After more than 30 years of research, there is a solid understanding of the consistency guarantees given by CPU systems. Unfortunately, the same is not yet true for GPUs. The growing popularity of general purpose GPU programming has been a call for action which industry players like Nvidia and Khronos have answered by formalizing their Ptx and Vulkan consistency models. These models give precise answers to questions about program's correctness. However, interpreting them still requires a level of expertise that escapes most developers, and the current tool support is insufficient.

We present Dartagnan , a bounded model checker (BMC) for concurrent programs under weak memory models. Its distinguishing feature is that the memory model is not implemented inside the tool but taken as part of the input. Dartagnan reads CAT , the standard language for memory models, which allows to define x86/ TSO , ARM v7, ARM v8, Power , C/C++, and Linux kernel concurrency primitives. BMC with memory models as inputs is challenging. One has to encode into SMT not only the program but also its semantics as defined by the memory model. What makes Dartagnan scale is its relation analysis, a novel static analysis that significantly reduces the size of the encoding. Dartagnan matches or even exceeds the performance of the model-specific verification tools Nidhugg and CBMC , as well as the performance of Herd , a CAT -compatible litmus testing tool. Compared to the unoptimized encoding, the speed-up is often more than two orders of magnitude.