Deep Learning Framework for Placement

Abstract

Nonlinear placement algorithms may need thousands of gradient descent iterations for convergence, which is computationally expensive. The literature has investigated multi-threading on CPUs using partitioning. The speedup ratio typically saturates at 5× with 2–6% quality degradation. GPU acceleration has been explored to accelerate a placement algorithm consisting of clustering, declustering, and nonlinear optimization. Around 15× speedup has been reported with less than 1% quality degradation. The speedup ratio is mostly limited by the sequential clustering and declustering steps in the algorithm. With the recent advances in nonlinear placement and deep learning, Lin et al. establish an analogy between the nonlinear placement problem and the neural network training problem. With such an analogy, deep learning frameworks/toolkits like PyTorch can be adopted to develop placement algorithms with native support to GPU acceleration and highly optimized kernel operators. Eventually, 30–40× speedup can be achieved for the nonlinear placement kernel. In this book chapter, we target this line of studies and survey how deep learning frameworks help placement development with high efficiency and performance. We also introduce how to customize kernel operators for further speedup as well as how to handle cutting-edge objectives and constraints with both conventional and machine learning techniques in such a framework.