ci: Clamp CUDA 'native' architecture to values supported by its toolkit

The `native` architecture compiles for the host's GPUs, but our CI jobs
may may run on hosts with GPUs newer than supported by their version of
the CUDA toolkit.  Add an undocumented environment variable to tell
CMake to clamp the native architecture to that supported by the toolkit.

Without this, we may try to compile for architectures not supported by
the CUDA Toolkit, which fails.  Since commit d1b48bfabd (CUDA: Add
support for CUDA_ARCHITECTURES=native, 2022-03-04), our CUDA 9.2 CI job
fails when it runs on a CI host with a GPU architecture newer than that
CUDA 9.2 supports.  Clamping the architecture level fixes that.

Do not document this clamp behavior publicly, at least for now.
Users can be responsible for building with a CUDA toolkit recent
enough to support their host's GPUs.

Issue: #22375

.gitlab/os-linux.yml

@@ -227,6 +227,7 @@
         GIT_CLONE_PATH: "$CI_BUILDS_DIR/cmake ci"
         CMAKE_ARCH: x86_64
         CTEST_LABELS: "CUDA"
+        CMAKE_CUDA_ARCHITECTURES_NATIVE_CLAMP: 1
 
 .cuda9.2_nvidia:
     extends: .cuda9.2
@@ -242,6 +243,7 @@
         GIT_CLONE_PATH: "$CI_BUILDS_DIR/cmake ci"
         CMAKE_ARCH: x86_64
         CTEST_LABELS: "CUDA"
+        CMAKE_CUDA_ARCHITECTURES_NATIVE_CLAMP: 1
 
 .cuda10.2_nvidia:
     extends: .cuda10.2
@@ -264,6 +266,7 @@
         GIT_CLONE_PATH: "$CI_BUILDS_DIR/cmake ci"
         CMAKE_ARCH: x86_64
         CTEST_LABELS: "CUDA"
+        CMAKE_CUDA_ARCHITECTURES_NATIVE_CLAMP: 1
 
 .cuda11.6_nvidia:
     extends: .cuda11.6

Modules/CMakeTestCUDACompiler.cmake

@@ -30,7 +30,27 @@ if(CMAKE_CUDA_ABI_COMPILED)
     OUTPUT_STRIP_TRAILING_WHITESPACE
     )
   if(_CUDA_ARCHS_RESULT EQUAL 0)
-    set(CMAKE_CUDA_ARCHITECTURES_NATIVE "${_CUDA_ARCHS_OUTPUT}")
+    if("$ENV{CMAKE_CUDA_ARCHITECTURES_NATIVE_CLAMP}")
+      # Undocumented hook used by CMake's CI.
+      # Clamp native architecture to version range supported by this CUDA.
+      list(GET CMAKE_CUDA_ARCHITECTURES_ALL 0  _CUDA_ARCH_MIN)
+      list(GET CMAKE_CUDA_ARCHITECTURES_ALL -1 _CUDA_ARCH_MAX)
+      set(CMAKE_CUDA_ARCHITECTURES_NATIVE "")
+      foreach(_CUDA_ARCH IN LISTS _CUDA_ARCHS_OUTPUT)
+        if(_CUDA_ARCH LESS _CUDA_ARCH_MIN)
+          set(_CUDA_ARCH "${_CUDA_ARCH_MIN}")
+        endif()
+        if(_CUDA_ARCH GREATER _CUDA_ARCH_MAX)
+          set(_CUDA_ARCH "${_CUDA_ARCH_MAX}")
+        endif()
+        list(APPEND CMAKE_CUDA_ARCHITECTURES_NATIVE ${_CUDA_ARCH})
+      endforeach()
+      unset(_CUDA_ARCH)
+      unset(_CUDA_ARCH_MIN)
+      unset(_CUDA_ARCH_MAX)
+    else()
+      set(CMAKE_CUDA_ARCHITECTURES_NATIVE "${_CUDA_ARCHS_OUTPUT}")
+    endif()
     list(REMOVE_DUPLICATES CMAKE_CUDA_ARCHITECTURES_NATIVE)
   else()
     if(NOT _CUDA_ARCHS_RESULT MATCHES "[0-9]+")

Modules/CUDA/architectures.cmake

@@ -66,7 +66,27 @@ if(NOT CMAKE_GENERATOR MATCHES "Visual Studio")
       )
   endif()
   if(_CUDA_ARCHS_RESULT EQUAL 0)
-    set(_CUDA_ARCHITECTURES_NATIVE "${_CUDA_ARCHS_OUTPUT}")
+    if("$ENV{CMAKE_CUDA_ARCHITECTURES_NATIVE_CLAMP}")
+      # Undocumented hook used by CMake's CI.
+      # Clamp native architecture to version range supported by this CUDA.
+      list(GET CMAKE_CUDA_ARCHITECTURES_ALL 0  _CUDA_ARCH_MIN)
+      list(GET CMAKE_CUDA_ARCHITECTURES_ALL -1 _CUDA_ARCH_MAX)
+      set(_CUDA_ARCHITECTURES_NATIVE "")
+      foreach(_CUDA_ARCH IN LISTS _CUDA_ARCHS_OUTPUT)
+        if(_CUDA_ARCH LESS _CUDA_ARCH_MIN)
+          set(_CUDA_ARCH "${_CUDA_ARCH_MIN}")
+        endif()
+        if(_CUDA_ARCH GREATER _CUDA_ARCH_MAX)
+          set(_CUDA_ARCH "${_CUDA_ARCH_MAX}")
+        endif()
+        list(APPEND _CUDA_ARCHITECTURES_NATIVE ${_CUDA_ARCH})
+      endforeach()
+      unset(_CUDA_ARCH)
+      unset(_CUDA_ARCH_MIN)
+      unset(_CUDA_ARCH_MAX)
+    else()
+      set(_CUDA_ARCHITECTURES_NATIVE "${_CUDA_ARCHS_OUTPUT}")
+    endif()
     list(REMOVE_DUPLICATES _CUDA_ARCHITECTURES_NATIVE)
   else()
     if (NOT _CUDA_ARCHS_RESULT MATCHES "[0-9]+")