The propagation of dependencies to CUDA libraries

[wangkuiyi]

I wrote a CMakeLists.txt file like the following

nv_library(gpu_info SRCS gpu_info.cc)
cc_library(malloc SRCS malloc.cc DEPS gpu_info)
cc_test(malloc_test SRCS malloc_test.cc DEPS malloc)

The linking stage of malloc_test complains that it cannot find CUDA symbols like _cudaFree and _cudaMalloc.

I noticed that if I switch calls to cc_library and cc_test into nv_library and nv_test, the build would succeed. However, what would imply that users would have to write a lot of if(WITH_GPU) in their CMakeLists.txt files.

Is it alright to change bazel.cmake to have the behavior that if any dependent relies on CUDA, all dependees rely on CUDA too?

Thanks!