resolve contentious AMD components

[dbarry9]

Pull Request Description

• framework: PAPI_DISABLE_COMPONENTS env var
• amd comps: contentious components in same config

This pull request resolves issues #416 and #478.

ROCm version >= 6.3.2 being the "cutoff" for making rocp_sdk active by default over rocm was chosen due to known bugs in the ROCProfiler SDK in prior releases.

These changes have been tested using ROCm 7.0.2 on the Frontier supercomputer, which contains the AMD MI250X architecture.

Author Checklist

• Description
  Why this PR exists. Reference all relevant information, including background, issues, test failures, etc
• Commits
  Commits are self contained and only do one thing
  Commits have a header of the form: module: short description
  Commits have a body (whenever relevant) containing a detailed description of the addressed problem and its solution
• Tests
  The PR needs to pass all the tests

[Treece-Burgess]

I am reviewing this PR.

[sbkathpe]

Yes, I believe this follows the spirit and intent of this env variable.

I would do a free(penv) though since this allocated space is no longer needed.

[sbkathpe]

Just to make sure, the "PAPI_DISABLE_COMPONENTS" code should NOT be compile-time dependent on "#if defined(DEFAULT_TO_ROCP_SDK)" but be unconditionally supported.
It is hard for me to tell if this condition is part of the final code or not.

[dbarry9]

Just to make sure, the "PAPI_DISABLE_COMPONENTS" code should NOT be compile-time dependent on "#if defined(DEFAULT_TO_ROCP_SDK)" but be unconditionally supported. It is hard for me to tell if this condition is part of the final code or not.

The code that parses PAPI_DISABLE_COMPONENTS and disables the components unconditionally has been added to PAPI_library_init in papi.c.

However, there is other code, which is contained within the #if defined in the rocm, rocp_sdk, rocm_smi, and amd_smi components. This code will be compile-time dependent due to the following conflicts:

• rocm conflicts with rocp_sdk (if both active)
• rocm_smi conflicts with amd_smi (if both active)

Example: Suppose the rocm and rocp_sdk components are both configured with a ROCm 7.0.1 module loaded. PAPI's configure chooses rocp_sdk to be active by default due to the ROCm version, and at the same time, PAPI's configure disables rocm by default using the #if defined(DEFAULT_TO_ROCP_SDK) inside of the rocm component code. However, if the user really wants the rocm component to be active, then they can set PAPI_DISABLE_COMPONENTS=rocp_sdk. Following this variable being set, in order for the rocm component to now be active, it needs to know that the rocp_sdk component is disabled (disabled unconditionally by PAPI_library_init). The rocm component knows the rocp_sdk component is disabled by parsing PAPI_DISABLE_COMPONENTS (very similarly to how PAPI_library_init does it), but the rocm component only checks whether rocp_sdk is listed in PAPI_DISABLE_COMPONENTS.

Summary: PAPI_DISABLE_COMPONENTS=x,y,z will indeed disable the components x,y, and z by PAPI_library_init parsing PAPI_DISABLE_COMPONENTS. In addition to this, PAPI_DISABLE_COMPONENTS is also parsed independently by certain components to resolve conflicts between specific components, but the actual disabling of components takes place unconditionally in PAPI_library_init.