rocp_sdk: enable delayed initialization

[dbarry9]

Pull Request Description

Revert the specific change in commit 258caab that removed delayed initialization from the rocp_sdk component.

This functionality is convenient for tools and applications which use PAPI internally.

These changes have been tested 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.

[Treece-Burgess]

Can you change sprintf to snprintf and then check the returned value, i.e.

int strLen = snprintf(_rocp_sdk_vector.cmp_info.disabled_reason, PAPI_HUGE_STR_LEN, "%s",  "Not initialized. Access component events to initialize it.");
if (strLen < 0 || strLen >= PAPI_HUGE_STR_LEN) {
    SUBDBG(...)
    return PAPI_EBUF;
}

[Treece-Burgess]

I assume the added "through PAPI_add_named_event()/PAPI_add_event()" was added due to the PAPI_EDELAY_INIT support and doing this workflow would initialize the component.

However, you can also call PAPI_enum_cmp_event(int *EventCode, PAPI_ENUM_FIRST, int cidx) to initialize the component. I would make that aware to users as well.

[Treece-Burgess]

From my testing, if I have the following workflow:

PAPI_library_init(PAPI_VER_CURRENT);
int cidx = API_get_component_index("rocp_sdk");

int EventCode = 0 | PAPI_NATIVE_MASK;
PAPI_enum_cmp_event(&EventCode, PAPI_ENUM_FIRST, cidx);

const PAPI_component_info_t *cmpinfo = PAPI_get_component_info(cidx);

The member variable of PAPI_component_info_t disabled_reason will still show "Not initialized. Access component events to initialize it." This should not be the case anymore as I have initialized rocp_sdk.

To resolve this you need to set disabled_reason to an empty string inside rocp_sdk_init_private (this is what I did in the cuda component see here).

[Treece-Burgess]

In rocp_sdk.c the following block of code exists:

int papi_errno = rocprofiler_sdk_init_pre();
    if (papi_errno != PAPI_OK) {
        _rocp_sdk_vector.cmp_info.initialized = 1;
        _rocp_sdk_vector.cmp_info.disabled = papi_errno;
        const char *err_string;
        rocprofiler_sdk_err_get_last(&err_string);
        snprintf(_rocp_sdk_vector.cmp_info.disabled_reason, PAPI_MAX_STR_LEN, "%s", err_string);
        return papi_errno;
    }

The call to rocprofiler_sdk_init_pre() should never fail as it just returns PAPI_OK, but for semantics the variable _rocp_sdk_vector.cmp_info.initialized should be set to 0 instead of 1 inside the conditional block. As if we did enter the block rocp_sdk would not be initialized.