OpenVINO Execution Provider

Accelerate ONNX models on Intel CPUs, GPUs and VPUs with ONNX Runtime and the Intel OpenVINO execution provider. Please refer to this page for details on the Intel hardware supported.

Contents

• Install
• Requirements
• Build
• Usage
• Support Coverage
• OpenVINO-EP samples Tutorials
• Blogs/Tutorials

Install

Pre-built packages and Docker images are published for ONNX Runtime with OpenVINO by Intel for each release.

• Python wheels: intel/onnxruntime
• Docker image: openvino/onnxruntime_ep_ubuntu18

Requirements

Build

For build instructions, please see the BUILD page.

Usage

C#

To use csharp api for openvino execution provider create a custom nuget package. Follow the instructions here to install prerequisites for nuget creation. Once prerequisites are installed follow the instructions to build openvino and add an extra flag --build_nuget to create nuget packages. Two nuget packages will be created Microsoft.ML.OnnxRuntime.Managed and Microsoft.ML.OnnxRuntime.Openvino.

Multi-threading for OpenVINO EP

OpenVINO Execution Provider enables thread-safe deep learning inference

Heterogeneous Execution for OpenVINO EP

The heterogeneous Execution enables computing for inference on one network on several devices. Purposes to execute networks in heterogeneous mode

To utilize accelerators power and calculate heaviest parts of network on accelerator and execute not supported layers on fallback devices like CPU To utilize all available hardware more efficiently during one inference

For more information on Heterogeneous plugin of OpenVINO, please refer to the Intel OpenVINO Heterogeneous Plugin.

Multi-Device Execution for OpenVINO EP

Multi-Device plugin automatically assigns inference requests to available computational devices to execute the requests in parallel. Potential gains are as follows:

• Improved throughput that multiple devices can deliver (compared to single-device execution)
• More consistent performance, since the devices can now share the inference burden (so that if one device is becoming too busy, another device can take more of the load)

For more information on Multi-Device plugin of OpenVINO, please refer to the Intel OpenVINO Multi Device Plugin.

Auto-Device Execution for OpenVINO EP

Use AUTO:<device 1><device 2>.. as the device name to delegate selection of an actual accelerator to OpenVINO. With the 2021.4 release, Auto-device internally recognizes and selects devices from CPU, integrated GPU and discrete Intel GPUs (when available) depending on the device capabilities and the characteristic of CNN models, for example, precisions. Then Auto-device assigns inference requests to the selected device.

From the application point of view, this is just another device that handles all accelerators in full system.

For more information on Auto-Device plugin of OpenVINO, please refer to the Intel OpenVINO Auto Device Plugin.

Model caching feature for OpenVINO EP

The model caching setting enables blobs with Myriadx(VPU) and as cl_cache files with iGPU.

Save/Load blob capability for Myriadx(VPU)

This feature enables users to save and load the blobs directly. These pre-compiled blobs can be directly loaded on to the specific hardware device target and inferencing can be done. This feature is only supported on MyriadX(VPU) hardware device target.

CL Cache capability for iGPU

Starting from version 2021.4 OpenVINO supports model caching.

This feature enables users to save and load the cl_cache files directly. These cl_cache files can be directly loaded on to igpu hardware device target and inferencing can be done. This feature is only supported on iGPU hardware device target.

There are two different methods of exercising this feature:

option 1. Enabling via Runtime options using c++/python API’s.

This flow can be enabled by setting the runtime config option ‘use_compiled_network’ to True while using the c++/python API’S. This config option acts like a switch to on and off the feature.

The blobs are saved and loaded from a directory named ‘ov_compiled_blobs’ from the executable path by default. This path however can be overridden using another runtime config option ‘blob_dump_path’ which is used to explicitly specify the path where you would like to dump and load the blobs (VPU) or cl_cache(iGPU) files from when already using the use_compiled_network(model caching) setting.

Refer to Configuration Options for more information about using these runtime options.

option 2. Importing the pre-compiled blobs directly from the path set by the user.

This flow enables users to import/load the pre-compiled blob directly if available readily. This option is enabled by explicitly setting the path to the blob using environment variables and setting the OV_USE_COMPILED_NETWORK flag to true.

This flow only works for MyriadX(VPU) device.

For Linux:

export OV_USE_COMPILED_NETWORK=1
export OV_BLOB_PATH=<path to the blob>
Example: export OV_BLOB_PATH=/home/blobs_dir/model.blob

For Windows:

set OV_USE_COMPILED_NETWORK=1
set OV_BLOB_PATH=<path to the blob>
Example: set OV_BLOB_PATH=\home\blobs_dir\model.blob

compile_tool:

The device specific Myriadx blobs can be generated using an offline tool called compile_tool from OpenVINO Toolkit.documentation.

Support for INT8 Quantized models

Starting from the OpenVINO EP 2021.4 Release, int8 models will be supported on CPU and GPU. However, int8 support won’t be available for VPU.

Support for Weights saved in external files

Starting from the OpenVINO EP 2021.4 Release, support for external weights is added. OpenVINO™ EP now supports ONNX models that store weights in external files. It is especially useful for models larger than 2GB because of protobuf limitations.

See the OpenVINO ONNX Support documentation.

Converting and Saving an ONNX Model to External Data: Use the ONNX API’s.documentation.

Example:

import onnx
onnx_model = onnx.load("model.onnx") # Your model in memory as ModelProto
onnx.save_model(onnx_model, 'saved_model.onnx', save_as_external_data=True, all_tensors_to_one_file=True, location='data/weights_data', size_threshold=1024, convert_attribute=False)

Note:

1. In the above script, model.onnx is loaded and then gets saved into a file called ‘saved_model.onnx’ which won’t have the weights but this new onnx model now will have the relative path to where the weights file is located. The weights file ‘weights_data’ will now contain the weights of the model and the weights from the original model gets saved at /data/weights_data.

2. Now, you can use this ‘saved_model.onnx’ file to infer using your sample. But remember, the weights file location can’t be changed. The weights have to be present at /data/weights_data

3. Install the latest ONNX Python package using pip to run these ONNX Python API’s successfully.

Support for IO Buffer Optimization

To enable IO Buffer Optimization we have to set OPENCL_LIBS, OPENCL_INCS environment variables before build. For IO Buffer Optimization, the model must be fully supported on OpenVINO and we must provide in the remote context cl_context void pointer as C++ Configuration Option. We can provide cl::Buffer address as Input using GPU Memory Allocator for input and output.

Example:

//Set up a remote context
cl::Context _context;
.....
// Set the context through openvino options
OrtOpenVINOProviderOptions options;
options.context = (void *) _context.get() ;
.....
//Define the Memory area
Ort::MemoryInfo info_gpu("OpenVINO_GPU", OrtAllocatorType::OrtDeviceAllocator, 0, OrtMemTypeDefault);
//Create a shared buffer , fill in with data
cl::Buffer shared_buffer(_context, CL_MEM_READ_WRITE, imgSize, NULL, &err);
....
//Cast it to void*, and wrap it as device pointer for Ort::Value
void *shared_buffer_void = static_cast<void *>(&shared_buffer);
Ort::Value inputTensors = Ort::Value::CreateTensor(
        info_gpu, shared_buffer_void, imgSize, inputDims.data(),
        inputDims.size(), ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT);

Configuration Options

OpenVINO EP can be configured with certain options at runtime that control the behavior of the EP. These options can be set as key-value pairs as below:-

Python API

Key-Value pairs for config options can be set using InferenceSession API as follow:-

session = onnxruntime.InferenceSession(<path_to_model_file>, providers=['OpenVINOExecutionProvider'], provider_options=[{Key1 : Value1, Key2 : Value2, ...}])

Note that the next release (ORT 1.10) will require explicitly setting the providers parameter if you want to use execution providers other than the default CPU provider (as opposed to the current behavior of providers getting set/registered by default based on the build flags) when instantiating InferenceSession.

C/C++ API

All the options shown below are passed to SessionOptionsAppendExecutionProvider_OpenVINO() API and populated in the struct OrtOpenVINOProviderOptions in an example shown below, for example for CPU device type:

OrtOpenVINOProviderOptions options;
options.device_type = "CPU_FP32";
options.enable_vpu_fast_compile = 0;
options.device_id = "";
options.num_of_threads = 8;
options.use_compiled_network = false;
options.blob_dump_path = "";
options.context = 0x123456ff;
SessionOptionsAppendExecutionProvider_OpenVINO(session_options, &options);

Summary of options

The following table lists all the available configuration options and the Key-Value pairs to set them:

Valid Hetero or Multi or Auto Device combinations: HETERO:,,... The can be any of these devices from this list ['CPU','GPU','MYRIAD','FPGA','HDDL']

A minimum of two DEVICE_TYPE’S should be specified for a valid HETERO or Multi-Device Build.

Example: HETERO:MYRIAD,CPU AUTO:GPU,CPU MULTI:MYRIAD,GPU,CPU

Other configuration settings

Onnxruntime Graph Optimization level

OpenVINO backend performs both hardware dependent as well as independent optimizations to the graph to infer it with on the target hardware with best possible performance. In most of the cases it has been observed that passing in the graph from the input model as is would lead to best possible optimizations by OpenVINO. For this reason, it is advised to turn off high level optimizations performed by ONNX Runtime before handing the graph over to OpenVINO backend. This can be done using Session options as shown below:-

Python API

options = onnxruntime.SessionOptions()
options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_DISABLE_ALL
sess = onnxruntime.InferenceSession(<path_to_model_file>, options)

C/C++ API

SessionOptions::SetGraphOptimizationLevel(ORT_DISABLE_ALL);

Deprecated: Dynamic device type selection

Note: This API has been deprecated. Please use the mechanism mentioned above to set the ‘device-type’ option. When ONNX Runtime is built with OpenVINO Execution Provider, a target hardware option needs to be provided. This build time option becomes the default target harware the EP schedules inference on. However, this target may be overriden at runtime to schedule inference on a different hardware as shown below.

Note. This dynamic hardware selection is optional. The EP falls back to the build-time default selection if no dynamic hardware option value is specified.

Python API

import onnxruntime
onnxruntime.capi._pybind_state.set_openvino_device("<harware_option>")
# Create session after this

This property persists and gets applied to new sessions until it is explicity unset. To unset, assign a null string (“”).

C/C++ API

Append the settings string “" to the EP settings string. Example shown below for the CPU_FP32 option:

std::string settings_str;
...
settings_str.append("CPU_FP32");
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_OpenVINO(sf, settings_str.c_str()));

Support Coverage

ONNX Layers supported using OpenVINO

The table below shows the ONNX layers supported and validated using OpenVINO Execution Provider.The below table also lists the Intel hardware support for each of the layers. CPU refers to Intel^® Atom, Core, and Xeon processors. GPU refers to the Intel Integrated Graphics. VPU refers to USB based Intel^® Movidius^TM VPUs as well as Intel^® Vision accelerator Design with Intel Movidius ^TM MyriadX VPU.

Topology Support

Below topologies from ONNX open model zoo are fully supported on OpenVINO Execution Provider and many more are supported through sub-graph partitioning

Image Classification Networks

Image Recognition Networks

Object Detection Networks

Image Manipulation Networks

*FPGA only runs in HETERO mode wherein the layers that are not supported on FPGA fall back to OpenVINO CPU.

OpenVINO-EP samples Tutorials

In order to showcase what you can do with the OpenVINO Execution Provider for ONNX Runtime, we have created a few samples that shows how you can get that performance boost you’re looking for with just one additional line of code.

Python API

Object detection with tinyYOLOv2 in Python

Object detection with YOLOv4 in Python

C/C++ API

Image classification with Squeezenet in CPP

Csharp API

Object detection with YOLOv3 in C#

Blogs/Tutorials

Overview of OpenVINO Execution Provider for ONNX Runtime

OpenVINO Execution Provider

Tutorial on how to use OpenVINO™ Execution Provider for ONNX Runtime Docker Containers

Docker Containers

Tutorial on how to use OpenVINO™ Execution Provider for ONNX Runtime python wheel packages

Python Pip Wheel Packages