Format Abstraction for Sparse Tensor Algebra Compilers
summary
Paper demonstrates an abstraction that expresses tensor formats as composable per-dimension level formats and a plugin-style code generator that emits concrete kernels by inlining format-provided implementations. Prototype extends the taco compiler with six level implementations (dense, CSR/CSF, COO, DIA, ELL, HASH), generates code for arbitrary format compositions, and matches or exceeds handwritten libraries in benchmarks — for example a single COO matrix-vector multiply runs up to 3.6× faster than convert-to-CSR baseline. The approach is modular and taco extension ships as prototype.
AI-written summary · about
chapters (6)
- 0:00The sparse tensor challenge
- 3:58Introducing the Taco compiler
- 8:11Representing tensor formats
- 13:24Automatic code generation
- 18:35Evaluation and results
- 21:42Q&A session
related talks
20mA Mechanized Algebra of Verified Data Structures for Optimizing Sparse Tensor Programs
18mCompilation of Shape Operators on Sparse Arrays
28mFunctional Sparse Tensor Compilation
15mCompilation of Sparse Array Programming Models
19mCompiler Support for Sparse Tensor Convolutions
17mA Sparse Iteration Space Transformation Framework for Sparse Tensor Algebra
11mExtending the Scope of the Array Abstraction beyond Sparsity
24mMapping Sparse Applications to Accelerated Computing Systems