SMT-Based Translation Validation for Machine Learning Compiler

Abstract Machine learning compilers are large software containing complex transformations for deep models, and any buggy transformation may cause a crash or silently bring regression to the prediction accuracy performance. This paper proposes an SMT-based translation validation framework Multi-Level IR (MLIR), compiler used by many compilers. It SMT encoding tailored that is over-approximation of FP arithmetic reduction operations. performs abstraction refinement if fails. We also propose new approach properties reductions in SMT. found mismatches between specification implementation MLIR, validated high-level , with proper splitting.