[Summer of Code] build liboneflow.so for Java

[zzk0]

构建

1. clone 指定分支代码: git clone -b libOneFlow_java https://github.com/zzk0/oneflow.git zzk-oneflow --depth=1
2. 安装依赖: sudo apt install -y libopenblas-dev nasm g++ gcc python3-pip cmake autoconf libtool
3. 安装 java 依赖: sudo apt install openjdk-11-jdk maven
4. mkdir build, cd build
5. cmake .. -DCMAKE_BUILD_TYPE=Debug -DBUILD_JNI=ON -C ../cmake/caches/cn/cuda.cmake -DBUILD_HWLOC=OFF
6. 构建 jar 包: make -j$(nproc) oneflow-api
7. 构建动态链接库: make -j$(nproc) oneflow_java
8. 导出到环境变量 LD_LIBRARY_PATH, export LD_LIBRARY_PATH=/path-to-project/build/oneflow/api/java:$LD_LIBRARY_PATH
9. 如果想要运行测试，需要进入到 oneflow/oneflow/java 文件夹，然后 mvn test。测试开始的时候，会下载测试需要的资源。

文件结构说明

Java 代码存放在 oneflow/java 目录下面，JNI 相关的代码存放在 oneflow/api/java 目录下面。

Java 代码

(transformers) percent1@ubuntu:~/zzk-oneflow$ tree oneflow/java/src/
oneflow/java/src/
├── main
│   └── java
│       └── org
│           └── oneflow
│               ├── DType.java
│               ├── InferenceSession.java
│               ├── OneFlow.java                # 和 JNI 对齐
│               ├── Option.java
│               ├── tensor
│               │   ├── FloatTensor.java
│               │   └── IntTensor.java
│               └── Tensor.java
└── test
    └── java
        └── org
            └── oneflow
                └── InferenceSessionTest.java   # 测试

JNI 代码

(transformers) percent1@ubuntu:~/zzk-oneflow$ tree oneflow/api/java/
oneflow/api/java/
├── CMakeLists.txt
├── env
│   └── env_api.h          # C API
├── job
│   └── job_api.h          # C API
├── library.cpp            # 从 Java 读取数据、调用封装的 API、发送数据
├── library.h              # 和 OneFlow.java 对齐
├── session
│   └── session_api.h      # C API
└── util
    └── jni_util.h         # 方便 JNI 编程，比如 jstring 和 string 互转，大端小端检测 

使用例子

@Test
public void example() {
    String jobName = "mlp_inference";

    Option option = new Option();
    option.setDeviceTag("gpu")
            .setControlPort(11245)
            .setSavedModelDir("mnist_test/models")
            .setModelVersion("1");

    InferenceSession inferenceSession = new InferenceSession(option);
    inferenceSession.open();

    float[] image = readImage("mnist_test/test_set/00000000_7.png");
    Tensor imageTensor = Tensor.fromBlob(image, new long[]{ 1, 1, 28, 28 });
    Tensor tagTensor = Tensor.fromBlob(new int[]{ 1 }, new long[]{ 1 });
    Map<String, Tensor> tensorMap = new HashMap<>();
    tensorMap.put("Input_14", imageTensor);
    tensorMap.put("Input_15", tagTensor);

    Map<String, Tensor> resultMap = inferenceSession.run(jobName, tensorMap);
    for (Map.Entry<String, Tensor> entry : resultMap.entrySet()) {
        Tensor resTensor = entry.getValue();
        float[] resFloatArray = resTensor.getDataAsFloatArray();
    }

    inferenceSession.close();
}