Are logic synthesis tools robust?

Abstract

A systematic investigation is presented about the robustness of logic synthesis tools to equivalence-preserving transformations of the input Verilog file. We have developed a framework that: 1) parses Verilog behavioral models into an abstract syntax tree; 2) generates random equivalence-preserving transformations on the syntax tree, and; 3) writes the transformed design back in Verilog format. The original and the transformed Verilog descriptions are then checked for equivalence and synthesized. Results show that average (peak) improvements in area of 2.5%(11%) and length of the critical path of 4%(13%) are achievable. Indeed these figures are comparable to recent advancements in logic synthesis ([17] [8] achieve 4.9%(23%) 5%(24%) improvements area-wise, respectively), signaling a relevant lack of robustness in synthesis tools. This lack of robustness suggests that new synthesis algorithms should be evaluated by measuring the average improvement on several transformed files to assess their real contributions to the quality of the results.

Sources of bias in EDA tools and its influence

“…These heuristics can be influenced by many external aspects and results of different quality can be produced under different circumstances. In this paper we will show an experimental evaluation of several tools, to illustrate that the situation is much more severe than previously reported [2], [3].…”

PBO-Based Test Compression

“…For this reason, a single measurement for one benchmark circuit and one algorithm configuration cannot bring reliable results [29], [30]. Therefore, except of the final experiments, we have conducted 1,000 measurements for each tested circuit and configuration, with the above-mentioned aspects set randomly.…”

On don’t cares in test compression

“…The result quality of most EDA processes based on local heuristics depends on random aspects coming from the input description [41,42]. This issue is discussed and documented in Section 4.1.…”

“…More importantly, the two distributions have different mean values, advantageously to the CPDCI. CPDCI also has smaller standard deviation, thus randomness plays a smaller role here, making CPDCI more robust than the standard SATCompress [41]. Nevertheless, the influence of randomness is still crucial (even though reduced in the CPDCI case), and worse results can be obtained by CPDCI accidentally, see Fig.…”

“…Despite of thorough investigations, we were not able to find a way of determining a ''good'' ordering, e.g., by preferring hard-to-detect faults. Actually, similar influences occur in almost all local search heuristics [41,42] and the problem is difficult to be fought. We reduce the influence of faults ordering by making the algorithm more robust and more efficient (see Section 4.1).…”