A Halide Backend for TorchInductor
summary
A Halide backend is added to TorchInductor as a reference backend that targets extensibility for kernel‑level DSLs and hardware vendors, integrated with TorchDynamo and sitting alongside Triton and C++ backends. Most models run correctly — about 90% of models on CPU — but early performance is mixed: some kernels are >4x slower on GPU versus Triton and ~1.2x slower on CPU, though a handcrafted Halide schedule brings a microbenchmark within 10% and on ARM about half the models are faster.
AI-written summary · about
chapters (8)
- 0:00Intro
- 0:10TorchCompile
- 1:40Why Halide
- 3:23Backend Comparison
- 4:17Current Status
- 5:24Code Comparison
- 6:47Results
- 7:48Conclusion
related talks
13mLightning Talk: Adding Backends for TorchInductor: Case Study with Intel GPU
24mHelion 1.0: A High-Level DSL for Performance Portable Kernels
07mPyTorch 2.0: TorchInductor
57mHalide: A Language for Fast, Portable Computation on Images and Tensors
28mHelion: A High-level DSL for Kernel Authoring
08mLightning Talk: Achieving SOTA GEMM Performance: A CuTeDSL Backend for PyTorch Induc...
15mLightning Talk: AOTInductor: Ahead-of-Time Compilation for PT2 Exported Models
12mLightning Talk: Accelerating Inference on CPU with Torch.Compile