AnyDSL: A Partial Evaluation Framework for Programming High-Performance Libraries
summary
AnyDSL presents a new programming system that uses online partial evaluation (via annotations) on a CPS-style IR to generate high-performance library code. It exposes compiler-known higher-order functions for target-specific code generation (vectorization/parallelization) so generic implementations are specialized at compile time for CPUs and GPUs. Evaluation shows three case studies—image processing, ray tracing, genome alignment—where one generic codebase yields CPU/GPU performance within ~10% of hand-tuned expert implementations.
AI-written summary · about
chapters (9)
- 0:00AnyDSL framework
- 4:06Domain-specific languages
- 9:13Classic meta-programming
- 10:41Domain-specific generation
- 13:14Image processing on GPU
- 16:31Domain services
- 18:45Raytracing engine
- 19:12Live demo & performance
- 21:07Memory management
related talks
22mDesign DSLs with xDSL
01h 00mBuilding DSLs on Top of KernelF
30mYin-Yang: Concealing the Deep Embedding of DSLs (Most Influential Paper)
24mDesigning a DSL That Perfectly Solves Your Problem
01h 04mEDSL Infinity Wars: Mainstreaming Symbolic Computation
35mDesigning Domain-Specific Languages: A Comparison of Three Approaches by Example
01h 05mFrom meta frameworks and transformations to distributed computing and more
01h 01mA partial history of partial evaluation