🚀 Feature

A generic API that encapsulates certain operators that require contexts. For example weights are prepacked in certain format.

Pitch

• Instead of using _xnnpack::conv2d_prepack and _xnnpack::conv2d_packed, prepacked::.., prepacked::..
  • Conv:
    • prepacked::conv2d_clamp_prepack -> Conv2dOpContext
    • prepacked::conv2d_clamp_run
  • Linear:
    • prepacked::linear_clamp_prepack -> LinearOpContext
    • prepacked::linear_clamp_run
  • OpContext:
    • Stores the engine info or backend info to be used by _packed ops during op execution.
  • Context creation during deserialization can use heuristics to decide which backend to chose from available ones.
    • e.g. if mobile has both XNNPACK and GPU backend available, some heuristics may help decided which backend to use and create context accordingly.
    • For now this can be just based on global context. e.g. if xnnpack backend enabled (and is only enabled on mobile) pack for xnnpack.