The OpenVX^™ Feature Set Definitions

The purpose is to define a minimal subset of OpenVX features that enable the construction and execution of OpenVX graphs, but it does not contain any specific vision-processing operations. Other feature sets build on this basic framework to add sufficient functionality to enable useful applications. The Base Feature Set is not a conformance feature set, and is defined for convenience of organization and explanation in this document.

The Base Feature Set requires support for the foundational vx_context, vx_reference, vx_graph, vx_kernel and vx_node objects. The vx_parameter and vx_meta_format objects provide the necessary functions to query parameters of the imported kernels, so they are also required.

The name of this feature set is vx_khr_base.

The Base Feature Set includes the following framework objects in their entirety, including all functions, macros, typedefs, and enumerations described in their respective sections in the main OpenVX specification:

The Base Feature Set also requires support for User Kernels as decribed in the main OpenVX specification in its entirety, including all functions, macros, typedefs, and enumerations described in the User Kernel section of the main specification.

To provide a basic set of vision processing functions. This set of functions is roughly equivalent to the set of functions available in version 1.1 of the OpenVX specification. In addition to the framework objects included in the Base Feature Set, the Vision Conformance Feature Set includes a set of data objects that the Vision functions operate upon and produce.

The Vision Conformance Feature Set includes all the functions and objects in the Base Feature Set, plus the following data objects and vision functions.

To provide a basic set of neural-network functions. This conformance feature set is roughly equivalent to the OpenVX neural network extension specification, plus the portions of the main specification needed to support these neural-network functions.

The Neural Network Conformance Feature Set includes all the functions and objects in the Base feature set, plus the following data objects and neural-network functions.

Provide a minimum set of functions to import and execute neural networks described in the NNEF standard format. Applications using this feature set will use the vxImportKernelFromURL function to import an NNEF file at the location of the URL to create an OpenVX kernel representing the neural network. This kernel can subsequently be used to create a node in an OpenVX graph, which can be executed using the normal OpenVX functions from the Base Feature Set. The inputs and outputs of the neural network node will be vx_tensor objects.

This feature set is dependent on the Base feature set and the tensor data object, which must also be supported in order to support this feature set.

The name of this feature set is vx_khr_nnef_import.

The NNEF Import Conformance Feature Set includes all the functions and objects in the Base feature set, support of the Kernel import extension vx_khr_import_kernel, which contains the vxImportKernelFromURL function, plus the following data objects and vision functions.

In the case where the imported neural network model defines a custom operation, namely an operation with known interfaces but unknown functionality, the implementation of such a custom operation must be provided as an OpenVX user kernel. This user kernel must be registered in the OpenVX context prior to the calling the vxImportKernelFromURL function. The registered user kernel must be consistent with the custom kernel declared in the NNEF model in term of:

The OpenVX implementation is responsible for detecting potential inconsistencies at vxImportKernelFromURL call time and/or at the time the imported kernel is instantiated as a node in an OpenVX graph. When importing from NNEF, the following correspondence is defined between the NNEF custom operation and the OpenVX user kernel that implements it:

To enable highly-efficient and compact manipulation of binary, i.e., U1, format images.

This feature set is dependent on the Vision Conformance Feature Set defined above, so an implementation must pass all the conformance tests for that feature set. In addition, the implementor may optionally enable the U1 conformance tests. If the implementation passes all the U1 conformance tests (as well as those for vision conformance) then the implementor can claim support for this binary image feature set.

The functions that are tested for U1 support by the U1 conformance tests are identified in "Inputs" and "Outputs" tables in the main OpenVX specification. Functions that require U1 support are indicated in the "U1" columns of these tables.

To provide an enhanced set of vision processing functions. This set of functions is roughly equivalent to the set of functions introduced in version 1.2 and later of the OpenVX specification.

This feature set is dependent on the Vision Conformance Feature Set defined above, so an implementation must pass all the conformance tests for that feature set. In addition, the implementor may optionally enable the enhanced vision conformance tests. If the implementation passes all the enhanced vision conformance tests (as well as those for regular vision conformance) then the implementor can claim support for this enhanced vision feature set.

The safety-critical environment (for example ISO26262) requires an implementation to satisfy rigorous demands for deployment. For development, an implementation must satisfy the lesser demands of a software tool used to create such a deployment. This section defines a deployment feature set, which is generally a subset of the entire OpenVX specification. In this context, the entire set of OpenVX features can be referred to as the development feature set that is run in a development environment with a full set of debug tools, as opposed to the deployment feature set that runs on an embedded target device with limited resources. A developer may use the full set of features to create and export a graph, and then for deployment this graph is imported by a program that only uses features in the Deployment Feature Set.

A safety-critical implementation of OpenVX must include a Development Feature Set that passes the conformance test suite for one or more of the Conformance feature sets described above. It must also implement and pass the conformance tests for the Import/Export extension. There are no special additional tests for the Deployment Feature Set or for Safety-Critical implementations generally.

The safety-critical environment requires that graphs must execute in a deterministic, reproducible way. It is up to the implementation to guarantee this behavior in some way. The implementation-dependent behaviors must be defined and documented by the implementation.

The Deployment Feature Set requires only a subset of features described in the “Basic Features” and “Administrative Features” sections of the specification, as well as the vx_khr_ix extension.

A Venn diagram of the relationship between the feature sets is shown below. Details are provided in the following sections.

The following table lists the required deployment features. Since the Deployment Feature Set does not support graphs to be constructed, none of the vxXXXNode() functions in the "Vision Functions" section of the specifications are listed in this table.

Some safety-critical environments may enforce software development guidelines (for example MISRA C:2012) to facilitate code quality, safety, security, portability and reliability. In order to meet such guidelines, developers may modify OpenVX standard header files without deviating from the OpenVX specification.

Refer to https://www.khronos.org/registry/OpenVX for the OpenVX standard header packages.