Merge topic 'FindCUDAToolkit-implicit-dirs'

4316d4dcfd FindCUDAToolkit: Search all of `nvcc` implicit includes and library dirs

Acked-by: Kitware Robot <kwrobot@kitware.com>
Merge-request: !8835

7 files changed, 170 insertions, 23 deletions

diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index c13ba72..3c1c9cc 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -350,6 +350,16 @@ t:cuda11.8-minimal-ninja:
     variables:
         CMAKE_CI_NO_MR: "true"
 
+t:cuda11.8-minimal-splayed-ninja:
+    extends:
+        - .cuda11.8_splayed_nvidia
+        - .cmake_test_linux_release
+        - .linux_x86_64_tags_cuda
+        - .run_dependent
+        - .needs_centos7_x86_64
+    variables:
+        CMAKE_CI_NO_MR: "true"
+
 t:hip5.5-nvidia:
     extends:
         - .hip5.5_nvidia
diff --git a/.gitlab/ci/configure_cuda11.8_splayed_nvidia.cmake b/.gitlab/ci/configure_cuda11.8_splayed_nvidia.cmake
new file mode 100644
index 0000000..519699b
--- /dev/null
+++ b/.gitlab/ci/configure_cuda11.8_splayed_nvidia.cmake
@@ -0,0 +1,3 @@
+set(CMake_TEST_CUDA "NVIDIA" CACHE STRING "")
+
+include("${CMAKE_CURRENT_LIST_DIR}/configure_external_test.cmake")
diff --git a/.gitlab/ci/env_cuda11.8_splayed_nvidia.sh b/.gitlab/ci/env_cuda11.8_splayed_nvidia.sh
new file mode 100644
index 0000000..38e788d
--- /dev/null
+++ b/.gitlab/ci/env_cuda11.8_splayed_nvidia.sh
@@ -0,0 +1,36 @@
+#
+# Splay the libraries and includes to emulate conda where
+# things are split between the host and build prefix
+#
+# /usr/local/cuda/include/crt/ -> /tmp/cuda/include/crt
+# /usr/local/cuda/lib64/stubs/ -> /tmp/cuda/stubs/
+# /usr/local/cuda/lib64/libcudart* -> /tmp/cuda/libs/
+#
+# Also reduce to minimal subset of libraries by removing
+# static libraries to emulate a minimal cuda install
+mkdir -p /tmp/cuda/libs
+mkdir -p /tmp/cuda/stubs
+mkdir -p /tmp/cuda/include
+
+mv /usr/local/cuda/lib64/libcuda* /tmp/cuda/libs
+mv /usr/local/cuda/lib64/stubs/ /tmp/cuda/stubs/
+mv /usr/local/cuda/include/crt/ /tmp/cuda/include/
+
+# patch the nvcc.profile to handle the splayed layout
+# which allows verification
+mv /usr/local/cuda/bin/nvcc.profile /usr/local/cuda/bin/nvcc.profile.orig
+echo "
+TOP              = \$(_HERE_)/..
+
+NVVMIR_LIBRARY_DIR = \$(TOP)/\$(_NVVM_BRANCH_)/libdevice
+
+LD_LIBRARY_PATH += \$(TOP)/lib:
+PATH            += \$(TOP)/\$(_NVVM_BRANCH_)/bin:\$(_HERE_):
+
+INCLUDES        +=  \"-I\$(TOP)/\$(_TARGET_DIR_)/include\" \$(_SPACE_) \"-I/tmp/cuda/include\" \$(_SPACE_)
+
+LIBRARIES        =+ \$(_SPACE_) \"-L\$(TOP)/\$(_TARGET_DIR_)/lib\$(_TARGET_SIZE_)\" \"-L/tmp/cuda/stubs/\" \"-L/tmp/cuda/libs\"
+
+CUDAFE_FLAGS    +=
+PTXAS_FLAGS     +=
+" > /usr/local/cuda/bin/nvcc.profile
diff --git a/.gitlab/os-linux.yml b/.gitlab/os-linux.yml
index e70dc07..8894057 100644
--- a/.gitlab/os-linux.yml
+++ b/.gitlab/os-linux.yml
@@ -362,6 +362,12 @@
         CMAKE_CONFIGURATION: cuda11.8_minimal_nvidia
         CTEST_NO_WARNINGS_ALLOWED: 1
 
+.cuda11.8_splayed_nvidia:
+    extends: .cuda11.8_minimal
+    variables:
+        CMAKE_CONFIGURATION: cuda11.8_splayed_nvidia
+        CTEST_NO_WARNINGS_ALLOWED: 1
+
 ### HIP builds
 
 .hip5.5:
diff --git a/Modules/FindCUDAToolkit.cmake b/Modules/FindCUDAToolkit.cmake
index 12aca8d..bd9e0ec 100644
--- a/Modules/FindCUDAToolkit.cmake
+++ b/Modules/FindCUDAToolkit.cmake
@@ -514,7 +514,7 @@ Result variables
     executable ``nvcc``.
 
 ``CUDAToolkit_INCLUDE_DIRS``
-    The path to the CUDA Toolkit ``include`` folder containing the header files
+    List of paths to all the CUDA Toolkit folders containing header files
     required to compile a project linking against CUDA.
 
 ``CUDAToolkit_LIBRARY_DIR``
@@ -579,13 +579,28 @@ Result variables
 #
 ###############################################################################
 
+function(_CUDAToolkit_build_include_dirs result_variable default_paths_variable)
+  set(content "${${default_paths_variable}}")
+  set(${result_variable} "${content}" PARENT_SCOPE)
+endfunction()
+
+function(_CUDAToolkit_build_library_dirs result_variable default_paths_variable)
+  set(content "${${default_paths_variable}}")
+  set(${result_variable} "${content}" PARENT_SCOPE)
+endfunction()
+
 # The toolkit is located during compiler detection for CUDA and stored in CMakeCUDACompiler.cmake as
-# CMAKE_CUDA_COMPILER_TOOLKIT_ROOT and CMAKE_CUDA_COMPILER_LIBRARY_ROOT.
+# - CMAKE_CUDA_COMPILER_TOOLKIT_ROOT
+# - CMAKE_CUDA_COMPILER_LIBRARY_ROOT
+# - CMAKE_CUDA_COMPILER_LIBRARY_DIRECTORIES_FROM_IMPLICIT_LIBRARIES
+# - CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES
 # We compute the rest based on those here to avoid re-searching and to avoid finding a possibly
 # different installation.
 if(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT)
   set(CUDAToolkit_ROOT_DIR "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}")
   set(CUDAToolkit_LIBRARY_ROOT "${CMAKE_CUDA_COMPILER_LIBRARY_ROOT}")
+  _CUDAToolkit_build_library_dirs(CUDAToolkit_IMPLICIT_LIBRARY_DIRECTORIES CMAKE_CUDA_HOST_IMPLICIT_LINK_DIRECTORIES)
+  _CUDAToolkit_build_include_dirs(CUDAToolkit_INCLUDE_DIRECTORIES CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES)
   set(CUDAToolkit_BIN_DIR "${CUDAToolkit_ROOT_DIR}/bin")
   set(CUDAToolkit_NVCC_EXECUTABLE "${CUDAToolkit_BIN_DIR}/nvcc${CMAKE_EXECUTABLE_SUFFIX}")
   set(CUDAToolkit_VERSION "${CMAKE_CUDA_COMPILER_TOOLKIT_VERSION}")
@@ -622,11 +637,45 @@ else()
         # NVIDIA HPC SDK, and distro's splayed layouts
         execute_process(COMMAND ${CUDAToolkit_NVCC_EXECUTABLE} "-v" "__cmake_determine_cuda"
           OUTPUT_VARIABLE _CUDA_NVCC_OUT ERROR_VARIABLE _CUDA_NVCC_OUT)
+        message(CONFIGURE_LOG
+          "Executed nvcc to extract CUDAToolkit information:\n${_CUDA_NVCC_OUT}\n\n")
         if(_CUDA_NVCC_OUT MATCHES "\\#\\$ TOP=([^\r\n]*)")
           get_filename_component(CUDAToolkit_BIN_DIR "${CMAKE_MATCH_1}/bin" ABSOLUTE)
+          message(CONFIGURE_LOG
+            "Parsed CUDAToolkit nvcc location:\n${CUDAToolkit_BIN_DIR}\n\n")
         else()
           get_filename_component(CUDAToolkit_BIN_DIR "${CUDAToolkit_NVCC_EXECUTABLE}" DIRECTORY)
         endif()
+        if(_CUDA_NVCC_OUT MATCHES "\\#\\$ INCLUDES=([^\r\n]*)")
+          separate_arguments(_nvcc_output NATIVE_COMMAND "${CMAKE_MATCH_1}")
+          foreach(line IN LISTS _nvcc_output)
+            string(REGEX REPLACE "^-I" "" line "${line}")
+            get_filename_component(line "${line}" ABSOLUTE)
+            list(APPEND _cmake_CUDAToolkit_include_directories "${line}")
+          endforeach()
+          message(CONFIGURE_LOG
+            "Parsed CUDAToolkit nvcc implicit include information:\n${_cmake_CUDAToolkit_include_directories}\n\n")
+
+          set(_cmake_CUDAToolkit_include_directories "${_cmake_CUDAToolkit_include_directories}" CACHE INTERNAL "CUDAToolkit internal list of include directories")
+        endif()
+        if(_CUDA_NVCC_OUT MATCHES "\\#\\$ LIBRARIES=([^\r\n]*)")
+          include(${CMAKE_ROOT}/Modules/CMakeParseImplicitLinkInfo.cmake)
+          set(_nvcc_link_line "cuda-fake-ld ${CMAKE_MATCH_1}")
+          CMAKE_PARSE_IMPLICIT_LINK_INFO("${_nvcc_link_line}"
+                                   _cmake_CUDAToolkit_implicit_link_libs
+                                   _cmake_CUDAToolkit_implicit_link_directories
+                                   _cmake_CUDAToolkit_implicit_frameworks
+                                   _nvcc_log
+                                   "${CMAKE_CUDA_IMPLICIT_OBJECT_REGEX}"
+                                   LANGUAGE CUDA)
+          message(CONFIGURE_LOG
+          "Parsed CUDAToolkit nvcc implicit link information:\n${_nvcc_log}\n${_cmake_CUDAToolkit_implicit_link_directories}\n\n")
+          unset(_nvcc_link_line)
+          unset(_cmake_CUDAToolkit_implicit_link_libs)
+          unset(_cmake_CUDAToolkit_implicit_frameworks)
+
+          set(_cmake_CUDAToolkit_implicit_link_directories "${_cmake_CUDAToolkit_implicit_link_directories}" CACHE INTERNAL "CUDAToolkit internal list of implicit link directories")
+        endif()
         unset(_CUDA_NVCC_OUT)
 
         set(CUDAToolkit_BIN_DIR "${CUDAToolkit_BIN_DIR}" CACHE PATH "" FORCE)
@@ -642,6 +691,15 @@ else()
       endif()
     endif()
 
+    if(DEFINED _cmake_CUDAToolkit_include_directories)
+      _CUDAToolkit_build_include_dirs(_cmake_CUDAToolkit_contents _cmake_CUDAToolkit_include_directories)
+      set(CUDAToolkit_INCLUDE_DIRECTORIES "${_cmake_CUDAToolkit_contents}" PARENT_SCOPE)
+    endif()
+    if(DEFINED _cmake_CUDAToolkit_implicit_link_directories)
+      _CUDAToolkit_build_library_dirs(_cmake_CUDAToolkit_contents _cmake_CUDAToolkit_implicit_link_directories)
+      set(CUDAToolkit_IMPLICIT_LIBRARY_DIRECTORIES "${_cmake_CUDAToolkit_contents}" PARENT_SCOPE)
+    endif()
+
     if(CUDAToolkit_BIN_DIR)
       get_filename_component(CUDAToolkit_ROOT_DIR ${CUDAToolkit_BIN_DIR} DIRECTORY ABSOLUTE)
       set(CUDAToolkit_ROOT_DIR "${CUDAToolkit_ROOT_DIR}" PARENT_SCOPE)
@@ -885,18 +943,27 @@ if(NOT CUDAToolkit_TARGET_DIR)
   set(_CUDAToolkit_Pop_Prefix True)
 endif()
 
-# CUDAToolkit_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 "${CUDAToolkit_TARGET_DIR}/include/cuda_runtime.h")
-  set(CUDAToolkit_INCLUDE_DIR "${CUDAToolkit_TARGET_DIR}/include")
-elseif(NOT CUDAToolkit_FIND_QUIETLY)
-  message(STATUS "Unable to find cuda_runtime.h in \"${CUDAToolkit_TARGET_DIR}/include\" for CUDAToolkit_INCLUDE_DIR.")
+
+# We don't need to verify the cuda_runtime header when we are using `nvcc` include paths
+# as the compiler being enabled means the header was found
+if(NOT CUDAToolkit_INCLUDE_DIRECTORIES)
+  # Otherwise use CUDAToolkit_TARGET_DIR to guess where the `cuda_runtime.h` is located
+  # On a scattered installation /usr, on a non-scattered something like /usr/local/cuda or /usr/local/cuda-10.2/targets/aarch64-linux.
+  if(EXISTS "${CUDAToolkit_TARGET_DIR}/include/cuda_runtime.h")
+    set(CUDAToolkit_INCLUDE_DIRECTORIES "${CUDAToolkit_TARGET_DIR}/include")
+  else()
+    message(STATUS "Unable to find cuda_runtime.h in \"${CUDAToolkit_TARGET_DIR}/include\" for CUDAToolkit_INCLUDE_DIRECTORIES.")
+  endif()
 endif()
 
 # The NVHPC layout moves math library headers and libraries to a sibling directory and it could be nested under
 # the version of the CUDA toolchain
 # Create a separate variable so this directory can be selectively added to math targets.
-if(NOT EXISTS "${CUDAToolkit_INCLUDE_DIR}/cublas_v2.h")
+find_path(CUDAToolkit_CUBLAS_INCLUDE_DIR cublas_v2.h PATHS
+  "${CUDAToolkit_INCLUDE_DIRECTORIES}"
+  NO_DEFAULT_PATH)
+
+if(NOT CUDAToolkit_CUBLAS_INCLUDE_DIR)
   file(REAL_PATH "${CUDAToolkit_TARGET_DIR}" CUDAToolkit_MATH_INCLUDE_DIR)
   cmake_path(APPEND CUDAToolkit_MATH_INCLUDE_DIR "../../math_libs/")
   if(EXISTS "${CUDAToolkit_MATH_INCLUDE_DIR}/${CUDAToolkit_VERSION_MAJOR}.${CUDAToolkit_VERSION_MINOR}/")
@@ -905,22 +972,26 @@ if(NOT EXISTS "${CUDAToolkit_INCLUDE_DIR}/cublas_v2.h")
   cmake_path(APPEND CUDAToolkit_MATH_INCLUDE_DIR "include")
   cmake_path(NORMAL_PATH CUDAToolkit_MATH_INCLUDE_DIR)
 
-  if(NOT EXISTS "${CUDAToolkit_MATH_INCLUDE_DIR}/cublas_v2.h")
-    if(NOT CUDAToolkit_FIND_QUIETLY)
-      message(STATUS "Unable to find cublas_v2.h in either \"${CUDAToolkit_INCLUDE_DIR}\" or \"${CUDAToolkit_MATH_INCLUDE_DIR}\"")
-    endif()
-    unset(CUDAToolkit_MATH_INCLUDE_DIR)
+  find_path(CUDAToolkit_CUBLAS_INCLUDE_DIR cublas_v2.h PATHS
+    "${CUDAToolkit_INCLUDE_DIRECTORIES}"
+    )
+  if(CUDAToolkit_CUBLAS_INCLUDE_DIR)
+    list(APPEND CUDAToolkit_INCLUDE_DIRECTORIES "${CUDAToolkit_CUBLAS_INCLUDE_DIR}")
   endif()
 endif()
+unset(CUDAToolkit_CUBLAS_INCLUDE_DIR CACHE)
+unset(CUDAToolkit_CUBLAS_INCLUDE_DIR)
 
 # Find the CUDA Runtime Library libcudart
 find_library(CUDA_CUDART
   NAMES cudart
+  PATHS ${CUDAToolkit_IMPLICIT_LIBRARY_DIRECTORIES}
   PATH_SUFFIXES lib64 lib/x64
 )
 find_library(CUDA_CUDART
   NAMES cudart
-  PATH_SUFFIXES lib64/stubs lib/x64/stubs
+  PATHS ${CUDAToolkit_IMPLICIT_LIBRARY_DIRECTORIES}
+  PATH_SUFFIXES lib64/stubs lib/x64/stubs lib/stubs stubs
 )
 
 if(NOT CUDA_CUDART AND NOT CUDAToolkit_FIND_QUIETLY)
@@ -937,7 +1008,7 @@ endif()
 include(${CMAKE_CURRENT_LIST_DIR}/FindPackageHandleStandardArgs.cmake)
 find_package_handle_standard_args(CUDAToolkit
   REQUIRED_VARS
-    CUDAToolkit_INCLUDE_DIR
+    CUDAToolkit_INCLUDE_DIRECTORIES
     CUDA_CUDART
     CUDAToolkit_BIN_DIR
   VERSION_VAR
@@ -946,7 +1017,6 @@ find_package_handle_standard_args(CUDAToolkit
 
 unset(CUDAToolkit_ROOT_DIR)
 mark_as_advanced(CUDA_CUDART
-                 CUDAToolkit_INCLUDE_DIR
                  CUDAToolkit_NVCC_EXECUTABLE
                  CUDAToolkit_SENTINEL_FILE
                  )
@@ -954,7 +1024,7 @@ mark_as_advanced(CUDA_CUDART
 #-----------------------------------------------------------------------------
 # Construct result variables
 if(CUDAToolkit_FOUND)
-  set(CUDAToolkit_INCLUDE_DIRS ${CUDAToolkit_INCLUDE_DIR})
+  set(CUDAToolkit_INCLUDE_DIRS "${CUDAToolkit_INCLUDE_DIRECTORIES}")
   get_filename_component(CUDAToolkit_LIBRARY_DIR ${CUDA_CUDART} DIRECTORY ABSOLUTE)
 
   # Build search paths without any symlinks
@@ -976,6 +1046,10 @@ if(CUDAToolkit_FOUND)
     endblock()
   endif()
 
+  if(DEFINED CUDAToolkit_IMPLICIT_LIBRARY_DIRECTORIES)
+    list(APPEND CUDAToolkit_LIBRARY_SEARCH_DIRS "${CUDAToolkit_IMPLICIT_LIBRARY_DIRECTORIES}")
+  endif()
+
   # If no `CUDAToolkit_LIBRARY_ROOT` exists set it based on CUDAToolkit_LIBRARY_DIR
   if(NOT DEFINED CUDAToolkit_LIBRARY_ROOT)
     foreach(CUDAToolkit_search_loc IN LISTS CUDAToolkit_LIBRARY_DIR CUDAToolkit_BIN_DIR)
@@ -989,7 +1063,8 @@ if(CUDAToolkit_FOUND)
     unset(CUDAToolkit_possible_lib_root)
   endif()
 endif()
-
+unset(CUDAToolkit_IMPLICIT_LIBRARY_DIRECTORIES)
+unset(CUDAToolkit_INCLUDE_DIRECTORIES)
 
 #-----------------------------------------------------------------------------
 # Construct import targets
@@ -1257,6 +1332,7 @@ if(CUDAToolkit_FOUND)
   _CUDAToolkit_find_and_add_import_lib(OpenCL)
 endif()
 
+unset(CUDAToolkit_LIBRARY_SEARCH_DIRS)
 if(_CUDAToolkit_Pop_ROOT_PATH)
   list(REMOVE_AT CMAKE_FIND_ROOT_PATH 0)
   unset(_CUDAToolkit_Pop_ROOT_PATH)
diff --git a/Tests/Cuda/Toolkit/CMakeLists.txt b/Tests/Cuda/Toolkit/CMakeLists.txt
index 8432b71..c2989f0 100644
--- a/Tests/Cuda/Toolkit/CMakeLists.txt
+++ b/Tests/Cuda/Toolkit/CMakeLists.txt
@@ -26,9 +26,11 @@ set(should_exist
   CUDAToolkit_LIBRARY_ROOT
   )
 foreach (cuda_loc_var IN LISTS should_exist)
-  if(NOT EXISTS "${${cuda_loc_var}}")
-    message(FATAL_ERROR "${cuda_loc_var} variable is expected to be set to valid path")
-  endif()
+  foreach (entry IN LISTS ${cuda_loc_var})
+    if(NOT EXISTS "${entry}")
+      message(FATAL_ERROR "${cuda_loc_var} variable is expected to be set to valid path")
+    endif()
+  endforeach()
 endforeach()
 
 
diff --git a/Tests/CudaOnly/Toolkit/CMakeLists.txt b/Tests/CudaOnly/Toolkit/CMakeLists.txt
index e0801a3..506fc9f 100644
--- a/Tests/CudaOnly/Toolkit/CMakeLists.txt
+++ b/Tests/CudaOnly/Toolkit/CMakeLists.txt
@@ -1,5 +1,5 @@
 cmake_minimum_required(VERSION 3.15)
-project(CudaOnlyToolkit CUDA)
+project(CudaOnlyToolkit LANGUAGES CUDA)
 find_package(CUDAToolkit REQUIRED)
 
 if(NOT DEFINED CUDAToolkit_VERSION)
@@ -13,8 +13,22 @@ message(STATUS "CUDAToolkit_VERSION_PATCH: ${CUDAToolkit_VERSION_PATCH}")
 message(STATUS "CUDAToolkit_BIN_DIR: ${CUDAToolkit_BIN_DIR}")
 message(STATUS "CUDAToolkit_INCLUDE_DIRS: ${CUDAToolkit_INCLUDE_DIRS}")
 message(STATUS "CUDAToolkit_LIBRARY_DIR: ${CUDAToolkit_LIBRARY_DIR}")
+message(STATUS "CUDAToolkit_LIBRARY_ROOT: ${CUDAToolkit_LIBRARY_ROOT}")
 message(STATUS "CUDAToolkit_NVCC_EXECUTABLE ${CUDAToolkit_NVCC_EXECUTABLE}")
 
+set(should_exist
+  CUDAToolkit_BIN_DIR
+  CUDAToolkit_INCLUDE_DIRS
+  CUDAToolkit_LIBRARY_DIR
+  CUDAToolkit_LIBRARY_ROOT
+  )
+foreach (cuda_loc_var IN LISTS should_exist)
+  foreach (entry IN LISTS ${cuda_loc_var})
+    if(NOT EXISTS "${entry}")
+      message(FATAL_ERROR "${cuda_loc_var}[${entry}] variable is expected to be set to valid path")
+    endif()
+  endforeach()
+endforeach()
 
 set(cuda_libs cudart cuda_driver cublas cufft cufftw curand cusolver cusparse)
 if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 10.1)