1.5.1  OMPT

The OMPT interface, which is intended for first-party tools, provides the following:

• A mechanism to initialize a first-party tool;
• Routines that enable a tool to determine the capabilities of an OpenMP implementation;
• Routines that enable a tool to examine OpenMP state information associated with a thread;
• Mechanisms that enable a tool to map implementation-level calling contexts back to their source-level representations;
• A callback interface that enables a tool to receive notification of OpenMP events;
• A tracing interface that enables a tool to trace activity on OpenMP target devices; and
• A runtime library routine that an application can use to control a tool.

OpenMP implementations may differ with respect to the thread states that they support, the mutual exclusion implementations that they employ, and the OpenMP events for which tool callbacks are invoked. For some OpenMP events, OpenMP implementations must guarantee that a registered callback will be invoked for each occurrence of the event. For other OpenMP events, OpenMP implementations are permitted to invoke a registered callback for some or no occurrences of the event; for such OpenMP events, however, OpenMP implementations are encouraged to invoke tool callbacks on as many occurrences of the event as is practical. Section 19.2.4 specifies the subset of OMPT callbacks that an OpenMP implementation must support for a minimal implementation of the OMPT interface.

With the exception of the omp_control_tool runtime library routine for tool control, all other routines in the OMPT interface are intended for use only by tools and are not visible to applications. For that reason, a Fortran binding is provided only for omp_control_tool; all other OMPT functionality is described with C syntax only.