The ROCmPackage is not a build system but a helper package. Like CudaPackage, it provides standard variants, dependencies, and conflicts to facilitate building packages using GPUs though for AMD in this case.

You can find the source for this package (and suggestions for setting up your compilers.yaml and packages.yaml files) at


This package provides the following variants:

  • rocm

    This variant is used to enable/disable building with rocm. The default is disabled (or False).

  • amdgpu_target

    This variant supports the optional specification of the AMD GPU architecture. Valid values are the names of the GPUs (e.g., gfx701), which are maintained in the amdgpu_targets property.


This package defines basic rocm dependencies, including llvm and hip.


Conflicts are used to prevent builds with known bugs or issues. This package already requires that the amdgpu_target always be specified for rocm builds. It also defines a conflict that prevents builds with an amdgpu_target when rocm is disabled.

Refer to Conflicts for more information on package conflicts.


This package provides one custom helper method, which is used to build standard AMD hip compiler flags.


This built-in static method returns the appropriately formatted --amdgpu-target build option for hipcc.

This method must be explicitly called when you are creating the arguments for your build in order to use the values.


This helper package can be added to your package by adding it as a base class of your package. For example, you can add it to your CMakePackage-based package as follows:

 class MyRocmPackage(CMakePackage, ROCmPackage):
     # Ensure +rocm and amdgpu_targets are passed to dependencies
     depends_on('mydeppackage', when='+rocm')
     for val in ROCmPackage.amdgpu_targets:
         depends_on('mydeppackage amdgpu_target={0}'.format(val),

     def cmake_args(self):
         spec = self.spec
         args = []
         if '+rocm' in spec:
             # Set up the hip macros needed by the build
             rocm_archs = spec.variants['amdgpu_target'].value
             if 'none' not in rocm_archs:
             # Ensure build with hip is disabled
         return args

assuming only on the ENABLE_HIP, HIP_ROOT_DIR, and HIP_HIPCC_FLAGS macros are required to be set and the only dependency needing rocm options is mydeppackage. You will need to customize the flags as needed for your build.

This example also illustrates how to check for the rocm variant using self.spec and how to retrieve the amdgpu_target variant’s value using self.spec.variants['amdgpu_target'].value.

All five packages using ROCmPackage as of January 2021 also use the CudaPackage. So it is worth looking at those packages to get ideas for creating a package that can support both cuda and rocm.