Multi-domain clock skew scheduling-aware register placement to optimize clock distribution network

Abstract

Abstract. Multi-domain clock skew scheduling is a cost effective technique for performance improvement. However, the required wire length and area overhead due to phase shifters for realizing such clock scheduler may be considerable if registers are placed without considering assigned skews. Focusing on this issue, in this paper, we propose a skew scheduling-aware register placement algorithm that enables clock tree optimization by considering domains assigned to registers in placement. Our experimental results show that the proposed approach remarkably decreases clock wire length and clock network power consumption at the cost of a slight increase in total wire length.

Register relocation to optimize clock network for multi-domain clock skew scheduling

“…It is observed that our register relocation achieves a significant cost reduction in both BST and RT clock network structures, which is up to 50 percents in clock wire length, much more than [8]. However, the increase on total wire length is NOT negligible in the trade-off, and so the dynamic power estimation on each circuit is shown in Fig.3.…”

“…Second, since the partitioning results of all the cells including registers might be refined, the amount of affected cells might be large and their displacements might be huge to have different timing in data paths, where the assumption for clock skew scheduling was violated. Therefore, after register location refinement, either setup or hold constraints in some timing paths might be violated, which is the major drawback in [8].…”

“…MohammadZadeh et al [8] proposed a multi-domain clock skew scheduling-aware register placement algorithm recently. They presented a clock distribution network construction with few clock skew domains, and tried to place registers into a "legal" partition whose specified skew domain matched.…”

“…Registers could be relocated within their permissible displacement range to guarantee no timing constraint violations, while minimizing the distance to the nearby registers in the same skew domain. Compared to [8], our proposed method improves physical locality of registers, which does not depend on some specified clock network structures but could be exploited by general clock constructions. This paper is organized as follows.…”

“…In the literature, there are two ways to follow: physical location-aware clock skew scheduling [4,5,7] and clock skew scheduling-aware placement [8].…”