CUDA: Fail fast if CMAKE_CUDA_ARCHITECTURES doesn't work during detection

Also re-ordered the code to avoid testing flags for other compilers, since
we know the vendor before full detection.

Help/release/dev/cuda-fail-fast.rst

@@ -2,4 +2,5 @@ cuda-fail-fast
 --------------
 
 * If ``CUDA`` compiler detection fails with user-specified
-  :variable:`CMAKE_CUDA_ARCHITECTURES` an error is raised.
+  :variable:`CMAKE_CUDA_ARCHITECTURES` or :variable:`CMAKE_CUDA_HOST_COMPILER`
+  an error is raised.

Modules/CMakeDetermineCUDACompiler.cmake

@@ -206,34 +206,35 @@ if(NOT CMAKE_CUDA_COMPILER_ID_RUN)
     endif()
   endif()
 
-  # First try with the user-specified architectures.
+  # Append 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_${arch_name}")
+      string(APPEND clang_test_flags " --cuda-gpu-arch=sm_${arch_name}")
+      string(APPEND nvcc_test_flags " -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}")
+    # If the user has specified architectures we'll want to fail during compiler detection if they don't work.
+    set(CMAKE_CUDA_COMPILER_ID_REQUIRE_SUCCESS ON)
   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_${arch}")
-  endforeach()
+  if(CMAKE_CUDA_COMPILER_ID STREQUAL "Clang")
+    if(NOT CMAKE_CUDA_ARCHITECTURES)
+      # Clang doesn't automatically select an architecture supported by the SDK.
+      # Try in reverse order of deprecation with the most recent at front (i.e. the most likely to work for new setups).
+      foreach(arch "20" "30" "52")
+        list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${clang_test_flags} --cuda-gpu-arch=sm_${arch}")
+      endforeach()
+    endif()
 
-  # Finally also try the default.
-  list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${clang_test_flags}")
+    # If the user specified CMAKE_CUDA_ARCHITECTURES this will include all the architecture flags.
+    # Otherwise this won't include any architecture flags and we'll fallback to Clang's defaults.
+    list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${clang_test_flags}")
+  elseif(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
+    list(APPEND CMAKE_CUDA_COMPILER_ID_TEST_FLAGS_FIRST "${nvcc_test_flags}")
+  endif()
 
   # 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.