Permalink
080394f Aug 30, 2018
1 contributor

Users who have contributed to this file

64 lines (48 sloc) 4.9 KB

ONNX Interface for Framework Integration (ONNXIFI)

ONNXIFI is a cross-platform API for loading and executing ONNX graphs on optimized backends. High-level frameworks and applications can use this API to execute neural network and machine learning models. Hardware vendors can implement this API to expose specialized hardware accelerators and highly optimized software infrastructure to the users.

Core Features

  • Standardized interface for neural network inference on special-purpose accelerators (NPUs), CPUs, GPUs, DSPs, and FPGAs
  • Based on widely supported technologies
    • C API for function calls
    • ONNX format for passing model graphs
    • NCHW tensor layout for passing inputs and outputs
  • Dynamic discovery of available backends for model execution
    • Multiple backends from different vendors can co-exist on the same system
  • Dynamic discovery of supported ONNX Operators on each backend

Optional Features:

  • Graphs with variable-shape inputs and/or outputs
  • Graphs with data-dependendent output shapes

How to Use ONNX Interface for Framework Integration

  1. (Optional) Use onnxifi_load to dynamically load the ONNX Interface for Framework Integration library.
  2. Call onnxGetBackendIDs to get stable identifiers of available backends. Note: it is possible there are no backends installed in the system.
  3. Call onnxGetBackendInfo to check additional information about any available backend.
  4. Call onnxGetBackendCompatibility to check which operations within your model can run on the backend.
  5. Call onnxInitBackend to initialize a backend, then call onnxInitGraph to offload one or more model graphs to the backend.
  6. Call onnxSetGraphIO to set locations and shapes for inputs and outputs of a graph.
  7. Initialize an inputFence structure of type onnxMemoryFenceV1: set tag to ONNXIFI_TAG_MEMORY_FENCE_V1, type to ONNXIFI_SYNCHRONIZATION_EVENT, and call onnxInitEvent to initiaze the event member.
  8. Initialize an outputFence structure of type onnxMemoryFenceV1: set tag to ONNXIFI_TAG_MEMORY_FENCE_V1, type to ONNXIFI_SYNCHRONIZATION_EVENT, and event to null.
  9. Call onnxRunGraph with the initialized inputFence and outputFence structures to enable execution of the graph. The call to onnxRunGraph will populate event member of the outputFence with a newly created event object, asynchronously execute the graph once inputFence's event is signalled, and then signal the outputFence's event.
  10. Call onnxSignalEvent with event member of inputFence to signal to the backend that the inputs are ready to be consumed.
  11. Call onnxWaitEvent (alternatively, repeatedly call onnxGetEventState in a loop until the event state is ONNXIFI_EVENT_STATE_SIGNALLED) with event member of outputFence to wait until graph outputs are ready to be consumed. Release events for inputs and outputs using onnxReleaseEvent.
  12. If your model works with fixed-size inputs and outputs, and shape and location of inputs and outputs does not change, one call to onnxSetGraphIO is sufficient for multiple onnxRunGraph calls. The previous call to onnxRunGraph, however, must have finished before a user calls onnxRunGraph again, because concurrent execution with the same input and output locations is not allowed. For models with variable-size inputs or outputs, you'd need to call onnxSetGraphIO before each onnxRunGraph call.
  13. When done using the model, release the model graph(s) with onnxReleaseGraph, then release the backend with onnxReleaseBackend and backend ID with onnxReleaseBackendID.

How to Implement ONNX Interface for Framework Integration

The minimum functionality an ONNXIFI implementation must provide is the following:

  • Support ONNX 1.0 model format.
    • There is no minimum list of Operators a backed has to support.
  • Support graph inputs / outputs in CPU memory.
  • Support graph inputs / outputs with fixed shape, specified in GraphProto message.

Discovery

Vendor-provided libraries should adhere to some rules to ensure discovery by ONNX-supported frameworks and applications:

  1. The libraries must be installed in the following directories:
  • GNU/Linux: user-installed system library directory (typically /usr/lib)
  • macOS: /opt/onnx/lib
  • Windows: system directory (typically C:\Windows\System32)
  1. Filenames of vendor-specific libraries must follow the rule below:
  • On Windows, library filename must match wildcard onnxifi-*.dll
  • On macOS, library filename must match wildcard libonnxifi-*.dylib
  • On Linux and other OSes, library filename must match wildcard libonnxifi-*.so

Extensions

Hardware vendors are welcome to add their own extensions to ONNX backend interface. The backend interface offers several extension mechanisms:

  • Experimental, exotic, or vendor-specific operators can be supported in a private domain using NodeProto.domain attribute.
  • Vendor-provided ONNXIFI implementation can expose additional functions