Specification and verification of atomic operations in GPGPU programs

Abstract

We propose a specification and verification technique based on separation logic to reason about data race freedom and functional correctness of GPU kernels that use atomic operations as synchronisation mechanism. Our approach exploits the notion of resource invariant from Concurrent Separation Logic (CSL) to capture the behaviour of atomic operations. However, because of the different memory levels in the GPU architecture, we adapt this notion of resource invariant to these memory levels, i.e., group resource invariants capture the behaviour of atomic operations that access locations in local memory, while kernel resource invariants capture the behaviour of atomic operations that access locations in global memory. We show soundness of our approach and we provide tool support that enables us to verify kernels from standard benchmarks suites.