Compiler for target machine independent optimization of data movement, ownership transfer and device control

摘要

A method and apparatus designed to expose the power of manipulating data transfer and ownership to the compiler has been given. The key ideas behind the XDP methodology are its separation of data transfer from local computation, its non-blocking semantics to allow overlapping of communication with computation, and its unified treatment of data and ownership transfer. In addition, XDP offers the compiler a convenient platform for doing optimizations involving data movement by providing mechanisms for delayed communication binding and generating generalized compute rules. Compile-time data and data ownership transfers are accomplished independent of the target machine architecture. The run-time symbol table given here to support XDP is implementable as an extension to most high-level compiler intermediate languages. The applicability of XDP is quite general, and is not restricted to the optimization of communication for distributed memory machines. For instance, it can be used to optimize data transfers across different levels of a memory hierarchy.