to the Inside Java Newscast where we cover recent developments in the OpenJDK community. I'm Nicolai Parlog, Java developer advocate at Oracle, and today I got four topics for you:
- With JDK 16 out the door, we can take some time to take a closer at two of the exciting new additions, the incubating vector API and a detail on records.
- Then we'll peek towards JDK 17, where a new JEP just landed and the release schedule was published.
Ready? Then let's dive right in!
We'll start with the vector API. Recently, Paul Sandoz, Java Architect at Oracle, and Sandhya Viswanathan, Principal Software Engineer at Intel, gave a talk on this topic at Oracle Live. I highly recommend to watch it! Here, I'll cover the motivation for and current status of the API, so you can safely skip the first couple of minutes - I'll leave a time-stamped link in the description below.
▚SIMD Programming & Vector Instructions
So what is this API all about? The Vector API deals with SIMD computing, which is short for Single Instruction, Multiple Data.
That means applying an operation (a single instruction), not just to one set of operands, but to multiple sets in parallel (that's where multiple data comes from). This is done by specialized CPU hardware offering so-called vector instructions that execute these operations in about the same number of cycles for several sets of operands as it would take to execute the same operation on a single set of operands. For example, instead of adding a single pair of numbers to one result, a vector instruction might add 8, 16, or more pairs to as many results without taking much longer.
Most common CPU architectures offer such vector instructions and the just-in-time compiler's auto-vectorizer already makes use of them wherever it can. The problem is that it's far from perfect in identifying and reliably optimizing such scenarios. So for performance-sensitive algorithms that adhere to the SIMD model, the new vector API offers a specialized programming interface. It takes a "what you see is what you get" approach where the API closely resembles common vector instructions. That guarantees that the just-in-time compiler can generate optimal hardware instructions across all supported CPU architectures.
Typical applications are linear algebra, image processing, character decoding - really anything that's heavy on basic arithmetic and needs to apply that to a lot of independent inputs.
The vector API is a cooperation between Oracle and Intel, spearheaded by Paul Sandoz and Sandhya Viswanathan.
It was first released in JDK 16 and is incubating in a module currently named jdk.incubator.vector.
That means in order to use it, you need to enable it with the command line flag
--add-modules=jdk.incubator.vector at compile and run time.
The API will keep incubating for quite some time for three main reasons:
- To gather feedback - so if you have a use case for it, you can help make it better by trying it out and reporting your findings to the Project Panama mailing list (link in the description).
- To provide more architecture-specific implementations, for example for ARM.
- To benefit from Project Valhalla's primitive objects once they arrive and from Panama's foreign memory.
At this point, I'd love to go into details and explain how the new types
VectorSpecies interact, or what a species is for that manner.
Also, what a shape is, why it's important, and how you pick the correct one for your CPU architecture.
But I don't have time for it here, so you should really check out Paul's and Sandhya's talk.
Beyond what I just outlined, Paul goes on to give an example how the vector API makes the JDK more performant and maintainable.
Sandhya has a lot more and pretty cool examples from dot product and matrix multiplication to image manipulation and Mandelbrot generation.
If, after that talk, you still don't have enough, check out Inside Java Podcast Episode 7 with John Rose, Java Virtual Machine Architect at Oracle, and Paul Sandoz as well as FizzBuzz – SIMD Style!, a blog post by Gunnar Morling, software engineer at Red Hat - links to both below.
Records were finalized in Java 16, so if you're on the recent release, you can go ahead and use them in production. To productively use them, though, you need more than just the language - tools and dependencies need to play along as well. For a data-centric feature like records that puts serialization high up on that list.
Quick aside, when I say serialization, I'm not just talking about Java's onboard serialization mechanism that turns instances into a byte stream and vice versa. That one of course already works perfectly fine with records. No, I'm talking about the wider concept where an instance gets turned into any kind of external representation, for example JSON or XML. These and many more formats are under the purview of a number of frameworks and they need to be made aware of records as well.
So to give records a leg up, Chris Hegarty and Julia Boes, both working at the Java Platform Group at Oracle, engaged with three popular Java-based serialization frameworks, namely Jackson, Kryo, and XStream. For Jackson, they helped with reviews, for Kryo and XStream they provided pull requests. Thanks (in part) to their contributions, the recent versions of these three frameworks now support records. By the way, as far as I'm aware so does Apache Johnzon.
If you're interested to learn more about why records and serialization are such a good match or how Julia and Chris implemented these features, check out their blog post on Inside Java - link below.
I really like the changes that went into JDK 16 and will make sure to cover more of them in the coming episodes. If you don't want to miss that, make sure to subscribe. But I also want to take a look at the near future, which is JDK 17. Two things happened here recently.
Then, the release schedule was finalized: JDK 17 will be forked from the main line on June 10th. That means from that point on, development will prioritize bug fixes over new features. The final release candidate is planned for August 19th and general availability for September 14th.
And that's it for today on the Inside Java Newscast. If you have any questions about what I covered in this episode, ask ahead in the comments below and if you like this kind of content help us spread the word with a like or sharing it with friends and colleagues. I'll see you again in two weeks. So long...