/*

* Copyright (c) 2019, Oracle and/or its affiliates. All rights reserved.

* Licensed under the MIT License.

*/

/**

* A Java interface to the ONNX Runtime.

*

* <p>Provides access to the same execution backends as the C library. Non-representable types in

* Java (such as fp16) are converted into the nearest Java primitive type when accessed through this

* API.

*

* <p>There are two shared libraries required: <code>onnxruntime</code> and <code>onnxruntime4j_jni

* </code>. The loader is in {@link ai.onnxruntime.OnnxRuntime} and the logic is in this order:

*

* <ol>

* <li>The user may signal to skip loading of a shared library using a property in the form <code>

* onnxruntime.native.LIB_NAME.skip</code> with a value of <code>true</code>. This means the

* user has decided to load the library by some other means.

* <li>The user may specify an explicit location of all native library files using a property in

* the form <code>onnxruntime.native.path</code>. This uses {@link java.lang.System#load}.

* <li>The user may specify an explicit location of the shared library file using a property in

* the form <code>onnxruntime.native.LIB_NAME.path</code>. This uses {@link

* java.lang.System#load}.

* <li>The shared library is autodiscovered:

* <ol>

* <li>If the shared library is present in the classpath resources, load using {@link

* java.lang.System#load} via a temporary file. Ideally, this should be the default use

* case when adding JAR's/dependencies containing the shared libraries to your

* classpath.

* <li>If the shared library is not present in the classpath resources, then load using

* {@link java.lang.System#loadLibrary}, which usually looks elsewhere on the filesystem

* for the library. The semantics and behavior of that method are system/JVM dependent.

* Typically, the <code>java.library.path</code> property is used to specify the

* location of native libraries.

* </ol>

* </ol>

*

* For troubleshooting, all shared library loading events are reported to Java logging at the level

* FINE.

*

* <p>Note that CUDA, ROCM, DNNL, OpenVINO and TensorRT are all "shared library execution providers"

* and must be stored either in the directory containing the ONNX Runtime core native library, or as

* a classpath resource. This is because these providers are loaded by the ONNX Runtime native

* library itself and the Java API cannot control the loading location.

*/

package ai.onnxruntime;