Advanced Concepts

Use these pages when your application needs more than the basic Model, Graph, and Run workflow. They explain the contracts and runtime behavior behind richer Neat applications.

Application Design

Design how your application is composed and how it exposes outputs.

• GraphsCompose models, nodes, named inputs and outputs, branches, and runs.
• Use a MIPI CameraBuild a source-owned camera-to-model graph with libcamera input and CVU preprocessing.
• Send VideoStream video output from a Neat application over H.264 RTP/UDP.
• Send JSON MetadataPublish structured application metadata over UDP JSON.

Execution Model

Understand how work is scheduled, threaded, and mapped onto the pipeline layer.

• Timing ModelUnderstand sync and async execution, push/pull, and when work happens.
• Threading ModelKnow which threads exist and where application code may run.
• GStreamer LayerLearn what Neat abstracts and when raw GStreamer details matter.

Data & Model Contracts

Understand the tensor, memory, and model contracts that pipelines rely on.

• Data FormatsMap format tokens to tensor layout, shape, and plane semantics.
• Dtype ContractFollow how tensor precision changes across preprocess, MLA, and postprocess.
• Memory ModelUnderstand zero-copy buffers, physical addresses, and cache behavior.

Model Runtime

Go deeper into compiled model artifacts and the hardware backends that execute them.

• MPK ContractSee how compiled model archives define inference stages and contracts.
• Processor BackendsUnderstand A65, CVU, MLA, MLASHM, APU, TVM, and M4 roles.
• CVU KernelsReview preprocess and postprocess graph building blocks.