CustomHalide – A new plugin of clang for loop optimization

Nowadays, Polyhedral and Halide techniques optimize complex nested loops in deep learning frameworks’ basic operator compilation process. However, faced with many algorithm platforms, a single framework cannot be used for deployment purposes. In some scenarios, the complex deployment environment of the framework is also an important consideration. At this point, users are more interested in whether there is an optimization tool chain that automatically optimizes the simple described algorithm directly to solve the needs of customized deployment while bringing performance benefits. This paper presents an automatic optimization toolchain CustomHalide based on Halide automatic scheduling algorithm and Clang AST. The core of CustomHaide is the tool Loop-convert, which realizes loop parameters’ extraction from source code, converting parameters, calling the Halide auto-scheduler, integration and compilation of optimized code. Loop-convert provides a fully automated flow from C/C++ source code to optimized programs. After testing, CustomHalide has a better effect on identifying and extracting source code than Polly. At the same time, the average performance of CustomHalide is 3% higher than that of Halide both on CPU and GPU in the case of known bound information. The average compilation time of CustomHalide is one third of TVM with the same end-to-end performance. Finally, through the bound setting strategy, CustomHalide supports loop optimization with unknown bound information, and its average performance is about 50% higher than that of the program optimized by clang-O3.