Compiling Structured Tensor Algebra
summary
Framework compiles tensor algebra while symbolically computing sparsity and redundancy at compile time, using an intermediate language STUR and a compiler called StructTensor that separates Boolean index logic from real arithmetic and emits C++ code. Structure inference follows formal, sound rules and reduces iteration domains (e.g., covariance / polynomial regression kernels) so benchmarks show 1–4 orders of magnitude speedups versus TensorFlow/PyTorch/TACO on tested kernels. Inputs must currently supply sparsity/redundancy patterns; analysis time is small compared to compilation.
AI-written summary · about
chapters (9)
- 0:00Structured tensor algebra intro
- 1:25Why structured tensor computations
- 3:55Symbolic tensor structures
- 6:46Redundancy in matrices
- 8:12STAR framework architecture
- 9:15Code generation for regression
- 13:14Performance results
- 15:04Conclusion and future work
- 15:49Question and answer session
related talks
15mCompiling and Compressing Structured Tensors
32mCompilation of Dynamic Sparse Tensor Algebra
28mFunctional Sparse Tensor Compilation
14mCompressed and Parallelized Structured Tensor Algebra
19mCompiler Support for Sparse Tensor Convolutions
15mCompilation of Sparse Array Programming Models
17mA Sparse Iteration Space Transformation Framework for Sparse Tensor Algebra
19mOptimizations and abstractions for sparse machine learning