Homomorphic encryption (HE) is an encryption scheme that provides arithmetic operations on the encrypted data without doing decryption. For Ring-based HE, an encryption scheme that uses arithmetic operations on a polynomial ring as building blocks, performance improvement of unit HE operations has been achieved by two kinds of efforts. The first one is through accelerating the building blocks, polynomial operations. However, it does not facilitate optimizations across polynomial operations such as fusing two polynomial operations. The second one is implementing highly optimized HE operations in an amalgamated manner. The written codes have superior performance, but they are hard to maintain. To resolve these challenges, we propose HEaaN.MLIR, a compiler that performs optimizations across polynomial operations. Also, we propose Poly and ModArith, compiler intermediate representations (IRs) for integer polynomial arithmetic and modulus arithmetic on integer arrays. HEaaN.MLIR has compiler optimizations that are motivated by manual optimizations that HE developers do. These include optimizing modular arithmetic operations, fusing loops, and vectorizing integer arithmetic instructions. HEaaN.MLIR can parse a program consisting of the Poly and ModArith instructions and generate a high-performance, multithreaded machine code for a CPU. Our experiment shows that the compiled operations outperform heavily optimized open-source and commercial HE libraries by up to 3.06x in a single thread and 4.55x in multiple threads.

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