How do I add HIP support to existing cmake project ?

[boxerab]

I have cmake project - would like to add some HIP files to it. So, I guess the compiler needs
to change for the project ? What is the best way of doing this ?

[bensander]

Easiest approach would be to change the compiler to "hipcc" - this will automatically detect the plaform and use the right compiler/includes/define for either ROCm or CUDA. With this approach, code will be compiled with HIPCC. Another approach sometimes used for large projects is to partition the files which define or call GPU kernels and compile only those files with hipcc, and the rest with a different host compiler such as g++.

[boxerab]

Thanks, will give that a try.

[boxerab]

Solution from cmake user group is to include a find_package for hcc, because it doesn't lie on a default path:

find_package(hcc PATHS /opt/rocm/hcc)

and then set the compiler in cmake:

CXX=hcc

Perhaps this can be added to the documentation somewhere ?

[pfultz2]

Solution from cmake user group is to include a find_package for hcc, because it doesn't lie on a default path:

Thats just for the hcc runtime. You can do find_package(hip PATHS /opt/rocm) for the HIP runtime, but its only supported for hcc right now. If you want to support both cuda and hcc, you can just set the compiler to the hipcc wrapper with:

CXX=/opt/rocm/bin/hipcc cmake ..

[ragerdl]

Personally, it'd be nice to see a section in https://rocm-documentation.readthedocs.io/en/latest/Programming_Guides/HIP-porting-guide.html on how to port a CUDA-oriented CMakeLists.txt to be HCC-centric. Of course, many things would be nice....

[gargrahul]

@ragerdl please check https://github.com/ROCm-Developer-Tools/HIP/tree/master/samples/2_Cookbook/12_cmake_hip_add_executable

We can refer this from porting guide.