On 28 August at 10:00 in MF14, Gijs Leemrijse will defend his MSc thesis titled “Towards relaxed memory semantics for the Autonomous Data Language”.
This work presents an alternative operational semantics for the Autonomous Data Language (AuDaLa) with relaxed memory consistency and incoherent memory. We show how the memory operations of our semantics can be safely mapped onto the NVIDIA PTX virtual ISA and demonstrate that our semantics performs faster than the original when executing AuDaLa programs on GPUs. We translate our operational semantics into an axiomatic memory consistency model and formally check, for a bounded program size, its correspondence with PTX’s memory consistency model using the Alloy model finder. We conclude by presenting AuDaLaC, our compiler targeting the CUDA platform, with which we explore several different strategies to compile AuDaLa programs. We demonstrate in several case studies that AuDaLa implementations can perform faster than sequential implementations.