metadsl: separating API from execution
summary
Claim: metadsl builds a typed, deep-embedded expression graph from normal-looking Python array calls and compiles it to targets like Tensorflow or LLVM; demo shows notebook visualization and backend integrations. Implementation uses decorator-wrapped functions + pattern-matching rewrite rules to perform optimizations (example: index-add fusion and constant-folding) before lowering. It works as an experimental proof of concept but is not a runtime replacement: control-flow tracing and debugging remain rough and compilation adds overhead and engineering complexity.
AI-written summary · about
chapters (17)
- 0:00Introduction to metadsl
- 1:07Python code needs speed
- 2:57Von Neumann in Python
- 4:46Hardware beyond CPU frequency
- 6:51Performance optimization bottleneck
- 8:17Domain-specific languages for hardware
- 11:51API and execution separation values
- 13:17Protocols and static typing in Python
- 16:51Separating API from execution
- 20:50Building expression graphs
- 23:20Evaluating expressions and backends
- 26:32Modular API design
- 28:28Domain-specific languages and embeddings
- 30:12API data versus code data
- 32:40Control flow in functional API
- 34:25Empowering users across stack
- 35:11Q&A
related talks
38mCreating a DSL
21mDataEngBytes 2023 - SYD-T3-04 - Abhinav Goyal
42mGame of Threads - You Spawn or You Die
30mConcrete Syntax Metapatterns
01h 57mMPS Talk Series 2021. Day 4
46mWriting a DSL (Domain Specific Language) for PowerShell
57mSoftware Architecture Lessons from Multidimensional Spaces Explorations
16mBringing the WebAssembly Standard up to Speed with SpecTec