CUDA: Add CUDA_ARCHITECTURES target property

Simplifies CUDA target architecture handling.

Required for Clang support as Clang doesn't automatically select a supported architecture.
We detect a supported architecture during compiler identification and set CMAKE_CUDA_ARCHITECTURES to it.

Introduces CMP0104 for backwards compatibility with manually setting code generation flags with NVCC.

Implements #17963.

1 files changed, 31 insertions, 0 deletions

diff --git a/Help/policy/CMP0104.rst b/Help/policy/CMP0104.rst
new file mode 100644
index 0000000..ca2c571
--- /dev/null
+++ b/Help/policy/CMP0104.rst
@@ -0,0 +1,31 @@
+CMP0104
+-------
+
+Initialize :variable:`CMAKE_CUDA_ARCHITECTURES` when
+:variable:`CMAKE_CUDA_COMPILER_ID <CMAKE_<LANG>_COMPILER_ID>` is ``NVIDIA``.
+Raise an error if :prop_tgt:`CUDA_ARCHITECTURES` is empty.
+
+:variable:`CMAKE_CUDA_ARCHITECTURES` introduced in CMake 3.18 is used to
+initialize :prop_tgt:`CUDA_ARCHITECTURES`, which passes correct code generation
+flags to the CUDA compiler.
+
+Previous to this users had to manually specify the code generation flags. This
+policy is for backwards compatibility with manually specifying code generation
+flags.
+
+The ``OLD`` behavior for this policy is to not initialize
+:variable:`CMAKE_CUDA_ARCHITECTURES` when
+:variable:`CMAKE_CUDA_COMPILER_ID <CMAKE_<LANG>_COMPILER_ID>` is ``NVIDIA``.
+Empty :prop_tgt:`CUDA_ARCHITECTURES` is allowed.
+
+The ``NEW`` behavior of this policy is to initialize
+:variable:`CMAKE_CUDA_ARCHITECTURES` when
+:variable:`CMAKE_CUDA_COMPILER_ID <CMAKE_<LANG>_COMPILER_ID>` is ``NVIDIA``
+and raise an error if :prop_tgt:`CUDA_ARCHITECTURES` is empty during generation.
+
+This policy was introduced in CMake version 3.18.  CMake version
+|release| warns when the policy is not set and uses ``OLD`` behavior.
+Use the :command:`cmake_policy` command to set it to ``OLD`` or ``NEW``
+explicitly.
+
+.. include:: DEPRECATED.txt