Setup a JavaCL Maven Project

This quick tutorial will get you through setting up a JavaCL project.

You can either create a plain Maven project by hand and modify its pom.xml as below, or use the fully automated way that uses a Maven archetype :

Create a new project

Choose Maven / Project from Archetype :

In the Maven Archetype screen, click on Add, fill-in the details then click on Ok :

Group Id = com.nativelibs4java
Archetype Id = javacl-simple-tutorial
Version = 1.0-SNAPSHOT
Repository = http://oss.sonatype.org/content/groups/public

Select the custom archetype you’ve just created and click on Next, then Finish :

With IntelliJ IDEA

This will work with the open source Community Edition as well as with the Ultimate Edition. If you haven’t tried it yet, IntelliJ IDEA is a great IDE to work with…

It’s very easy to create the sources for this tutorial with IDEA :

Make sure to install Maven and set the M2_HOME environment variable properly
Create a new project

Select Create a project from scratch :

Choose a name for your tutorial project and select Maven Module for the type of the project, then click on Next :

Tick Create from archetype and click on Add archetype, then fill-in the details :

Group Id = com.nativelibs4java
Archetype Id = javacl-simple-tutorial
Version = 1.0-SNAPSHOT
Repository = http://oss.sonatype.org/content/groups/public

Click Ok and Finish : you’re done

With Maven in command-line

First, please make sure you’ve properly installed Maven.

Then simply type the following commands in a shell :

With Maven 3.0.3+ :
With previous versions of Maven :

This will generate a directory with the following layout :

JavaCLTutorial
|__ pom.xml
|__ src/
     |__ test 
     |__ main/
          |__ java/
          |    |__ tutorial/
          |         |__ JavaCLTutorial1.java
          |         |__ JavaCLTutorial2.java
          |         |__ JavaCLTutorial3.java
          |__ opencl/
               |__ tutorial/
                    |__ TutorialKernels.cl

The three versions of the JavaCLTutorial class correspond to the progression of this tutorial, as you’ll see below.

A very simple OpenCL addition program

This is the OpenCL equivalent of the traditional Hello-world example : we’re just going to perform parallel piece-wise additions on two float vectors, storing the results on a third float vector.

It contains two parts :

an OpenCL source code that will be compiled at run-time and run on an OpenCL device (CPU or GPU)
a Java host program that sets up the JavaCL context, reads and compiles the OpenCL source code and calls it with appropriately initialized arguments.

The Maven `pom.xml` project file

This file contains a single cross-platform JavaCL dependency :

  <repositories>
    <repository>
      <id>sonatype</id>
      <name>Sonatype OSS Snapshots Repository</name>
      <url>http://oss.sonatype.org/content/groups/public</url>
    </repository>
  </repositories>
  
  <dependencies>
    <dependency>
      <groupId>com.nativelibs4java</groupId>
      <artifactId>javacl</artifactId>
      <version>1.0-SNAPSHOT</version>
    </dependency>
  </dependencies>

It also contains the configuration needed by the JavaCL Generator plugin, which will help make your kernels programming experience a lot (type-)safer and more enjoyable (see below).

The `add_floats` OpenCL kernel

Here’s our first kernel in src/main/opencl/tutorial/TutorialKernels.cl :

__kernel void add_floats(__global const float* a, __global const float* b, __global float* out, int n) 
{
    int i = get_global_id(0);
    if (i >= n)
        return;

    out[i] = a[i] + b[i];
}

It’s very simple : it takes the global id of this execution (it will be executed many times in parallel, with only the global id differing between two executions) and adds values found in a and b at index i, storing the result in out at index i.

The JavaCL host program

Here’s the contents of src/main/java/tutorial/JavaCLTutorial1.java :

package tutorial;

import com.nativelibs4java.opencl.*;
import com.nativelibs4java.opencl.util.*;
import com.nativelibs4java.util.*;
import org.bridj.Pointer;
import static org.bridj.Pointer.*;
import static java.lang.Math.*;

public class JavaCLTutorial1 {
    public static void main(String[] args) {
        CLContext context = JavaCL.createBestContext();
        CLQueue queue = context.createDefaultQueue();
        ByteOrder byteOrder = context.getByteOrder();
        
        int n = 1024;
        Pointer<Float>
            aPtr = allocateFloats(n).order(byteOrder),
            bPtr = allocateFloats(n).order(byteOrder);

        for (int i = 0; i < n; i++) {
            aPtr.set(i, (float)cos(i));
            bPtr.set(i, (float)sin(i));
        }

        // Create OpenCL input buffers (using the native memory pointers aPtr and bPtr) :
        CLBuffer<Float> 
            a = context.createBuffer(Usage.Input, aPtr),
            b = context.createBuffer(Usage.Input, bPtr);

        // Create an OpenCL output buffer :
        CLBuffer<Float> out = context.createBuffer(Usage.Output, n);

        // Read the program sources and compile them :
        String src = IOUtils.readText(JavaCLTutorial1.class.getResource("TutorialKernels.cl"));
        CLProgram program = context.createProgram(src);

        // Get and call the kernel :
        CLKernel addFloatsKernel = program.createKernel("add_floats");
        addFloatsKernel.setArgs(a, b, out, n);
        CLEvent addEvt = addFloatsKernel.enqueueNDRange(queue, new int[] { n });
        
        Pointer<Float> outPtr = out.read(queue, addEvt); // blocks until add_floats finished

        // Print the first 10 output values :
        for (int i = 0; i < 10 && i < n; i++)
            System.out.println("out[" + i + "] = " + outPtr.get(i));
        
    }
}

Running

If you’re using Maven in command-line, type this in the shell in project’s folder (otherwise, click on Run in your IDE of choice ;-)) :

mvn compile exec:java -Dexec.mainClass=com.mycompany.JavaCLTutorial1

The end of the output will look like this (please allow some time the first time Maven runs : it will download many files, but won’t do it again ;-)) :

out[0] = 1.0
out[1] = 1.3817732
out[2] = 0.49315056
out[3] = -0.8488725
out[4] = -1.4104462
out[5] = -0.6752621
out[6] = 0.6807548
out[7] = 1.4108889
out[8] = 0.84385824
out[9] = -0.49901175

Refining that simple code…

Use the JavaCL program wrapper generator to make the code even simpler

Saw that “javacl-generator” plugin configuration in the pom.xml file ?

JavaCL Generator is a tool that parses any .cl file present in src/main/opencl and creates a wrapper class that only accepts the correct argument types and numbers, instead of the all-forgiving CLKernel.setArgs(Object…) that might make your program crash at runtime if you used an incorrect argument type (or missed an argument).

As the Maven pom.xml file you’ve got from the archetype already contains the correct configuration for the generator, you can simply modify the host code as follows (see JavaCLTutorial2.java):

        /* 
        This code is no longer needed :
        // Read the program sources and compile them :
        String src = IOUtils.readText(JavaCLTutorial.class.getResource("TutorialKernels.cl"));
        CLProgram program = context.createProgram(src);
        
        // Get and call the kernel :
        CLKernel addFloatsKernel = program.createKernel("add_floats");
        addFloatsKernel.setArgs(a, b, out, n);
        CLEvent evt = addFloatsKernel.enqueueNDRange(queue, new int[] { n });
        */

        // Instantiate the auto-generated program wrapper and call the kernel :
        TutorialKernels kernels = new TutorialKernels(context);
        CLEvent addEvt = kernels.add_floats(queue, a, b, out, n, new int[] { n }, null);

        Pointer<Float> outPtr = out.read(queue, addEvt); // blocks until add_floats finished

Simple, isn’t it ?

The program wrappers will be regenerated automatically at each compilation, so they’ll keep in sync with your OpenCL kernels !

Perform (and chain) more calculations in OpenCL

In the code above, we’ve initialized the a and b buffers with pointers to memory that we’ve filled by hand in Java… Why not fill it straight in OpenCL (i.e. on the GPU, if the context uses a GPU device) ?

Let’s consider this other kernel in src/main/opencl/TutorialKernels.cl :

__kernel void fill_in_values(__global float* a, __global float* b, int n) 
{
    int i = get_global_id(0);
    if (i >= n)
        return;

    a[i] = cos((float)i);
    b[i] = sin((float)i);
}

With the following changes in the Java code (see JavaCLTutorial3.java) :

        // Create OpenCL input and output buffers :
        CLBuffer<Float> 
            a = context.createBuffer(Usage.InputOutput, n), // a and b are now read AND written to
            b = context.createBuffer(Usage.InputOutput, n),
            out = context.createBuffer(Usage.Output, n);

        // Instantiate the auto-generated program wrapper and chain calls to the two kernel :
        TutorialKernels kernels = new TutorialKernels(context);

        CLEvent fillEvt = kernels.fill_in_values(queue, a, b, n, new int[] { n }, null);
        CLEvent addEvt = kernels.add_floats(queue, a, b, out, n, new int[] { n }, null, fillEvt);

        Pointer<Float> outPtr = out.read(queue, addEvt); // blocks until add_floats finished

See how we’ve chained dependent operations through events (operations are executed asynchronously from the Java program) :

the add_floats kernel execution must wait for the fill_in_values execution to finish
the out.read operation must wait for the add_floats execution to finish.

Wrap up

That’s it ! In this short tutorial, you’ve seen how to :

setup a simple JavaCL project (in command-line, with Netbeans or IntelliJ IDEA)
allocate OpenCL buffers
call and chain executions of simple OpenCL kernels
use the JavaCL Generator to get free compile-time checks of OpenCL kernel arguments and less boilerplate code

Stop Thinking, Just Do!