Create an empty project:

$ xmake create -P test -l cuda
$ cd test
$ xmake

-- define target
target("cuda_console")
    set_kind("binary")
    add_files("src/*.cu")
    -- generate SASS code for SM architecture of current host
    add_cugencodes("native")
    -- generate PTX code for the virtual architecture to guarantee compatibility
    add_cugencodes("compute_30")

NOTE

Starting with v2.2.7, the default build will enable device-link. (see Separate Compilation and Linking of CUDA C++ Device Code) If you want to disable device-link, you can set it with set_policy("build.cuda.devlink", false).

NOTE

Device functions in cuda source files should be device-linked once and only once. On targets with kind binary or shared xmake will automatically perform the device-link which takes the static libraries they depend into account, while for static targets by default will not be device-linked. However, if the final binary or shared target do not contain any cuda files, the device-link stage could be missing, resulting in an undefined reference error. In this case the static target should be set add_values("cuda.build.devlink", true) manually.

Xmake will detect Cuda SDK automatically and we can also set the SDK directory (or SDK version for default installations) manually.

$ xmake f --cuda=/usr/local/cuda-9.1/
$ xmake f --cuda=9.1
$ xmake

If you want to known more information, you can see #158.