accelerator/rocm: some minor fixes

[edgargabriel]

• memset the IPC handles to zero before calling hipIpcGetMemHandle (and the event counter part)
• get_buffer_id: fix the buffer_id argument passed to the hip call.

[wenduwan]

I just noticed that we didn't provide constructor/destructor

OBJ_CLASS_INSTANCE(
    opal_accelerator_stream_t,
    opal_object_t,
    NULL,
    NULL);

OBJ_CLASS_INSTANCE(
    opal_accelerator_event_t,
    opal_object_t,
    NULL,
    NULL);

OBJ_CLASS_INSTANCE(
    opal_accelerator_ipc_handle_t,
    opal_object_t,
    NULL,
    NULL);

OBJ_CLASS_INSTANCE(
    opal_accelerator_ipc_event_handle_t,
    opal_object_t,
    NULL,
    NULL);

Would it be better to reset the object in the constructor instead?

[edgargabriel]

@wenduwan I am not sure I follow your comment, can you please explain a bit more?

[wenduwan]

Sorry I got confused. I just realized hipIpcMemHandle_t is not an OPAL object.

My original comment was about the opal_accelerator_rocm_ipc_event_handle_t object and its parent class - I figured that we are not initializing the object to zero values because we did not define the constructor.

[wenduwan]

I cannot test the change but it looks reasonable.