_COMPILER_LIBRARY_ROOT}\"")
    if(CMAKE_CROSSCOMPILING)
      # Need to pass the host target and include directories if we're crosscompiling.
      string(APPEND clang_test_flags " --sysroot=\"${CMAKE_SYSROOT}\" --target=${CMAKE_CUDA_COMPILER_TARGET}")
    endif()
  endif()

  # First try with the user-specified architectures.
  if(CMAKE_CUDA_ARCHITECTURES)
    set(clang_archs "${clang_test_flags}")
    set(nvcc_archs "${nvcc_test_flags}")

    foreach(arch ${CMAKE_CUDA_ARCHITECTURES})
      # Strip specifiers as PTX vs binary doesn't matter.
      string(REGEX MATCH "[0-9]+" arch_name "${arch}")
      string(APPEND clang_archs " --cuda-gpu-arch=sm_${arch_name}")
      string(APPEND nvcc_archs " -gencode=arch=compute_${arch_name},code=sm_${arch_name}")
      list(APPEND tested_architectures "${arch_name}")
    endforeach()

    list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${clang_archs}")
    list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${nvcc_archs}")
  endif()

  # Fallback default NVCC flags.
  list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST ${nvcc_test_flags})

  # Clang doesn't automatically select an architecture supported by the SDK.
  # Try in reverse order of deprecation with the most recent at front (i.e. the most likely to work for new setups).
  foreach(arch "20" "30" "52")
    list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${clang_test_flags} --cuda-gpu-arch=sm_${arch}")
  endforeach()

  # Finally also try the default.
  list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${clang_test_flags}")

  # We perform compiler identification for a second time to extract implicit linking info and host compiler for NVCC.
  # We also use it to verify that CMAKE_CUDA_ARCHITECTURES and additionally on Clang that CUDA toolkit path works.
  # The latter could be done during compiler testing in the future to avoid doing this for Clang.
  # We need to unset the compiler ID otherwise CMAKE_DETERMINE_COMPILER_ID() doesn't work.
  set(CMAKE_CUDA_COMPILER_ID)
  set(CMAKE_CUDA_PLATFORM_ID)
  file(READ ${CMAKE_ROOT}/Modules/CMakePlatformId.h.in
    CMAKE_CUDA_COMPILER_ID_PLATFORM_CONTENT)

  CMAKE_DETERMINE_COMPILER_ID(CUDA CUDAFLAGS CMakeCUDACompilerId.cu)

  if(${CMAKE_GENERATOR} MATCHES "Visual Studio")
    # Now that we have the path to nvcc, we can compute the toolkit root.
    get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${CMAKE_CUDA_COMPILER}" DIRECTORY)
    get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}" DIRECTORY)
    set(CMAKE_CUDA_COMPILER_LIBRARY_ROOT "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}")
  endif()

  _cmake_find_compiler_sysroot(CUDA)
endif()

set(_CMAKE_PROCESSING_LANGUAGE "CUDA")
include(CMakeFindBinUtils)
include(Compiler/${CMAKE_CUDA_COMPILER_ID}-FindBinUtils OPTIONAL)
unset(_CMAKE_PROCESSING_LANGUAGE)

if(MSVC_CUDA_ARCHITECTURE_ID)
  set(SET_MSVC_CUDA_ARCHITECTURE_ID
    "set(MSVC_CUDA_ARCHITECTURE_ID ${MSVC_CUDA_ARCHITECTURE_ID})")
endif()

if(${CMAKE_GENERATOR} MATCHES "Visual Studio")
  set(CMAKE_CUDA_HOST_LINK_LAUNCHER "${CMAKE_LINKER}")
  set(CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES "")
  set(CMAKE_CUDA_HOST_IMPLICIT_LINK_DIRECTORIES "")
  set(CMAKE_CUDA_HOST_IMPLICIT_LINK_FRAMEWORK_DIRECTORIES "")

  # We do not currently detect CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES but we
  # do need to detect CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT from the compiler by
  # looking at which cudart library exists in the implicit link libraries passed
  # to the host linker.
  if(CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT MATCHES "link\\.exe [^\n]*cudart_static\\.lib")
    set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT "STATIC")
  elseif(CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT MATCHES "link\\.exe [^\n]*cudart\\.lib")
    set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT "SHARED")
  else()
    set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT "NONE")
  endif()
  set(_SET_CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT
    "set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT \"${CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT}\")")
elseif(CMAKE_CUDA_COMPILER_ID STREQUAL "Clang")
  if(NOT CMAKE_CUDA_ARCHITECTURES)
    # Find the architecture that we successfully compiled using and set it as the default.
    string(REGEX MATCH "-target-cpu sm_([0-9]+)" dont_care "${CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT}")
    set(detected_architecture "${CMAKE_MATCH_1}")
  else()
    string(REGEX MATCHALL "-target-cpu sm_([0-9]+)" target_cpus "${CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT}")

    foreach(cpu ${target_cpus})
      string(REGEX MATCH "-target-cpu sm_([0-9]+)" dont_care "${cpu}")
      list(APPEND architectures "${CMAKE_MATCH_1}")
    endforeach()
  endif()

  # Find target directory. Account for crosscompiling.
  if(CMAKE_CROSSCOMPILING)
    if(CMAKE_SYSTEM_PROCESSOR STREQUAL "armv7-a")
      # Support for NVPACK
      set(_CUDA_TARGET_NAME "armv7-linux-androideabi")
    elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "arm")
      set(_CUDA_TARGET_NAME "armv7-linux-gnueabihf")
    elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64")
      if(ANDROID_ARCH_NAME STREQUAL "arm64")
        set(_CUDA_TARGET_NAME "aarch64-linux-androideabi")
      else()
        set(_CUDA_TARGET_NAME "aarch64-linux")
      endif()
    elseif(CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
      set(_CUDA_TARGET_NAME "x86_64-linux")
    endif()

    if(EXISTS "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}/targets/${_CUDA_TARGET_NAME}")
      set(_CUDA_TARGET_DIR "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}/targets/${_CUDA_TARGET_NAME}")
    endif()
  else()
    set(_CUDA_TARGET_DIR "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}")
  endif()

  # We can't use find_library() yet at this point, so try a few guesses.
  if(EXISTS "${_CUDA_TARGET_DIR}/lib64")
    set(_CUDA_LIBRARY_DIR "${_CUDA_TARGET_DIR}/lib64")
  elseif(EXISTS "${_CUDA_TARGET_DIR}/lib/x64")
    set(_CUDA_LIBRARY_DIR "${_CUDA_TARGET_DIR}/lib/x64")
  elseif(EXISTS "${_CUDA_TARGET_DIR}/lib")
    set(_CUDA_LIBRARY_DIR "${_CUDA_TARGET_DIR}/lib")
  else()
    message(FATAL_ERROR "Unable to find _CUDA_LIBRARY_DIR based on _CUDA_TARGET_DIR=${_CUDA_TARGET_DIR}")
  endif()

  # _CUDA_TARGET_DIR always points to the directory containing the include directory.
  # On a scattered installation /usr, on a non-scattered something like /usr/local/cuda or /usr/local/cuda-10.2/targets/aarch64-linux.
  if(EXISTS "${_CUDA_TARGET_DIR}/include/cuda_runtime.h")
    set(_CUDA_INCLUDE_DIR "${_CUDA_TARGET_DIR}/include")
  else()
    message(FATAL_ERROR "Unable to find cuda_runtime.h in \"${_CUDA_TARGET_DIR}/include\" for _CUDA_INCLUDE_DIR.")
  endif()

  # Clang does not add any CUDA SDK libraries or directories when invoking the host linker.
  # Add the CUDA toolkit library directory ourselves so that linking works.
  # The CUDA runtime libraries are handled elsewhere by CMAKE_CUDA_RUNTIME_LIBRARY.
  set(CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES "${_CUDA_INCLUDE_DIR}")
  set(CMAKE_CUDA_HOST_IMPLICIT_LINK_DIRECTORIES "${_CUDA_LIBRARY_DIR}")
  set(CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES "")
  set(CMAKE_CUDA_HOST_IMPLICIT_LINK_FRAMEWORK_DIRECTORIES "")
elseif(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
  set(_nvcc_log "")
  string(REPLACE "\r" "" _nvcc_output_orig "${CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT}")
  if(_nvcc_output_orig MATCHES "#\\\$ +PATH= *([^\n]*)\n")
    set(_nvcc_path "${CMAKE_MATCH_1}")
    string(APPEND _nvcc_log "  found 'PATH=' string: [${_nvcc_path}]\n")
    string(REPLACE ":" ";" _nvcc_path "${_nvcc_path}")
  else()
    set(_nvcc_path "")
    string(REPLACE "\n" "\n    " _nvcc_output_log "\n${_nvcc_output_orig}")
    string(APPEND _nvcc_log "  no 'PATH=' string found in nvcc output:${_nvcc_output_log}\n")
  endif()
  if(_nvcc_output_orig MATCHES "#\\\$ +LIBRARIES= *([^\n]*)\n")
    set(_nvcc_libraries "${CMAKE_MATCH_1}")
    string(APPEND _nvcc_log "  found 'LIBRARIES=' string: [${_nvcc_libraries}]\n")
  else()
    set(_nvcc_libraries "")
    string(REPLACE "\n" "\n    " _nvcc_output_log "\n${_nvcc_output_orig}")
    string(APPEND _nvcc_log "  no 'LIBRARIES=' string found in nvcc output:${_nvcc_output_log}\n")
  endif()

  set(_nvcc_link_line "")
  if(_nvcc_libraries)
    # Remove variable assignments.
    string(REGEX REPLACE "#\\\$ *[^= ]+=[^\n]*\n" "" _nvcc_output "${_nvcc_output_orig}")
    # Encode [] characters that break list expansion.
    string(REPLACE "[" "{==={" _nvcc_output "${_nvcc_output}")
    string(REPLACE "]" "}===}" _nvcc_output "${_nvcc_output}")
    # Split lines.
    string(REGEX REPLACE "\n+(#\\\$ )?" ";" _nvcc_output "${_nvcc_output}")
    foreach(line IN LISTS _nvcc_output)
      set(_nvcc_output_line "${line}")
      string(REPLACE "{==={" "[" _nvcc_output_line "${_nvcc_output_line}")
      string(REPLACE "}===}" "]" _nvcc_output_line "${_nvcc_output_line}")
      string(APPEND _nvcc_log "  considering line: [${_nvcc_output_line}]\n")
      if("${_nvcc_output_line}" MATCHES "^ *nvlink")
        string(APPEND _nvcc_log "    ignoring nvlink line\n")
      elseif(_nvcc_libraries)
        if("${_nvcc_output_line}" MATCHES "(@\"?tmp/a\\.exe\\.res\"?)")
          set(_nvcc_link_res_arg "${CMAKE_MATCH_1}")
          set(_nvcc_link_res "${CMAKE_PLATFORM_INFO_DIR}/CompilerIdCUDA/tmp/a.exe.res")
          if(EXISTS "${_nvcc_link_res}")
            file(READ "${_nvcc_link_res}" _nvcc_link_res_content)
            string(REPLACE "${_nvcc_link_res_arg}" "${_nvcc_link_res_content}" _nvcc_output_line "${_nvcc_output_line}")
          endif()
        endif()
        string(FIND "${_nvcc_output_line}" "${_nvcc_libraries}" _nvcc_libraries_pos)
        if(NOT _nvcc_libraries_pos EQUAL -1)
          set(_nvcc_link_line "${_nvcc_output_line}")
          string(APPEND _nvcc_log "    extracted link line: [${_nvcc_link_line}]\n")
        endif()
      endif()
    endforeach()
  endif()

  if(_nvcc_link_line)
    if("x${CMAKE_CUDA_SIMULATE_ID}" STREQUAL "xMSVC")
      set(CMAKE_CUDA_HOST_LINK_LAUNCHER "${CMAKE_LINKER}")
    else()
      #extract the compiler that is being used for linking
      separate_arguments(_nvcc_link_line_args UNIX_COMMAND "${_nvcc_link_line}")
      list(GET _nvcc_link_line_args 0 _nvcc_host_link_launcher)
      if(IS_ABSOLUTE "${_nvcc_host_link_launcher}")
        string(APPEND _nvcc_log "  extracted link launcher absolute path: [${_nvcc_host_link_launcher}]\n")
        set(CMAKE_CUDA_HOST_LINK_LAUNCHER "${_nvcc_host_link_launcher}")
      else()
        string(APPEND _nvcc_log "  extracted link launcher name: [${_nvcc_host_link_launcher}]\n")
        find_program(_nvcc_find_host_link_launcher
          NAMES ${_nvcc_host_link_launcher}
          PATHS ${_nvcc_path} NO_DEFAULT_PATH)
        find_program(_nvcc_find_host_link_launcher
          NAMES ${_nvcc_host_link_launcher})
        if(_nvcc_find_host_link_launcher)
          string(APPEND _nvcc_log "  found link launcher absolute path: [${_nvcc_find_host_link_launcher}]\n")
          set(CMAKE_CUDA_HOST_LINK_LAUNCHER "${_nvcc_find_host_link_launcher}")
        else()
          string(APPEND _nvcc_log "  could not find link launcher absolute path\n")
          set(CMAKE_CUDA_HOST_LINK_LAUNCHER "${_nvcc_host_link_launcher}")
        endif()
        unset(_nvcc_find_host_link_launcher CACHE)
      endif()
    endif()

    #prefix the line with cuda-fake-ld so that implicit link info believes it is
    #a link line
    set(_nvcc_link_line "cuda-fake-ld ${_nvcc_link_line}")
    CMAKE_PARSE_IMPLICIT_LINK_INFO("${_nvcc_link_line}"
                                   CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES
                                   CMAKE_CUDA_HOST_IMPLICIT_LINK_DIRECTORIES
                                   CMAKE_CUDA_HOST_IMPLICIT_LINK_FRAMEWORK_DIRECTORIES
                                   log
                                   "${CMAKE_CUDA_IMPLICIT_OBJECT_REGEX}")

    # Detect CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT from the compiler by looking at which
    # cudart library exists in the implicit link libraries passed to the host linker.
    # This is required when a project sets the cuda runtime library as part of the
    # initial flags.
    if(";${CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES};" MATCHES [[;cudart_static(\.lib)?;]])
      set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT "STATIC")
    elseif(";${CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES};" MATCHES [[;cudart(\.lib)?;]])
      set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT "SHARED")
    else()
      set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT "NONE")
    endif()
    set(_SET_CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT
      "set(CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT \"${CMAKE_CUDA_RUNTIME_LIBRARY_DEFAULT}\")")

    file(APPEND ${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeOutput.log
      "Parsed CUDA nvcc implicit link information from above output:\n${_nvcc_log}\n${log}\n\n")
  else()
    file(APPEND ${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeError.log
      "Failed to parse CUDA nvcc implicit link information:\n${_nvcc_log}\n\n")
    message(FATAL_ERROR "Failed to extract nvcc implicit link line.")
  endif()
endif()

# CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES is detected above as the list of
# libraries that the CUDA compiler implicitly passes to the host linker.
# CMake invokes the host linker directly and so needs to pass these libraries.
# We filter out those that should not be passed unconditionally both here
# and from CMAKE_CUDA_IMPLICIT_LINK_LIBRARIES in CMakeTestCUDACompiler.
set(CMAKE_CUDA_IMPLICIT_LINK_LIBRARIES_EXCLUDE
  # The CUDA runtime libraries are controlled by CMAKE_CUDA_RUNTIME_LIBRARY.
  cudart        cudart.lib
  cudart_static cudart_static.lib
  cudadevrt     cudadevrt.lib

  # Dependencies of the CUDA static runtime library on Linux hosts.
  rt
  pthread
  dl
  )
list(REMOVE_ITEM CMAKE_CUDA_HOST_IMPLICIT_LINK_LIBRARIES ${CMAKE_CUDA_IMPLICIT_LINK_LIBRARIES_EXCLUDE})

if(CMAKE_CUDA_COMPILER_SYSROOT)
  string(CONCAT _SET_CMAKE_CUDA_COMPILER_SYSROOT
    "set(CMAKE_CUDA_COMPILER_SYSROOT \"${CMAKE_CUDA_COMPILER_SYSROOT}\")\n"
    "set(CMAKE_COMPILER_SYSROOT \"${CMAKE_CUDA_COMPILER_SYSROOT}\")")
else()
  set(_SET_CMAKE_CUDA_COMPILER_SYSROOT "")
endif()

# Determine CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES
if(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
  set(CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES)
  string(REPLACE "\r" "" _nvcc_output_orig "${CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT}")
  if(_nvcc_output_orig MATCHES "#\\\$ +INCLUDES= *([^\n]*)\n")
    set(_nvcc_includes "${CMAKE_MATCH_1}")
    string(APPEND _nvcc_log "  found 'INCLUDES=' string: [${_nvcc_includes}]\n")
  else()
    set(_nvcc_includes "")
    string(REPLACE "\n" "\n    " _nvcc_output_log "\n${_nvcc_output_orig}")
    string(APPEND _nvcc_log "  no 'INCLUDES=' string found in nvcc output:${_nvcc_output_log}\n")
  endif()
  if(_nvcc_includes)
    # across all operating system each include directory is prefixed with -I
    separate_arguments(_nvcc_output NATIVE_COMMAND "${_nvcc_includes}")
    foreach(line IN LISTS _nvcc_output)
      string(REGEX REPLACE "^-I" "" line "${line}")
      get_filename_component(line "${line}" ABSOLUTE)
      list(APPEND CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES "${line}")
    endforeach()

    file(APPEND ${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeOutput.log
      "Parsed CUDA nvcc include information from above output:\n${_nvcc_log}\n${log}\n\n")
  else()
    file(APPEND ${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeOutput.log
      "Failed to detect CUDA nvcc include information:\n${_nvcc_log}\n\n")
  endif()

  # Parse default CUDA architecture.
  cmake_policy(GET CMP0104 _CUDA_CMP0104)
  if(NOT CMAKE_CUDA_ARCHITECTURES AND _CUDA_CMP0104 STREQUAL "NEW")
    string(REGEX MATCH "arch[ =]compute_([0-9]+)" dont_care "${CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT}")
    set(detected_architecture "${CMAKE_MATCH_1}")
  elseif(CMAKE_CUDA_ARCHITECTURES)
    string(REGEX MATCHALL "-arch compute_([0-9]+)" target_cpus "${CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT}")

    foreach(cpu ${target_cpus})
      string(REGEX MATCH "-arch compute_([0-9]+)" dont_care "${cpu}")
      list(APPEND architectures "${CMAKE_MATCH_1}")
    endforeach()
  endif()
endif()

# If the user didn't set the architectures, then set them to a default.
# If the user did, then make sure those architectures worked.
if(DEFINED detected_architecture AND "${CMAKE_CUDA_ARCHITECTURES}" STREQUAL "")
  set(CMAKE_CUDA_ARCHITECTURES "${detected_architecture}" CACHE STRING "CUDA architectures")

  if(NOT CMAKE_CUDA_ARCHITECTURES)
    message(FATAL_ERROR "Failed to find a working CUDA architecture.")
  endif()
elseif(architectures)
  # Sort since order mustn't matter.
  list(SORT architectures)
  list(SORT tested_architectures)

  # We don't distinguish real/virtual architectures during testing.
  # For "70-real;70-virtual" we detect "70" as working and tested_architectures is "70;70".
  # Thus we need to remove duplicates before checking if they're equal.
  list(REMOVE_DUPLICATES tested_architectures)

  if(NOT "${architectures}" STREQUAL "${tested_architectures}")
    message(FATAL_ERROR
      "The CMAKE_CUDA_ARCHITECTURES:\n"
      "  ${CMAKE_CUDA_ARCHITECTURES}\n"
      "do not all work with this compiler.  Try:\n"
      "  ${architectures}\n"
      "instead.")
  endif()
endif()

# configure all variables set in this file
configure_file(${CMAKE_ROOT}/Modules/CMakeCUDACompiler.cmake.in
  ${CMAKE_PLATFORM_INFO_DIR}/CMakeCUDACompiler.cmake
  @ONLY
)

# Don't leak variables unnecessarily to user code.
unset(_CUDA_INCLUDE_DIR CACHE)
unset(_CUDA_NVCC_EXECUTABLE CACHE)
unset(_CUDA_LIBRARY_DIR)
unset(_CUDA_TARGET_DIR)
unset(_CUDA_TARGET_NAME)

set(CMAKE_CUDA_COMPILER_ENV_VAR "CUDACXX")
set(CMAKE_CUDA_HOST_COMPILER_ENV_VAR "CUDAHOSTCXX")