CUDA (nvcc) cross compiler

[dat-tran-matternet]

I'm trying to get a cross compiler for CUDA nvcc ( cross compile cuda code for jetson on my x86 host). I tried the following in my image but don't see the cross compiiler for nvcc. thanks

TOOLCHAIN_TARGET_TASK_append_tegra = "\
    cuda-compiler \
    cuda-libraries \
    cuda-nvcc \
    cuda-nvcc-dev \
    opencv-dev \
 "

I build the Target SDK like so: bitbake -c populate_sdk $IMAGE

[ichergui]

Hey

Use TOOLCHAIN_HOST_TASK instead of TOOLCHAIN_TARGET_TASK

• Add the following line in your conf/local.conf file

TOOLCHAIN_HOST_TASK_append = " nativesdk-packagegroup-cuda-sdk-host"

[madisongh]

If you use cmake's CUDA language support for your project, the SDK environment should take care of this for you. Otherwise, take a look at the environment variables we set as part of the SDK setup, once you get the host-side packagegroup into your SDK -- I believe the CUDAFLAGS variable (which you can reference in a makefile) woul dhave the necessary settings.

[dat-tran-matternet]

Thanks for your help guys. I was able to get it working with running:
nvcc main.cu $CUDAFLAGS -ccbin=${TARGET_PREFIX}g++ -O3 -o cuda-test

I plan to use CMake in the future but want to make sure my setup is correct with a simple hello world example.
Thanks again.

[madisongh]

Ah yeah, you do need to pass -ccbin as well. The environment variable CUDAHOSTCXX can be used as the value for that.

[dat-tran-matternet]

You guys are awesome! Here is the updated command with your suggestions:

nvcc main.cu $CUDAFLAGS -ccbin=${CUDAHOSTCXX} -O3 -o cuda-test

[ichergui]

I'm happy that all works fine for you.