Résumé

Logic Synthesis From Partial Specifications (LSFPS) is the problem of finding the hardware implementation of a Boolean function from a partial knowledge of its care set. The elements missing from the specifications are named don't knows. The exact solution of LSFPS is the minimum size circuit of the corresponding problem in which the don't knows set is void. Hence, in addition to the traditional objective of size minimization, the goal is to maximize the test accuracy, i.e., the accuracy of the circuit when evaluated over a subset of the don't knows. This problem is relevant because efficient solutions can lead to hardware friendly machine learning models, not relying on black-box approaches. Indeed, LSFPS maps directly to the problem of the automatic generation of optimized topologies for Binarized Neural Networks. Furthermore, combining the exact solution with modern logic synthesis techniques would unlock unprecedented optimization capabilities. Previous works proved the effectiveness of approximate logic synthesis (ALS) for designing circuits with high test accuracy. Nonetheless, these methods sacrifice accuracy on the specifications, which banishes them from the legitimate candidates for LSFPS. In this paper, we propose accuracy recovery, a procedure to map an approximate version of the circuit to a new one that satisfies the exact functionality of the specifications. The proposed approach relies on an extension of a disjoint support decomposition algorithm. Relative experiments on the IWLS2020 benchmarks show that, on average, the addition of the designed decomposition to a synthesis flow reduces by 17.38% the number of gates and by 12.02% the depth. The usage of accuracy recovery, based on such a decomposition, yields a 95.73% accuracy in the binary MNIST problem, beating the state-of-the-art in ALS of 92.76%.

Détails