C2TACO: Lifting Tensor Code to TACO
summary
C2TACO synthesizes TACO tensor-DSL programs from C by using enumerative program synthesis guided by automatically generated IO examples and static code features (tensor order, program length, operators). The tool synthesizes correct TACO for 95% of benchmarks from a 61-program real-world suite, outperforms a Transformer NMT baseline, TF-Coder and ChatGPT on coverage, and produces portable code that yields 1.79× CPU and 24.11× GPU speedups on average; current limitations include sparse-tensor and high-order indexing support.
AI-written summary · about
chapters (10)
- 0:00Introduction to C2TACO and machine learning tensor algebra
- 2:29Lifting legacy tensor code challenges
- 4:19Synthesis from input-output
- 6:50Bottom-up enumerative search
- 11:09Comparison with NMT and others
- 13:41Evaluation on benchmarks
- 16:15Error analysis and limits
- 18:27Performance gains after lifting
- 19:37Conclusion and future work
- 20:40Q&A
related talks
19mGuided Tensor Lifting
17mA Sparse Iteration Space Transformation Framework for Sparse Tensor Algebra
25mFormat Abstraction for Sparse Tensor Algebra Compilers
19mMosaic: An Interoperable Compiler for Tensor Algebra
24mVerified Tensor-Program Optimization Via High-Level Scheduling Rewrites
15mTACCL: Guiding Collective Algorithm Synthesis using Communication Sketches
15mCompilation of Sparse Array Programming Models
20mCompiling Structured Tensor Algebra