CUDA: Add abstraction for cuda runtime selection

Fixes #17559
Replace our hard-coded default of cudart=static with a first-class abstraction to select the runtime library from an enumeration of logical names.

2 files changed, 30 insertions, 0 deletions

diff --git a/Help/prop_tgt/CUDA_RUNTIME_LIBRARY-VALUES.txt b/Help/prop_tgt/CUDA_RUNTIME_LIBRARY-VALUES.txt
new file mode 100644
index 0000000..a6d7050
--- /dev/null
+++ b/Help/prop_tgt/CUDA_RUNTIME_LIBRARY-VALUES.txt
@@ -0,0 +1,9 @@
+``None``
+  Link with ``-cudart=none`` or equivalent flag(s) to use no CUDA
+  runtime library.
+``Shared``
+  Link with ``-cudart=shared`` or equivalent flag(s) to use a
+  dynamically-linked CUDA runtime library.
+``Static``
+  Link with ``-cudart=static`` or equivalent flag(s) to use a
+  statically-linked CUDA runtime library.
diff --git a/Help/prop_tgt/CUDA_RUNTIME_LIBRARY.rst b/Help/prop_tgt/CUDA_RUNTIME_LIBRARY.rst
new file mode 100644
index 0000000..0782765
--- /dev/null
+++ b/Help/prop_tgt/CUDA_RUNTIME_LIBRARY.rst
@@ -0,0 +1,21 @@
+CUDA_RUNTIME_LIBRARY
+--------------------
+
+Select the CUDA runtime library for use by compilers targeting the CUDA language.
+
+The allowed case insensitive values are:
+
+.. include:: CUDA_RUNTIME_LIBRARY-VALUES.txt
+
+Contents of ``CUDA_RUNTIME_LIBRARY`` may use
+:manual:`generator expressions <cmake-generator-expressions(7)>`.
+
+If this property is not set then CMake uses the default value
+``Static`` to select the CUDA runtime library.
+
+.. note::
+
+  This property has effect only when the ``CUDA`` language is enabled. To
+  control the CUDA runtime linking when only using the CUDA SDK with the
+  ``C`` or ``C++`` language we recommend using the :module:`FindCUDAToolkit`
+  module.