Minor build improvements

* Allow specifying multiple GPU architectures. E.g.:
  cmake -DEIGEN_CUDA_COMPUTE_ARCH="60;70"
* Pass CUDA SDK path to clang. Without it it will default to /usr/local/cuda
which may not be the right location, if cmake was invoked with
-DCUDA_TOOLKIT_ROOT_DIR=/some/other/CUDA/path

1 files changed, 19 insertions, 4 deletions

diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index e8e1dc832..e6c757275 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -250,7 +250,11 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA)
     set(CUDA_NVCC_FLAGS "-ccbin ${CMAKE_C_COMPILER}" CACHE STRING "nvcc flags" FORCE)
   endif()
   if(EIGEN_TEST_CUDA_CLANG)
-   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 --cuda-gpu-arch=sm_${EIGEN_CUDA_COMPUTE_ARCH}")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
+    string(APPEND CMAKE_CXX_FLAGS " --cuda-path=${CUDA_TOOLKIT_ROOT_DIR}")
+    foreach(ARCH IN LISTS EIGEN_CUDA_COMPUTE_ARCH)
+        string(APPEND CMAKE_CXX_FLAGS " --cuda-gpu-arch=sm_${ARCH}")
+    endforeach()
   endif()
 
   set(EIGEN_CUDA_RELAXED_CONSTEXPR "--expt-relaxed-constexpr")
@@ -265,7 +269,11 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA)
     set(EIGEN_CUDA_CXX11_FLAG "")
   endif()
 
-  set(CUDA_NVCC_FLAGS  "${EIGEN_CUDA_CXX11_FLAG} ${EIGEN_CUDA_RELAXED_CONSTEXPR} -arch compute_${EIGEN_CUDA_COMPUTE_ARCH} -Xcudafe \"--display_error_number\" ${CUDA_NVCC_FLAGS}")
+  set(NVCC_ARCH_FLAGS)
+  foreach(ARCH IN LISTS EIGEN_CUDA_COMPUTE_ARCH)
+    string(APPEND NVCC_ARCH_FLAGS " -gencode arch=compute_${ARCH},code=sm_${ARCH}")
+  endforeach()
+  set(CUDA_NVCC_FLAGS  "${EIGEN_CUDA_CXX11_FLAG} ${EIGEN_CUDA_RELAXED_CONSTEXPR} -Xcudafe \"--display_error_number\" ${NVCC_ARCH_FLAGS} ${CUDA_NVCC_FLAGS}")
   cuda_include_directories("${CMAKE_CURRENT_BINARY_DIR}" "${CUDA_TOOLKIT_ROOT_DIR}/include")
   set(EIGEN_ADD_TEST_FILENAME_EXTENSION "cu")
 
@@ -276,8 +284,15 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA)
   ei_add_test(cxx11_tensor_cast_float16_gpu)
   ei_add_test(cxx11_tensor_scan_gpu)
 
+  set(EIGEN_CUDA_OLDEST_COMPUTE_ARCH 9999)
+  foreach(ARCH IN LISTS EIGEN_CUDA_COMPUTE_ARCH)
+    if(${ARCH} LESS ${EIGEN_CUDA_OLDEST_COMPUTE_ARCH})
+      set(EIGEN_CUDA_OLDEST_COMPUTE_ARCH ${ARCH})
+    endif()
+  endforeach()
+
   # Contractions require arch 3.0 or higher
-  if (${EIGEN_CUDA_COMPUTE_ARCH} GREATER 29)
+  if (${EIGEN_CUDA_OLDEST_COMPUTE_ARCH} GREATER 29)
     ei_add_test(cxx11_tensor_device)
     ei_add_test(cxx11_tensor_gpu)
     ei_add_test(cxx11_tensor_contract_gpu)
@@ -285,7 +300,7 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA)
   endif()
 
   # The random number generation code requires arch 3.5 or greater.
-  if (${EIGEN_CUDA_COMPUTE_ARCH} GREATER 34)
+  if (${EIGEN_CUDA_OLDEST_COMPUTE_ARCH} GREATER 34)
     ei_add_test(cxx11_tensor_random_gpu)
   endif()