Arm NN and the plug-in framework
summary
Step‑by‑step guide to adding a hardware backend to Arm NN: use the Graph Builder API to build a network, call the optimizer which assigns layers to backends (it chooses the first supported backend in your provided list and falls back), implement a workload factory that creates synchronous NQ workload executions, and optionally implement create memory manager with acquire/release to allocate device buffers. Support optimize subgraph to merge or emit precompiled layers, register the backend via the backend registry, and provide build files for Linux/Android.
AI-written summary · about
chapters (17)
- 0:00Intro
- 1:45Arm NN Overview (continued)
- 4:22Arm NN Supported Layers
- 4:47Graph Builder
- 5:45Connecting layers
- 6:45A simple example
- 7:44Establishing connections
- 8:09Tensorinfo
- 14:47Optimizer (continued)
- 17:08Runtime
- 19:47Run inference (continued)
- 23:19Subgraphs
- 24:46Workloads
- 27:27Workload Implementation
- 30:28Custom backends (continued)
- 39:59Backend optimization PreCompiled layers
- 43:29Memory Management continued
related talks
38mAutoML for Efficient Neural Architecture Design
42mAI and Machine Learning BoF
32mTowards new deep learning abstractions on top of exist frameworks
16mAccelerating Web AI on
38mNEURAL NETWORKS ON YOUR MOBILE PHONE. DIVING INTO THE NN API
23mDeep dive in Nx Backends
37mDeep Learning Neural Network Acceleration at the Edge
32mDeploy AI with Simulink