CoolerSpace: A Language for Physically Correct and Computationally Efficient Color(…)
summary
A Python library that defines a domain-specific language for color physics, CoolerSpace uses a type-driven physical units type system to forbid mixing incompatible color spaces and to reject underdetermined casts (e.g., LMS → spectrum). It compiles to ONNX and runs equality-saturation optimization via egg to reorder and constant-fold tensor ops (equality saturation via egg). Implementation yields no runtime overhead: raw CoolerSpace is up to 5.7× faster than prior Python systems, optimizations add up to 1.4×, with benchmarks showing ~15% geometric-mean gain and up to 44% best-case speedup.
AI-written summary · about
chapters (13)
- 0:00CoolerSpace introduction
- 1:06Color program challenges
- 2:29Color space mismatches
- 4:58Performance motivation
- 6:03CoolerSpace compilation
- 7:48Type system for colors
- 10:13Equality saturation optimization
- 12:16Benchmarking CoolerSpace
- 13:40Resources and closing
- 14:06Q&A: Precision issues
- 14:48Q&A: Future integration
- 15:11Q&A: Equality saturation role
- 16:14Q&A: Polymorphism in types
related talks
19mSpEQ: Translation of Sparse Codes using Equivalences
20mExploiting Sparsity and Redundancy
28mFunctional Sparse Tensor Compilation
30mA Compiler-First Planetary Compute Engine: Automatic differentiable and performance(…)
13mMind the Abstraction Gap: Bringing Equality Saturation to Real-World ML Compilers
17mA Sparse Iteration Space Transformation Framework for Sparse Tensor Algebra
24mMapping Sparse Applications to Accelerated Computing Systems
28mSpEQ: Translation of Sparse Codes using Equivalences