Same API, Different Execution
summary
Claims a small Python-first framework called meta DSL records expressions and rewrites them so the same array API runs with different executors. Demonstrations show a numpy backend executing eagerly, a symbolic compilation path that emits strings/IR for TensorFlow or LLVM/XLA, and a concrete index-first optimization that reduces work. Goal is community-maintainable translations across hardware; integration notes mention MLIR and practical examples that compile and execute rather than remaining theoretical.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 0:20Python code example
- 1:05Von Neumann in Python
- 3:15Hardware slowdown and multi-core
- 6:30Von Neumann bottleneck
- 8:17TensorFlow's graph approach
- 9:00Hardware fragmentation
- 10:26Open source needs
- 11:51Meta DSL project
- 13:20Computation representation
- 15:48API versus execution
- 18:44Symbolic compilation
- 20:52Abstract arrays
- 21:34Accessible high-performance computing
- 22:37Q&A
related talks
25m#6 Open Source Forum 2019 Adam Gibson
26mBuilding the Next Generation of Open Source AI Tooling
43mLeveraging DSLs (Almost) for Free
17mComposable Data Management Systems
26mA FAIR Case for a Live Computational Commons
16mKeynote: Accelerating AI/ML Development with Open Source Software
15mLibreLane: Looking to the future
29mThe Windy Road From Research to Real-World Impact