Tuning Approximate Computations with Constraint-Based Type Inference

Unreliable hardware can lead to great gains in energy efficiency, but it can be difficult to reason about how unreliable each operation in a computation may feasibly be. To make approximate computing viable, we need tools that can help programmers derive precision–energy trade-offs for individual fine-grained operations while reasoning about the collective impact on the result quality. We formulate the problem of precision tuning as type inference over a system of types parameterized on their accuracy. Our type inference system generates numerical constraints and uses an SMT solver to produce parameters for unspecified types. Programmers can choose to provide explicit types where they make sense and depend on inference where the appropriate accuracy parameter is unclear. Remaining research challenges are discussed.