MIREncoder: Multi-modal IR-based Pretrained Embeddings for Performance Optimizations
This addresses the need for efficient, reusable pre-trained models in high-performance computing to optimize compilable code, though it is incremental as it builds on existing IR-based methods.
The paper tackles the problem of improving performance optimization for parallel workloads by proposing MIREncoder, a multi-modal IR-based auto-encoder that generates pre-trained embeddings for downstream tasks, and it shows that this approach outperforms state-of-the-art methods while reducing overhead.
One of the primary areas of interest in High Performance Computing is the improvement of performance of parallel workloads. Nowadays, compilable source code-based optimization tasks that employ deep learning often exploit LLVM Intermediate Representations (IRs) for extracting features from source code. Most such works target specific tasks, or are designed with a pre-defined set of heuristics. So far, pre-trained models are rare in this domain, but the possibilities have been widely discussed. Especially approaches mimicking large-language models (LLMs) have been proposed. But these have prohibitively large training costs. In this paper, we propose MIREncoder, a M}ulti-modal IR-based Auto-Encoder that can be pre-trained to generate a learned embedding space to be used for downstream tasks by machine learning-based approaches. A multi-modal approach enables us to better extract features from compilable programs. It allows us to better model code syntax, semantics and structure. For code-based performance optimizations, these features are very important while making optimization decisions. A pre-trained model/embedding implicitly enables the usage of transfer learning, and helps move away from task-specific trained models. Additionally, a pre-trained model used for downstream performance optimization should itself have reduced overhead, and be easily usable. These considerations have led us to propose a modeling approach that i) understands code semantics and structure, ii) enables use of transfer learning, and iii) is small and simple enough to be easily re-purposed or reused even with low resource availability. Our evaluations will show that our proposed approach can outperform the state of the art while reducing overhead.