OpenTimer v2: A New Parallel Incremental Timing Analysis Engine

GPU-accelerated static timing analysis

“…The resulting task graph in terms of encapsulated function calls and data dependency is extremely vast and complex [6]. To alleviate the long runtime, recent work has explored parallel and distributed frameworks to speed up STA [3,4,7,8]. For example, Huang et al develop a timing analysis engine, OpenTimer, using a task dependency graph to represent timing propagation tasks and their precedence [4,6,7].…”

“…Due to the threading overhead and irregular computational patterns of STA, performance of CPU-based multi-threading usually saturates at around 8-16 threads [4,8]. To break the performance bottleneck, GPU acceleration for timing analysis is further explored [8,17].…”

“…Previous work in academic research and commercial tool development has proposed various parallel STA algorithms [3][4][5][6][7][8][9][10][11][12]. Each of these algorithms has their own pros and cons, but nearly all of them are architecturally constrained by the multithreaded paradigm on a manycore central processing units (CPUs) platform.…”

“…However, developing a highperformance STA algorithm that harnesses the power of CPU-GPU collaborative computing is an extremely challenging job. Computing a timing graph involves irregular memory access and significant diverse computational patterns, including graph-oriented computing, dynamic data structures, branch-and-bound, and recursion [4]. The resulting task graph in terms of encapsulated function calls and functional dependencies is vast and complex.…”

“…We develop GPU-efficient data structures and acceleration kernels to offload critical parts of STA computing to GPU. We have implemented our algorithms on OpenTimer, an open-source STA engine developed by Huang et al [4]. The core design philosophy of our algorithm is generally applicable and can be implemented for other STA frameworks.…”