MUSCAT: MUS-based Circuit Approximation Technique

Two-Level and Multilevel Approximate Logic Synthesis

“…Furthermore, Witschen at al. [49] proposes an ALS approach that generates valid-by-construction approximate circuits. In this work, the approximate circuit is generated in a single algorithm pass, different from other multilevel ALS approaches that iteratively select local approximations which do not exceed the error constraint.…”

“…ALS method Data structure Error metric [20,22] Boolean network ER, WCE [21] Boolean network WCE, WCRE [23][24][25] Boolean network ER, MAE [32,33] Boolean network ER [36] Boolean network BER [37,47] Boolean network ER, MRE [40,41] Boolean network WBF, MAE [42,49] Boolean network WCE [26] AIG WCE [38] AIG WCE, MSE [39,46] AIG ER [43] AIG WBF, ER, WCE [44] AIG ER, MAE [27,45] AIG ER, MAE, MRE [48] AIG ER, MAE, WCE [28] BDD ER, MAE, WCE [29] BDD ER, MAE, MRE [30,31] BDD BER [34] LUT ER [35] LUT BER, ER circuits in a sequential design, also presenting a particular error calculation technique for this scenario.…”

“…The approximation strategy of multilevel ALS approaches is more challenging to define. The most recent works proposed using more complex techniques, such as genetic algorithms [35,36], machine learning [47], SAT [49], and exact synthesis [48]. As error controlling corresponds to an expensive part of the ALS flow, some works also focus on optimizing the error estimation by using new estimation techniques [24,60], selecting multiple changes in one pass [26], and even approximating the whole circuit in a validby-construction form [49].…”

“…The most recent works proposed using more complex techniques, such as genetic algorithms [35,36], machine learning [47], SAT [49], and exact synthesis [48]. As error controlling corresponds to an expensive part of the ALS flow, some works also focus on optimizing the error estimation by using new estimation techniques [24,60], selecting multiple changes in one pass [26], and even approximating the whole circuit in a validby-construction form [49]. In that way, we believe that the main challenge is selecting a good approximation candidate for a node or a subcircuit that optimizes locally the circuit, introducing low error but without dedicating too much time evaluating the impact on the error metric of each possible candidate.…”