Authors:
Abstract: Mainstream deep learning frameworks are commonly implemented by invoking underlying high performance tensor libraries on various architectures. However, as these libraries provide increasingly complex semantics including operator fusions, in-place operations, and various memory layouts, the gap between mathematical deep learning models and the underlying libraries becomes larger. In this paper, inspired by the classic problem of Instruction Selection, we design a theorem solver guided exhausted search algorithm to select functions for complex tensor computations. Preliminary results with some micro-benchmarks and a real model show that our approach can outperform both Tensorflow and Tensor Comprehensions at run time.
Best Poster Finalist (BP): no
Poster: PDF
Poster summary: PDF
Back to Poster Archive Listing