Abstract

Graphics processing units (GPUs) are highly parallel processors that are now commonly used in the acceleration of a wide range of computationally intensive tasks. The focus of this talk is on verifying race-freedom of data-dependent GPU kernels, whose data or control flow are dependent on the input of the program. Kernels in this small but important class are difficult to verify because they require reasoning about shared state manipulated by many parallel threads. Existing verification techniques for GPU kernels achieve soundness and scalability by using a two-thread reduction and making the contents of the shared state nondeterministic each time threads synchronise at a barrier, to account for all possible thread interactions. This abstraction is too coarse for data-dependent kernels.

This talk will present barrier invariants, a new abstraction technique that allows key properties about the shared state of a kernel to be preserved across barriers during formal reasoning. By integrating this technique into the GPU Verify tool, we have verified kernels that were previously beyond existing verification techniques for GPU kernels.