A framework for fast test generation at the RTL
Kelson Gent, Akash Deepak Agrawal, Michael S. Hsiao · 2017
We present a framework for high quality functional test generation at the RTL. The method utilizes static learning to derive cross cycle relationships which helps the search algorithm make inferences about the potential execution path in the next cycle. Additionally, dynamic taint analysis is adapted to model behavioral propagation. Finally, an ant colony optimization swarm intelligence algorithm is combined with an operator level behavioral coverage metric. We show that the method can be used to generate high quality functional test patterns at the RTL. By leveraging RTL code for test generation, we can generate test vectors with comparable levels of coverage up to 100× times faster than gate level ATPG and more than 10× faster than mixed level generation techniques. Additionally, without gate level information, we are able to approach the highest levels of coverage seen by previous techniques and in the case of circuits with deep narrow states, we are able to greatly exceed gate level ATPG performance including generating patterns for b12word and FABSCALAR, which prior ATPG techniques are unable to sufficiently handle.