Compressed and Parallelized Structured Tensor Algebra
summary
DASTAC generates compressed linear indexes from polyhedral structure using the Barvinok algorithm, emits affine code and parallel loops via MLIR, and adds automatic data layout compression so tensors store only needed elements. The framework produces specialized kernels that match or beat expert hand-tuned code, reporting 0.16x–44.83x single-thread and 1.37x–243.78x multi-thread speedups, and supports transparent compress/decompress paths to interoperate with dense representations.
AI-written summary · about
chapters (8)
- 0:00Structured tensor algebra intro
- 1:51Structure encoding and inference
- 4:01Compressed tensor storage
- 6:30Parallelization and memory
- 7:31Performance comparisons
- 8:34Automatic tensor compression
- 9:39Takeaways and future work
- 10:22Q&A
related talks
15mCompiling and Compressing Structured Tensors
20mCompiling Structured Tensor Algebra
20mA Mechanized Algebra of Verified Data Structures for Optimizing Sparse Tensor Programs
32mCompilation of Dynamic Sparse Tensor Algebra
28mFunctional Sparse Tensor Compilation
20mExploiting Sparsity and Redundancy
17mA Sparse Iteration Space Transformation Framework for Sparse Tensor Algebra
25mFormat Abstraction for Sparse Tensor Algebra Compilers
