Mosaic: An Interoperable Compiler for Tensor Algebra
summary
Mosaic is a Taco-based compiler that binds tensor expressions to external library or compiler functions via a scheduling API, substitutes bound subexpressions with calls, and generates the rest as fused code; it registers 38 external functions (≈20 lines each), maps 11 expressions across 7 systems, uses a Z3 solver for index constraints, and offers an automatic search and verification system. Results show up to 206× speedup over fused code in some cases and up to 3.57× the reverse.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:24Tensor Algebra Systems Today
- 2:05Sampling Dense Matrix Multiplication
- 3:26Performance
- 5:24Mosaic
- 6:26Taco
- 8:59Components
- 11:34Scheduling
- 14:38Questions
related talks
24mMapping Sparse Applications to Accelerated Computing Systems
28mC2TACO: Lifting Tensor Code to TACO
25mFormat Abstraction for Sparse Tensor Algebra Compilers
24mVerified Tensor-Program Optimization Via High-Level Scheduling Rewrites
22mTileLoom: Automatic Dataflow Planning for Tile-Based Languages on Spatial Dataflow Accel…
15mCompilation of Sparse Array Programming Models
01h 00mThe Shape of Things to Come
32mCompilation of Dynamic Sparse Tensor Algebra