Add CUDA complex sqrt.

This is to support scalar `sqrt` of complex numbers `std::complex<T>` on
device, requested by Tensorflow folks.

Technically `std::complex` is not supported by NVCC on device
(though it is by clang), so the default `sqrt(std::complex<T>)` function only
works on the host. Here we create an overload to add back the
functionality.

Also modified the CMake file to add `--relaxed-constexpr` (or
equivalent) flag for NVCC to allow calling constexpr functions from
device functions, and added support for specifying compute architecture for
NVCC (was already available for clang).

test/CMakeLists.txt

@@ -395,6 +395,12 @@ find_package(CUDA 5.0)
 if(CUDA_FOUND)
   
   set(CUDA_PROPAGATE_HOST_FLAGS OFF)
+  
+  set(EIGEN_CUDA_RELAXED_CONSTEXPR "--expt-relaxed-constexpr")
+  if (${CUDA_VERSION} STREQUAL "7.0")
+    set(EIGEN_CUDA_RELAXED_CONSTEXPR "--relaxed-constexpr")
+  endif()
+  
   if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang") 
     set(CUDA_NVCC_FLAGS "-ccbin ${CMAKE_C_COMPILER}" CACHE STRING "nvcc flags" FORCE)
   endif()
@@ -404,7 +410,12 @@ if(CUDA_FOUND)
     foreach(GPU IN LISTS EIGEN_CUDA_COMPUTE_ARCH)
       string(APPEND CMAKE_CXX_FLAGS " --cuda-gpu-arch=sm_${GPU}")
     endforeach()
+  else()
+    foreach(GPU IN LISTS EIGEN_CUDA_COMPUTE_ARCH)
+      string(APPEND CUDA_NVCC_FLAGS " -gencode arch=compute_${GPU},code=sm_${GPU}")
+    endforeach()
   endif()
+  string(APPEND CUDA_NVCC_FLAGS " ${EIGEN_CUDA_RELAXED_CONSTEXPR}")
   set(EIGEN_ADD_TEST_FILENAME_EXTENSION  "cu")
   
   ei_add_test(gpu_basic)