ROCm / HIP

HIP: C++ Heterogeneous-Compute Interface for Portability
https://rocmdocs.amd.com/projects/HIP/
MIT License
3.77k stars 539 forks source link

Question about install HIP to run on nvidia gpu #3665

Open huccpp opened 2 weeks ago

huccpp commented 2 weeks ago

I try to install HIP to run on nvidia gpu. I looked at the instruction here: https://rocm.docs.amd.com/projects/HIP/en/latest/install/install.html I am on ubuntu 20.04. Earlier I have installed nvidia hpc sdk, but the path of cuda it contains is not at /usr/local/cuda as written (it is at /opt/nvidia/.../cuda). How to specify the path for the apt install? And in this case, do I need to do the step sudo apt-get install ubuntu-drivers-common && sudo ubuntu-drivers autoinstall sudo reboot ? Thanks a lot.

harkgill-amd commented 1 week ago

Hi @huccpp, I believe your current installation at /opt/nvidia is due to cuda being installed from the Ubuntu repositories. If you install the toolkit from the official website, the installation should be done under /usr/local.

huccpp commented 1 week ago

Hello @harkgill-amd, Thanks for your reply. Earlier I installed the nvidia hpc sdk from https://developer.nvidia.com/hpc-sdk-downloads which includes things in cuda toolkits, but inside /opt/nvidia/ . The problem is that I wonder whether it is possible to install HIP without installing cuda toolkits as its contents are already in my nvidia hpc sdk installation (also to avoid possible conflicts). Thanks. huccpp

harkgill-amd commented 1 week ago

I know of users working around the official cuda toolkit dependency by adding a | nvidia-cuda-toolkit (>= 7.5) in the hip-runtime-nvidia .deb file, see https://github.com/ROCm/HIP/issues/820#issuecomment-910063554. Please note that this is not officially supported and may cause errors for dependencies that look for the cuda installation at /usr/local/cuda.

Not sure if the installation from the HPC SDK link satisfies either of these package requirements but it would be worth giving a try. If you could also share any errors encountered during the installation, that would be helpful in further investigating.

EDIT: You can also set the default CUDA installation path with the CUDA_PATH environment variable (ref).