Halide: A Language for Fast, Portable Computation on Images and Tensors

on YouTube ↗57m

" summary

Halide is a C++-embedded DSL that separates algorithm from schedule and uses tiling, vectorization and compute_at to trade locality, parallelism and redundant recomputation. With an LLVM backend it JITs or AOT-compiles pipelines, runs auto-schedulers (Mulipudi 2016, Lee 2018, atoms 2019 auto-scheduler), and is used in Google Pixel camera, Photoshop and YouTube. On a 4 MP image a tuned schedule reaches ~0.4 ms on four cores (≈251× faster than a naive 100 ms implementation).

AI-written summary · about

" chapters (47)

1. 0:00Introduction
2. 0:45What are DSLs
3. 2:06What is Halide
4. 2:50Halide Team
5. 4:11Summary
6. 5:04CPU Optimization
7. 6:08CPU Independent Work
8. 6:42Recompute Values
9. 7:56Locality
10. 8:58Dependency Chain
11. 9:55Optimized Libraries
12. 10:34Why C
13. 11:47Running Example
14. 11:57Two Stage Blur
15. 12:52Edge Cases
16. 13:21C Example
17. 13:48Boundary Extending
18. 14:38Horizontal Blur
19. 15:36Pixel Cache
20. 16:09Swap Loops
21. 16:46Parallelization
22. 17:18Is it portable
23. 17:54Is it correct
24. 18:09Is it maintainable
25. 18:31Define the algorithm
26. 19:09Blur pipeline
27. 20:53Questions
28. 22:50Schedule
29. 26:25Inlining
30. 27:19Vectorize
31. 28:40Tile
32. 30:42Bonus
33. 31:36Fastest Schedule
34. 33:06AutoSchedulers
35. 34:16AutoSchedulers 2019
36. 35:23Questions Answers
37. 38:09Compiler Infrastructure
38. 39:04Halide in C
39. 40:46JIP Mode
40. 41:20Halide Program
41. 42:43Ahead of Time Compilation
42. 43:33Generator Code
43. 45:18CrossCompile
44. 45:53CMake
45. 46:58Runtime
46. 48:02Conclusion
47. 48:49QA

" related talks

• 08mA Halide Backend for TorchInductorJason Ansel · PyTorch Conference 2024
• 15mEfficient Automatic Scheduling of Imaging & Vision Pipelines for the GPUJonathan Ragan-Kelley, Tzu-Mao Li, Andrew Adams, Luke Anderson, Karima Ma, Tian Jin · OOPSLA 2021
• 24mHelion 1.0: A High-Level DSL for Performance Portable KernelsOguz Ulgen · PyTorch Conference Europe 2026
• 28mHelion: A High-level DSL for Kernel AuthoringJason Ansel · PyTorch Conference 2025
• 01h 00mThe Shape of Things to ComeJared Roesch · ARRAY 2026
• 24mVerified Tensor-Program Optimization Via High-Level Scheduling RewritesAdam Chlipala, Jonathan Ragan-Kelley, Gilbert Louis Bernstein, Amanda Liu · POPL 2022
• 22mTileLoom: Automatic Dataflow Planning for Tile-Based Languages on Spatial Dataflow Accel…Wei Li, Cheng Tan, Zhenyu Bai, Heru Wang, Pranav Dangi, Zhiqiang Zhang, Huiying Lan, Weng-Fai Wong, Tulika Mitra · 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26)
• 41mHelion's New CuTe DSL and Pallas Backends and How We Built Them Via AgentsJason Ansel, Dunfan Lu · KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026