9 files changed, 45 insertions, 4 deletions
diff --git a/.gitlab/ci/configure_cuda10.2_clang.cmake b/.gitlab/ci/configure_cuda10.2_clang.cmake
index e13ca88..e2d5a9c 100644
--- a/.gitlab/ci/configure_cuda10.2_clang.cmake
+++ b/.gitlab/ci/configure_cuda10.2_clang.cmake
@@ -1,3 +1,4 @@
 set(CMake_TEST_CUDA "Clang" CACHE STRING "")
+set(CMake_TEST_CUDA_CUPTI "ON" CACHE STRING "")
 
 include("${CMAKE_CURRENT_LIST_DIR}/configure_external_test.cmake")
diff --git a/.gitlab/ci/configure_cuda10.2_nvidia.cmake b/.gitlab/ci/configure_cuda10.2_nvidia.cmake
index 519699b..2cb2950 100644
--- a/.gitlab/ci/configure_cuda10.2_nvidia.cmake
+++ b/.gitlab/ci/configure_cuda10.2_nvidia.cmake
@@ -1,3 +1,4 @@
 set(CMake_TEST_CUDA "NVIDIA" CACHE STRING "")
+set(CMake_TEST_CUDA_CUPTI "ON" CACHE STRING "")
 
 include("${CMAKE_CURRENT_LIST_DIR}/configure_external_test.cmake")
diff --git a/.gitlab/ci/configure_cuda11.6_nvidia.cmake b/.gitlab/ci/configure_cuda11.6_nvidia.cmake
index 519699b..2cb2950 100644
--- a/.gitlab/ci/configure_cuda11.6_nvidia.cmake
+++ b/.gitlab/ci/configure_cuda11.6_nvidia.cmake
@@ -1,3 +1,4 @@
 set(CMake_TEST_CUDA "NVIDIA" CACHE STRING "")
+set(CMake_TEST_CUDA_CUPTI "ON" CACHE STRING "")
 
 include("${CMAKE_CURRENT_LIST_DIR}/configure_external_test.cmake")
diff --git a/.gitlab/ci/configure_nvhpc_ninja.cmake b/.gitlab/ci/configure_nvhpc_ninja.cmake
index ca8ba93..07e47c4 100644
--- a/.gitlab/ci/configure_nvhpc_ninja.cmake
+++ b/.gitlab/ci/configure_nvhpc_ninja.cmake
@@ -1,4 +1,5 @@
 set(CMake_TEST_CUDA "NVIDIA" CACHE STRING "")
+set(CMake_TEST_CUDA_CUPTI "ON" CACHE STRING "")
 
 set(configure_no_sccache 1)
 
diff --git a/Modules/FindCUDAToolkit.cmake b/Modules/FindCUDAToolkit.cmake
index b4c596c..469a7ad 100644
--- a/Modules/FindCUDAToolkit.cmake
+++ b/Modules/FindCUDAToolkit.cmake
@@ -960,7 +960,7 @@ endif()
 # the version of the CUDA toolchain
 # Create a separate variable so this directory can be selectively added to math targets.
 find_path(CUDAToolkit_CUBLAS_INCLUDE_DIR cublas_v2.h PATHS
-  "${CUDAToolkit_INCLUDE_DIRECTORIES}"
+  ${CUDAToolkit_INCLUDE_DIRECTORIES}
   NO_DEFAULT_PATH)
 
 if(NOT CUDAToolkit_CUBLAS_INCLUDE_DIR)
@@ -973,7 +973,7 @@ if(NOT CUDAToolkit_CUBLAS_INCLUDE_DIR)
   cmake_path(NORMAL_PATH CUDAToolkit_MATH_INCLUDE_DIR)
 
   find_path(CUDAToolkit_CUBLAS_INCLUDE_DIR cublas_v2.h PATHS
-    "${CUDAToolkit_INCLUDE_DIRECTORIES}"
+    ${CUDAToolkit_INCLUDE_DIRECTORIES}
     )
   if(CUDAToolkit_CUBLAS_INCLUDE_DIR)
     list(APPEND CUDAToolkit_INCLUDE_DIRECTORIES "${CUDAToolkit_CUBLAS_INCLUDE_DIR}")
@@ -1246,7 +1246,7 @@ if(CUDAToolkit_FOUND)
 
   find_path(CUDAToolkit_CUPTI_INCLUDE_DIR cupti.h PATHS
       "${CUDAToolkit_ROOT_DIR}/extras/CUPTI/include"
-      "${CUDAToolkit_INCLUDE_DIRS}"
+      ${CUDAToolkit_INCLUDE_DIRS}
       PATH_SUFFIXES "../extras/CUPTI/include"
                     "../../../extras/CUPTI/include"
       NO_DEFAULT_PATH)
diff --git a/Tests/Cuda/CMakeLists.txt b/Tests/Cuda/CMakeLists.txt
index c737bcc..efe0358 100644
--- a/Tests/Cuda/CMakeLists.txt
+++ b/Tests/Cuda/CMakeLists.txt
@@ -14,6 +14,7 @@ add_cuda_test_macro(Cuda.MixedStandardLevels5 MixedStandardLevels5)
 add_cuda_test_macro(Cuda.NotEnabled CudaNotEnabled)
 add_cuda_test_macro(Cuda.SeparableCompCXXOnly SeparableCompCXXOnly)
 add_cuda_test_macro(Cuda.StubRPATH StubRPATH)
+set(Cuda.Toolkit_BUILD_OPTIONS -DHAS_CUPTI:BOOL=${CMake_TEST_CUDA_CUPTI})
 add_cuda_test_macro(Cuda.Toolkit Toolkit)
 add_cuda_test_macro(Cuda.IncludePathNoToolkit IncludePathNoToolkit)
 add_cuda_test_macro(Cuda.SharedRuntimePlusToolkit SharedRuntimePlusToolkit)
diff --git a/Tests/Cuda/Toolkit/CMakeLists.txt b/Tests/Cuda/Toolkit/CMakeLists.txt
index c2989f0..4255b82 100644
--- a/Tests/Cuda/Toolkit/CMakeLists.txt
+++ b/Tests/Cuda/Toolkit/CMakeLists.txt
@@ -79,6 +79,23 @@ endforeach()
 add_executable(Toolkit main.cpp)
 target_link_libraries(Toolkit PRIVATE CUDA::toolkit)
 
+if(HAS_CUPTI)
+  set(cupti_libs )
+  if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 10.2)
+    list(APPEND cupti_libs cupti nvperf_target)
+  endif()
+  if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 11.3)
+    list(APPEND cupti_libs pcsamplingutil)
+  endif()
+  foreach (cuda_lib IN LISTS cupti_libs)
+    if(NOT CUDA_${cuda_lib}_LIBRARY)
+      message(FATAL_ERROR "expected CUDAToolkit variable CUDA_${cuda_lib}_LIBRARY not found")
+    endif()
+    if(NOT TARGET CUDA::${cuda_lib})
+      message(FATAL_ERROR "expected CUDAToolkit target CUDA::${cuda_lib} not found")
+    endif()
+  endforeach()
+endif()
 # cupti is an optional component of the CUDA toolkit
 if(TARGET CUDA::cupti)
   add_executable(cupti cupti.cpp)
diff --git a/Tests/CudaOnly/CMakeLists.txt b/Tests/CudaOnly/CMakeLists.txt
index 32940eb..9f49498 100644
--- a/Tests/CudaOnly/CMakeLists.txt
+++ b/Tests/CudaOnly/CMakeLists.txt
@@ -12,6 +12,7 @@ add_cuda_test_macro(CudaOnly.ExportPTX CudaOnlyExportPTX)
 add_cuda_test_macro(CudaOnly.SharedRuntimePlusToolkit CudaOnlySharedRuntimePlusToolkit)
 add_cuda_test_macro(CudaOnly.StaticRuntimePlusToolkit CudaOnlyStaticRuntimePlusToolkit)
 add_cuda_test_macro(CudaOnly.Standard98 CudaOnlyStandard98)
+set(CudaOnly.Toolkit_BUILD_OPTIONS -DHAS_CUPTI:BOOL=${CMake_TEST_CUDA_CUPTI})
 add_cuda_test_macro(CudaOnly.Toolkit CudaOnlyToolkit)
 add_cuda_test_macro(CudaOnly.ToolkitBeforeLang CudaOnlyToolkitBeforeLang)
 add_cuda_test_macro(CudaOnly.ToolkitMultipleDirs CudaOnlyToolkitMultipleDirs)
diff --git a/Tests/CudaOnly/Toolkit/CMakeLists.txt b/Tests/CudaOnly/Toolkit/CMakeLists.txt
index 506fc9f..b2694bf 100644
--- a/Tests/CudaOnly/Toolkit/CMakeLists.txt
+++ b/Tests/CudaOnly/Toolkit/CMakeLists.txt
@@ -41,7 +41,6 @@ if(CUDAToolkit_VERSION_MAJOR VERSION_LESS 11)
   list(APPEND cuda_libs nvgraph)
 endif()
 
-
 # Verify that all the CUDA:: targets and variables exist
 foreach (cuda_lib IN LISTS cuda_libs)
   if(NOT CUDA_${cuda_lib}_LIBRARY)
@@ -81,5 +80,24 @@ foreach (cuda_lib nvrtc nvToolsExt OpenCL)
   endif()
 endforeach()
 
+if(HAS_CUPTI)
+  set(cupti_libs )
+  if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 10.2)
+    list(APPEND cupti_libs cupti nvperf_target)
+  endif()
+  if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 11.3)
+    list(APPEND cupti_libs pcsamplingutil)
+  endif()
+
+  foreach (cuda_lib IN LISTS cupti_libs)
+    if(NOT CUDA_${cuda_lib}_LIBRARY)
+      message(FATAL_ERROR "expected CUDAToolkit variable CUDA_${cuda_lib}_LIBRARY not found")
+    endif()
+    if(NOT TARGET CUDA::${cuda_lib})
+      message(FATAL_ERROR "expected CUDAToolkit target CUDA::${cuda_lib} not found")
+    endif()
+  endforeach()
+endif()
+
 add_executable(CudaOnlyToolkit main.cu)
 target_link_libraries(CudaOnlyToolkit PRIVATE CUDA::toolkit)