MATLAB Coder: Partial Evaluation in Practice

on YouTube ↗53m

" summary

MATLAB Coder converts MATLAB into embeddable C/C++ by using online partial evaluation and a type-driven engine that does function specialization on types and values to produce unboxed C types. It reimplements about 3,000 MATLAB functions, generates multiple specialized versions (example: four versions of sum), turns many runtime errors into compile-time errors, supports opt-in no-dynamic-memory for constrained targets, emits readable C for examples like a Kalman filter, and runs in seconds for small cases and minutes (≈6 min) for large builds.

AI-written summary · about

" chapters (20)

1. 0:00Intro
2. 0:50MATLAB is designed for prototyping
3. 2:33Our goal is to enable MATLAB in production
4. 7:37Focus: MATLAB Coder's "type inference" algorithm
5. 22:25Takeaways from the examples...
6. 23:36MATLAB Coder's Type Inference Engine
7. 25:26Types propagate bottom-up in each statement
8. 26:31Type inference visits statement in natural order
9. 26:57Multiple types assigned to the same variable cause a type
10. 27:21Constant folding and control-flow pruning help avoid type
11. 28:13Function calls produce new function specializations by recursively invoking type inference on the callee
12. 29:07Functions can be specialized not only on input types, but also on constant input values, demand-driven
13. 30:16Iteration over heterogeneous arrays is another use case for specialization
14. 31:01Complete loop unrolling for typing uses of heterogeneous arrays
15. 31:25Type Inference Engine Summary
16. 32:28Future work planned to make type inference more permissive
17. 36:59Compiling for embedded systems requires more than just type inference
18. 39:49Partial evaluation powers tools that enable running MATLAB "anywhere"
19. 40:53And powers MATLAB embedded in Simulink and Stateflow
20. 41:24You can deploy high-level languages to embedded systems

" related talks

• 59mCoq to C Translation with Partial Evaluation + Counterexample GenerationJohn Toman, Takeshi Tsukada, Naoki Kobayashi, Akira Tanaka, Hideto Ueno · PEPM 2021
• 09mCompilers for nothing, executables for free!Vaibhav Sagar · !!Con West 2020
• 28mCode Generation in LinneaHenrik Barthels, Paolo Bientinesi · ARRAY 2019
• 06mCode Generation: The Lisp WayJean-Louis Leroy · C++Now 2018
• 01h 01mA partial history of partial evaluationJulia Lawall, Peter Sestoft · PEPM 2022
• 01h 01m“How bad is Meta-Programming still today?"Peter Gottschling · CppCon 2016
• 24mCakes That Bake Cakes: Dynamic Computation in CakeMLMagnus O. Myreen, Yong Kiam Tan, Thomas Sewell, Ramana Kumar, Alexander Mihajlovic, Oskar Abrahamsson, Scott Owens · PLDI 2023
• 01h 06m“C++ Metaprogramming: Journey from simple to insanity and back"Fedor Pikus · CppCon 2015