Copyright 2013-2019 The Khronos Group Inc.

This specification is protected by copyright laws and contains material proprietary to Khronos. Except as described by these terms, it or any components may not be reproduced, republished, distributed, transmitted, displayed, broadcast or otherwise exploited in any manner without the express prior written permission of Khronos.

This specification has been created under the Khronos Intellectual Property Rights Policy, which is Attachment A of the Khronos Group Membership Agreement available at Khronos Group grants a conditional copyright license to use and reproduce the unmodified specification for any purpose, without fee or royalty, EXCEPT no licenses to any patent, trademark or other intellectual property rights are granted under these terms. Parties desiring to implement the specification and make use of Khronos trademarks in relation to that implementation, and receive reciprocal patent license protection under the Khronos IP Policy must become Adopters and confirm the implementation as conformant under the process defined by Khronos for this specification; see

Khronos makes no, and expressly disclaims any, representations or warranties, express or implied, regarding this specification, including, without limitation: merchantability, fitness for a particular purpose, non-infringement of any intellectual property, correctness, accuracy, completeness, timeliness, and reliability. Under no circumstances will Khronos, or any of its Promoters, Contributors or Members, or their respective partners, officers, directors, employees, agents or representatives be liable for any damages, whether direct, indirect, special or consequential damages for lost revenues, lost profits, or otherwise, arising from or in connection with these materials.

Khronos is a registered trademark, and OpenVX is a trademark of The Khronos Group Inc. OpenCL is a trademark of Apple Inc., used under license by Khronos. All other product names, trademarks, and/or company names are used solely for identification and belong to their respective owners.

1. Kernel Import Extension to OpenVX 1.3

1.1. Purpose

This document details an extension to OpenVX 1.3, and references some APIs and symbols that may be found in that API, at

Provide a way of importing an OpenVX kernel from a vendor binary specified by URL.

vx khr import kernel

The name of this extension is vx_khr_import_kernel.

1.2. Acknowledgements

This specification would not be possible without the contributions from this partial list of the following individuals from the Khronos Working Group and the companies that they represented at the time:

  • Radhakrishna Giduthuri - AMD

  • Niclas Danielsson, Axis Communications AB

  • Frank Brill, Cadence

  • Thierry Lepley, Cadence

  • Adam Herr, Intel

  • Jesse Villarreal, Texas Instruments

1.3. Example: AlexNet graph

In order to use a neural network in OpenVX graph, one may to use the process outlined below:

  • Import a pre-trained neural network kernel into the context from a vendor binary specified by URL. Use the vxImportKernelFromURL API to import the neural network kernel.

  • Create an OpenVX graph that will use the imported neural network kernel.

  • Create tensor objects for all neural network parameters (i.e., both input and output)

  • Instantiate a neural network node into the graph using the vxCreateGenericNode and vxSetParameterByIndex APIs.

  • Use the vxVerifyGraph API to verify and optimize the graph.

  • Run the OpenVX graph in a loop

#include <VX/vx_khr_import_kernel.h>

void AlexNet( )
    vx_uint32 num_params, i;
    vx_tensor tensors[MAX_TENSORS] = { NULL };

    // create OpenVX context
    vx_context context = vxCreateContext();

    // import neural network kernel
    const char * type = "vx_xyz_folder"; // XYZ's kernel binary container
    const char * url = "/assets/alexnet/"; // folder with AlexNet binary
    vx_kernel nn_kernel = vxImportKernelFromURL(context, type, url);

    // create OpenVX graph
    vx_graph graph = vxCreateGraph(context);

    // add neural network instance as a node in the OpenVX graph
    vx_node node = vxCreateGenericNode(graph, nn_kernel);

    // query number of parameters in imported kernel
    vxQueryKernel(nn_kernel, VX_KERNEL_PARAMETERS, num_params, sizeof(vx_uint32));

    // query parameters of kernel to create tensor objects and add to node
    for(i=0; i<num_params; i++)
        vx_type_e type;
        vx_parameter prm = vxGetKernelParameterByIndex(nn_kernel, i);
        vxQueryParameter(prm, VX_PARAMETER_TYPE, &type, sizeof(vx_type_e));

        if(VX_TYPE_TENSOR == type)
            vx_meta_format meta;
            vx_size num_dims;
            vx_size sizes[MAX_SIZES];
            vx_enum tensor_type;
            vx_int8 fixed_point_precision;

            vxQueryParameter(prm, VX_PARAMETER_META_FORMAT, &meta,

            // Query data needed to create tensor
            vxQueryMetaFormatAttribute(meta, VX_TENSOR_NUMBER_OF_DIMS,
                                       &num_dims, sizeof(vx_size));
            vxQueryMetaFormatAttribute(meta, VX_TENSOR_DIMS,
                                       &sizes, sizeof(sizes));
            vxQueryMetaFormatAttribute(meta, VX_TENSOR_DATA_TYPE,
                                       &tensor_type, sizeof(vx_enum));
            vxQueryMetaFormatAttribute(meta, VX_TENSOR_FIXED_POINT_PRECISION,
                                       &fixed_point_precision, sizeof(vx_int8));

            tensors[i] = vxCreateTensor(context, num_dims, sizes, tensor_type,
        vxSetParameterByIndex(node, i, tensors[i]);


    // verify graph

    // process graph with one batch at a time
    while( userGetNextJobInput(tensors[0]) == VX_SUCCESS )
        // execute the graph to run AlexNet

        // consume the output from AlexNet

    for(i=0; i<num_params; i++)

2. Module Documentation


2.1. Functions

2.1.1. vxImportKernelFromURL

Import a kernel from binary specified by URL.

vx_kernel vxImportKernelFromURL(
    vx_context                                  context,
    const vx_char*                              type,
    const vx_char*                              url);


  • [in] context - OpenVX context

  • [in] type - Vendor-specific identifier that indicates to the implementation how to interpret the url. For example, if an implementation can interpret the url as a file, a folder a symbolic label, or a pointer, then a vendor may choose to use "vx_<vendor>_file", "vx_<vendor>_folder", "vx_<vendor>_label", and "vx_<vendor>_pointer", respectively for this field. Container types starting with "vx_khr_" are reserved. Refer to vendor documentation for list of container types supported.

  • [in] url - URL to binary container.

Returns: A vx_kernel reference. Any possible errors preventing a successful import should be checked using vxGetStatus.


An implementation may provide several different error codes to give useful diagnostic information in the event of failure to create the context.


The name of kernel parameters can be queried using the vxQueryReference API with vx_parameter as ref and VX_REFERENCE_NAME as attribute.