2026.01.22 initial rocm presets

[dbarry9]

Pull Request Description

This pull request adds the following to the rocm component:

• functions to facilitate ROCm presets, and
• presets for FMA instructions of various precisions.

These changes have been tested on the AMD MI250X architecture.

Here is the relevant output from papi_avail on the Frontier supercomputer:

================================================================================
  PAPI Component Preset Events
================================================================================
       Name           Code    Avail Deriv Description (Note)
PAPI_ROCM_FP16_FMA 0x80000080  Yes   No   ROCM Half precision (FP16) FMA instructions
    :device=0
        Mandatory device qualifier [0,1,2,3,4,5,6,7]
--------------------------------------------------------------------------------
PAPI_ROCM_BF16_FMA 0x80000081  No    No   ROCM Half precision (BF16) FMA instructions
--------------------------------------------------------------------------------
PAPI_ROCM_FP32_FMA 0x80000082  Yes   No   ROCM Single precision (FP32) FMA instructions
    :device=0
        Mandatory device qualifier [0,1,2,3,4,5,6,7]
--------------------------------------------------------------------------------
PAPI_ROCM_FP64_FMA 0x80000083  Yes   No   ROCM Double precision (FP64) FMA instructions
    :device=0
        Mandatory device qualifier [0,1,2,3,4,5,6,7]
--------------------------------------------------------------------------------
PAPI_ROCM_FP_FMA   0x80000084  Yes   Yes  ROCM floating-point FMA instructions
    :device=0
        Mandatory device qualifier [0,1,2,3,4,5,6,7]
--------------------------------------------------------------------------------
PAPI_ROCM_FP8_OPS  0x80000085  No    No   ROCM 8-bit precision floating-point operations
================================================================================

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]

Q: In papi_rocm_presets.h you have 6 listed events, but here only 4 are defined. Are the other two for a separate architecture?

[Treece-Burgess]

Q: It appears that none of the events listed will need an instance qualifier appended to them. For the cuda component, you show the stats qualifier as well in the papi_avail output as those events do need the stats qualifier.

Say for a future case you add an event which does require an instance value, is that already handled with this update or would more work need to be done?

[Treece-Burgess]

nit: I would personally be more descriptive with the variable naming not just here, but in other instances as well.

name should correspond to the arch_name, but it is not immediately clear this is what we actually mean. Long term, I believe it would help others that come along and work on the component that may not be as familiar with it.

[Treece-Burgess]

From the docs HSA_AGENT_INFO_NAME is a null terminated char[64]. The variable you are passing is arch_name[PAPI_2MAX_STR_LEN] ([256]), I would swap to PAPI_MIN_STR_LEN.

[Treece-Burgess]

nit: Should ROCm not be AMD here?

[Treece-Burgess]

Q: From my understanding, hsa_iterate_agents will enumerate over the available agents and invoke the callback you have defined. Therefore, if we have 7 AMD devices this will iterate over all 7 calling your callback function. What if there are multiple mixed devices, i.e. MI250x and MI300A? From my testing you would only capture the very last one.

[Treece-Burgess]

nit: I would change the comments from /* to // as then it becomes easier to comment blocks of code out later if you are debugging.

However, you can use #if 0 #endif.

This is something we should probably discuss internally as a team though.

[Treece-Burgess]

When I use PAPI_get_event_info with either the preset rocm define or the preset rocm event code the member variable component_index from the structure PAPI_event_info_t shows a value of 0. This should not be the case for the rocm component.

[Treece-Burgess]

On Frontier, when I call PAPI_enum_event with the modifier PAPI_PRESET_ENUM_AVAIL I do not see PAPI_ROCM_FP16_FMA output even though this preset is available on the architecture. When I used the modifier PAPI_ENUM_EVENTS it did appear in the output.

[Treece-Burgess]

nit: If you add a SUBDBG to this I could see the nested approach; however, one option would be to:

papi_errno = rocm_init_comp_presets();
if (papi_errno == PAPI_ENOIMPL) {
    papi_errno = PAPI_OK;
}

return papi_errno;