Chet Haase's Blog

Swing Archives

Intermediate Images

"Freebird!"

I'm a little afraid of posting this blog, thinking that it could resulting in either

• So many and varied requests that it will be difficult to address them all

• No responses at all. (If a blog falls in the forest and noone reads it, does it still make a sound?)

Nevertheless, I'll forge ahead. After all, the whole point in my blogs/articles is to talk about stuff that developers want or need to know more about.

So, in the words of the unforgettable sourthern rock band and Java desktop client developers Lynyrd Skynyrd:

What is it you want to hear?

In other words, what would you like to see articles on? I have a few ideas kicking around in my head that I'd like to cover, and people have suggested a few more. But if you have other ideas that you would like to have considered, please tell me in the feedback section below.

Here are a few topics that I would like to cover sometime soon, to give you a feel for where things are headed. If you have opinions on these, feel free to post those below as well:

• Managed Images: I've written about Managed Images in various forums so far (such as the BufferedImage (Part 1 and Part 2), VolatileImage (Part 1 and Part 2), and Image Strategies articles). But maybe managed images deserve their own dedicated article.
• Old vs New: I believe that some of the confusion and performance problems in Java graphics usage comes from the difference between the old Java graphics APIs (those written in the jdk 1.x days, e.g. ImageProducer, ImageConsumer, PixelGrabber) and the new ones (starting from jdk 1.2, e.g. BufferedImage). I would like to write an article (or maybe a series of articles) on the differences between these APIs and their implementations. Specifically, I would like to examine how developers would do things (such as image pixel manipulation) on the old graphics APIs and how they would do things using the newer APIs (hopefully better, faster, and more efficiently).
• Server-based graphics: This is also probably a series of articles. One thread is a discussion of how to do graphics operations (such as image creation and manipulation) in server apps (when you are not displaying to screen, but rather to/from images only). Another thread involves rich clients for server-based applications (such as Web Services); how can we make it easier to write clients that interact on the network and with servers? These are all pretty complex topics, so understanding and explaining the issues may take some time. But in this increasingly connected world, this seems like a useful topic in graphics development.
• ImageIO Utilities: While ImageIO simplified many things in image reading/writing and manipulation, it complicated some operations that are simpler in older APIs. For example, displaying an animated GIF is simple when loaded through some of the old 1.x APIs such as Toolkit.getImage(). But when using ImageIO, the developer needs to do much more work to display the images correctly. It's that whole "with great power comes great responsibility" thing; the power of the ImageIO APIs and the information that it exposes to the developer also (for now) necessitates more work to make some operations succeed. I would like to explore some of these issues to see if we can derive some simple frameworks or approaches that developers can reuse in their work to simplify their applications.
• Image Scaling: This is related to the "Old vs. New" topic above; some operations in our APIs are not obvious. Also, even when there are obvious methods in the API for accomplishing tasks, sometimes it is not clear how these methods work, or that the best ways of accomplishing the tasks are perhaps not related to the most obvious means. For example, Image.getScaledInstance() is not, in general, the best way to scale an image, either performance- or quality-wise. The Graphics.drawImage() methods are much better in general, but they are not necessarily obvious to developers new to the Graphics class. I would like to cover some of these areas and discuss the merits of the different approaches.
• Rendering Hints: Wouldn't it be nice to have the hints discussed in a single area where it is more clear (than in the current javadocs) what happens when various hints are enabled?
• Transform Tips & Tricks: There are always questions about how to go about transforming things correctly or most efficiently. We could explore some of these topics in an article or series of articles.
• Animation: Some of the *Image* articles have touched on this topic, but perhaps an article dedicated to animation would be helpful. This article could cover everything from performance tips to timer usage.
  Done! I've just posted an article entitled Timing is Everything on java.net. It doesn't cover animation as a whole, but has some information about timers and utility code I developed to add higher-level functionality to the existing Java timers.
  
• Intermediate Images: This topic comes out of something we've discussed in our JavaOne talks before: using images to cache complex rendering. This can be a huge win for animation or performance-sensitive graphics apps. Some sample code and implementation details here might help people take advantage of this high-performance approach to Java graphics development.
  Done! An article on Intermediate Images is now up on java.sun.com.
  

That's all I can think of for now, although I tend to add to my internal list of blog TODOs frequently.

Threadaches

I find myself trying to multithread my life constantly. I've got so many things to do; surely there's a way I can multiplex the chores to get all of them done faster, right?

For example, I'll be brushing my teeth and realize I also need to comb my hair. I'm only using one hand to hold the toothbrush, so I reach for the comb with the other hand. Then I'll start combing my hair, at which point the other hand with the toothbrush stops moving, or (even worse) keeps moving, but in such as way that toothpaste starts running all over the place.

Now I've gone from two actions that would have been simple to perform serially to one interleaved complex action that's resulted in a toothpaste-and-spittle mess and an unruly mop on my head.

The problem here is that, despite the speed of our processors and the simplicity of our actions, some things in life simply cannot be multithreaded.

This is not to say that multithreading itself is unachievable or not worthwhile for some operations. As an example, consider breathing or blinking; if I had to stop any other action to make my eyes blink or to breathe, I'd never get anything done. Consider multiplexing actions while driving; if we couldn't do many things at once in this environment, we'd all be dead. (Of course, many people assume that this multiplexing-while- driving capability is universal and extends to complex conversations on phones while driving at 80 on the highway; I believe these folks will eventually be evolved out of our society, although they may take a few of us with them along the way).

So there are clearly some operations which are better done on separate threads; the trick is figuring out which ones are which if you don't want to be wiping up toothpaste off your clothes all the time. Or, in the case of your applications, if you don't want to be debugging thread deadlocks or performance problems due to thread abuse.

That reminds me; I was going to write about software in this article, not toothpaste and driving. I knew there was a reason I was posting this to a developer site...

Now, to get Technical

I've spent the better part of the last year thinking about multithreading problems, and I could probably spend the rest of my career doing the same (although it probably wouldn't be a long career since my head would pop off before too long). There are so many issues in this ugly space that to write about all of them would take too long. So for the purposes of a focused blog (although it's probably too late for that goal), let's just focus on one area of trouble for Java client developers: multi-threaded graphics.

When developers first discover the joys of multithreaded programming, it's like opening a new present on Christmas; the power of Java thread creation is that it is so easy to create and use new threads, that you can start taking advantage of multithreaded processing in a much easier fashion than you ever could before in previous programming languages. And the knowledge that your application can perform multiple tasks simultaneously, especially on systems with multiple processors or hyper-threaded cores, is huge. You no longer need to block on IO, waiting for the system to open a file. You no longer need to hang in your display code waiting for an image to downloaded over the network. You no longer need to hang the GUI while calculating a complex set of equations. In all of these cases, it is trivial to spawn a new thread to perform the bottlenecking operation, and use the result of that operation later as appropriate.

In fact, Java builds in much of this functionality for you, so you don't even need to do the work of multithreading some of your operations which may be blocking. For example, the old Toolkit Image loading methods automatically load images on a separate thread. The following code:

	Image img = Toolkit.getImage("duke.gif");

"Well heck," the Happy Developer says. "If two threads make my application that much faster, imagine how fast it'll be with ten!"

For example, suppose a developer finds that one of the more costly operations in their application is some rendering operation, like drawing some complex Shape. Every frame they have to draw numShapes of these shapes, like so:

	// Assume myShapes[] and numShapes are initialized appropriately elsewhere...
	public void paintComponent(Graphics g) {
	    for (int i = 0; i < numShapes; ++i) {
	    	g.draw(myShapes[i]);
	    }
	}

"Golly!", the Happy Developer says, "What if I use that cool Thread mechanism to speed this up?! Then it'll go way faster." They might write something like the following:

	class ShapeRenderer implements Runnable {
	    Graphics g;
	    Shape s;
	    ShapeRenderer(Graphics g, Shape s) {
	        this.graphics = g;
	        this.shape = s;
	    }
	    public void run() {
	        graphics.draw(s);
	    }
	}
	public void paintComponent(Graphics g) {
	    for (int i = 0; i < numShapes; ++i) {
	        Thread t = new Thread(new ShapeRenderer(g, myShapes[i]);
		t.start();
	    }
	}

They compile the code, run it hopefully ... and discover that it didn't fix their performance problem. In fact, they discover that the app is actually much slower than the original code.

What gives?

There are actually several different factors that can contribute to the performance of this particular approach, ranging from things that add no benefit to those factors that actually make multi-threading this application slower. Let's go through some of these factors, one by one:

1) Thread/object creation overhead
Okay, so we've all heard the mantra that object creation and destruction (and the associated garbage collection process) is actually pretty fast. Memory allocation and destruction on current garbage collection systems is almost free, in fact; it costs only the instructions involved in moving a pointer around (memory is assigned by reserving space on a pre-allocated heap, and is destroyed by another pointer move. Of course, there are more details here about how garbage collectors, but suffice it to say that simple creation/destruction of temporary objects is very, very cheap).

However, that doesn't mean to say that creating temporary objects is actually free. In particular, it doesn't mean that you want to create and initialize objects in your inner loop if you don't have to. In the above example, we are not only asking for temporary memory for the Thread and ShapeRenderer objects (a cheap operation), but we are also asking that those objects get created and initialized, which may not be so cheap, depending on the complexity of the objects involved and whatever initialization process they need to go through. In this case, the creation of a thread will probably involve a fair amount of processing at either/both the Java and native levels in order to create the underlying thread object.

The Happy Developer, realizing this, will of course take steps to minimize the temporary object creation. In this situation, they may realize that since the same thing happens every time through the loop, there is no reason that the applications needs to create the Thread and ShapeRendering objects every time through; they can just incur the overhead of creation one time and then reuse these objects whenever we need them.

I won't bother with an example here, just picture a variation of the above where the Thread and the ShapeRender objects are created only once. Then, inside the paintComponent() method, we need only update the Graphics object of each ShapeRenderer and then tell each Thread to do its thing.

Once again, the Happy Developer (this time with a slightly less huge smile of anticipation on their face) awaits the stunning results ... but discovers that this new variation is still worse than the original approach.

Things may be a bit better here than before; at least the application is not going through the contortions of creating and initializing the Thread and ShapeRenderer objects every time through the painting loop. But there's still something amiss in the messy threading details.

2) Thread Swapping

One of the hidden details of multithreaded programming is that the operating system has to go through a fair amount of work in order to run a separate thread. This is not much in the whole scheme of things (less than a millisecond, certainly), but it can add up when there are several threads involved.

For example, if you have ten threads all trying to do similar tasks at the same time and at the same priority, then the system will keep swapping the threads in and out trying to get the work done. This may not be as bad as swapping out each thread after just a couple of instructions in some round-robin fashion; we may get a fair chunk of work done in any given thread before we are swapped out. But the amount of work accomplished on that thread must be weighed against the work done to swap threads to know whether it was worthwhile having multiple threads to accomplish the task.

I wrote a test to see what thread-swapping overhead was. I ran in a tight loop, calling wait/notify to swap the threads back and forth. For 100,000 swaps, it took 1.3 seconds on that particular test system. This doesn't sound like much, but if you can imagine each thread trying to perform something simple like drawing a single line, the fact that we could only do 100,000 of these operations in that 1.3 seconds makes the thread swap overhead seem pretty significant. (For comparison purposes, I also timed calling a function 100,000 on the same system, which took only about 10 ms).

The cost of thread swapping overhead comes into play especially on systems where there are less computing resources available than there are threads than want those resources. This is an excellent, if obvious, segue into my next point...

3) Limited Thread Resources

The ideal case for multithreaded systems is having one processor per thread, or at least one processor available whenever a new thread needs processing power. For example, if you have a four CPU system and there are four threads all trying to run at the same time, they can each have a processor to themselves and get along swimmingly. There need be no thread-swapping overhead, as mentioned in point #2 above, because the threads do not have to be swapped out; they have full control over their CPU (at least while the process is running).

There are now hybrid systems where single chips can have multiple resources available for threads, such as the Hyper Threaded CPUs of various chips today. While these systems cannot dedicate an entire CPU to a particular thread, they have enough resources on any CPU to dedicate some of those resources to separate threads. There is still some overhead of thread swapping and contention here, but it is at least better than the single-CPU model.

The problem with the sample application above is that it does not take into account anything about the system when it creates a thread per Shape. Unless the application is running on a system that has as many CPUs (or at least an many thread resources, in the hyper threading case) as there are threads, then there is going to be contention in thread processing and thus overhead for swapping threads in and out.

Our Happy Developer might realize that the available thread processing resources could be a bottleneck. Suppose they know that, in general, their application will run on systems with at least 2 processors or one hyper-threaded processor, and they would still like to take advantage of that capability in multithreading their application. They may then change their app to use a model of exactly two rendering threads instead of the one-per-Shape model above. Now, instead of sending each Shape rendering operation through its own dedicated thread, it will queue up these operations on two separate threads. The code will be more complex for this approach, but still fairly straightforward. For brevity, I'll skip the example, but hopefully it's easy to picture this thread-sharing approach.

Again, the Happy Developer awaits patiently the results they know will stun the world. There is a slight faltering of their smile this time; they have met defeat too many times in the past. Still, they look forward to ... more failure.

Once again, the application fails to improve upon the original single-threaded approach. This time, they have eliminated much of the overhead in dealing with threads. And when running on a system than can process multiple threads simultaneously, they may even have eliminated thread-swapping overhead (or at least reduced it significantly).

Given the power of doing multiple tasks simultaneously, and the ability of the system to handle this simultaneity, what happened?

4) Graphics is Inherently Single-Threaded

Here's the sad reality of today's computing platforms; graphics hardware is inherently single-threaded.

This single-threaded approach is so ingrained in today's platforms that it is an assumption at the hardware, the driver, and even the API level (although some APIs are written to handle multi-threaded programming, they do not do a good job of compensating for the limitations below them and the hardware and driver level)..

While the hardware architectures, processors, operating systems, and languages have evolved to allow and encourage multi-threaded programming, the underlying graphics engines simply cannot do it.

This means that you may be able to easily write an application that performs graphics operations in multiple threads. And the system you are using (e.g., Java) may turn around and issue those graphics calls in separate threads. And the underlying systems that Java depends upon (e.g., X, DirectX, GDI, whatever) may be able to receive those calls from multiple threads. But when it finally gets down to the hardware, it has all been funneled through one pipe and there is no way to get any advantage by trying to use multiple threads at a higher level.

Just like the chain that is only as strong as its weakest link, an application is only as multithreaded as the systems it depends upon; in this case, if the graphics subsystem is single-threaded, then there is nothing you can do at the upper layers to make that system more multi-threaded-friendly.

Let's take a look at an example. I was working on this one just this week, playing around with various options in double-buffering. I wanted to play with the idea of writing to a buffer on one thread and copying that buffer to the onscreen window on a totally different thread. There were various reasons for this (and various possible gotchas), but it was at least worth an experiment.

I wrote my rendering loop to fill the buffer as fast as it could, with simple calls that mimicked a scrolling operation (copy part of the buffer to itself in a different location, fill in the rest with some color). After each operation, it would update a flag that told the system that the buffer contents were new (and should thus be copied to the onscreen window at some time).

I wrote my buffer-copying thread to occasionally wake up (every few milliseconds) and copy the buffer to the screen if the buffer had been changed since the last copy.

There are a few implementation details here (such as synchronizing on the update flag variable), but I have described the essential bits.

What I saw confused me at first. It ran pretty well on my development system. So I had someone else run it on their system. In that new environment, it basically froze the window for several seconds. It looked like we were not even getting our screen- updating loop, as if the system was not even kicking off the timer I had set.

In digging into it further, I found the artifact was more disturbing. We were being woken up correctly based on the timer, and were then attempting to update the screen. But the buffer-rendering loop had so completely filled the graphics pipeline with scroll/fill calls that we basically froze the system while those were being worked on by the underlying driver and hardware.

Here was a situation where:

• the language supported the threading approach
• the underlying subsystem (in this case, DirectDraw) supported rendering to/from multiple threads
• the operating system supported multiple threads

• the graphics rendering system was created as inherently single-threaded, so the fact that one thread could so completely fill the graphics pipeline meant that future threads that the operating system wanted to swap in would simply have to wait for the original thread.

The upshot of this whole diatribe on the graphics subsystem is that there is basically no gain to be had in changing the original sample application in the way that we did; since the underlying system is inherently single-threaded, we have nothing to gain and everything to lose by introducing potential thread overhead when everything will just end up in a single thread in the final rendering step in the hardware.

Conclusions, Thoughts, Powertool Accidents

The point of this article was to raise some issues to be aware of in multithreaded programming. Note that I am specifically not saying "Don't Do It!". There are lots of advantages to multithreaded approaches, some of them mentioned in the introduction above. For example, it is still a huge win to do time-consuming non-graphics tasks in a separate thread (such as IO or image loading) if you do not want to block the main thread (a big example in my desktop client world is the canonical "don't block the GUI thread" example; all blocking operations should happen elsewhere so that the user still sees a snappy GUI)..

And there are even cases where graphics operations can be effectively multi-threaded. For example, in a system that is using all software rendering (such as rendering destinations that are not hardware-accelerated, such as BufferedImage objects) running on systems that support multiple threads, there could be a huge win in having separate rendering threads. Imagine a multi-chip server that is producing separate images all in parallel, rendering to each with software loops, running each of those loops on separate processors; this is a major win for parallelism.

But what I am saying is: be aware of the issues in the platform and underlying systems when taking a multithreaded approach. And always test your application to see if you actually got the speedup you were anticipating.

Think of multithreaded programming kind of like a chainsaw. It can be an incredibly powerful tool that can dramatically reduce the time needed to perform some tasks. Or it can chop your hand off and cause a huge mess that's impossible to recover from. You need to know how to use it effectively to determine which one it will do for you.

ToolkitBufferedVolatileManagedImage Strategies

A common question seems to arise often from Java graphics developers about which image type or creation method to use. When exactly should you use VolatileImage? What is BufferedImage appropriate for? What about the old Toolkit images? And when is BufferStrategy more appropriate than one of these image types?

It's a pretty big topic, and the answer (like all truly great answers) is probably "It depends". But there are some general guidelines that can come in handy. And perhaps a description of what these different kinds of images and methods are all about might help.

1) Image Types

First of all, perhaps a short dictionary of image types might help:

• Toolkit Image: This is the oldest kind of image object in the Java API. These images are created and returned by the old 1.x APIs such as Applet.getImage() and Toolkit.createImage(). These images are created from a pointer to a data source, such as a GIF or JPG file, and return an object of type Image. They are useful for convenient loading and storage of image data for display, but getting at the actual pixel data or manipulating it is not as easy.
• BufferedImage: This is an image type that was created in the jdk1.2 API. They were created for easier and more powerful manipulation of the actual pixel data in an image. At first, there was no way to load an image from a data source directly into a BufferedImage; these images were used, instead, to create an arbitrary buffer for pixel data, which you could then write to, read from, or display conveniently. The main way to get actual pixel data into a BufferedImage object at first was through use of rendering operations (after getting a Graphics object for the BufferedImage), or by manually setting the pixel data through methods in BufferedImage, WritableRaster, and DataBuffer. With the advent of the ImageIO API (see below) in jdk1.4, it became possible to create a BufferedImage object directly from a data source, just like Toolkit images (only these BufferedImage objects are writable, unlike their Toolkit image cousins).
• VolatileImage: This image type was created in jdk 1.4 as a means of creating and managing accelerated image memory. One of the problems with hardware acceleration for images is that, on some platforms, accelerated memory can be deleted out from under you at any time. This is obviously not what you want for your typical image data. To work around that, the VolatileImage API was created to provide a notification mechanism so that you know when an image must be re-rendered due to data loss. VolatileImage objects are not loaded from image data, but are just created as empty pixel buffers (much as the initial BufferedImage objects were (see above)); to get loaded image data into a VolatileImage, applications must load the image data through some non-Volatile means, get the Graphics object for the VolatileImage, and then copy the data into the Graphics object using drawImage().
• Managed Images: These image objects are not specific objects or APIs in Java, but are rather a concept of how we accelerate image operations. A "managed image" is one that you create through any of the normal image creation/loading methods and which we try to accelerate for you internally, by creating an accelerated mirror copy of the image data. This type of image can benefit from hardware acceleration without falling prey to the "surface loss" issues mentioned above for VolatileImage. I'll talk more about managed images and hardware acceleration later in the article.

That's it for the basic image types. Now let's talk about how we actually create and use these image objects.

2) Who you gonna call?

Whenever I want to give myself a fright about the complexity of our APIs, I simply ponder the vast array of choices that face developers who simply want to create an image. I'm sure I'm missing some here, but let's see...

Applet:
	getImage(url)
	getImage(url, name)

BufferedImage:
	new BufferedImage(colorModel, raster, premultiplied, properties)
	new BufferedImage(width, height, type)
	new BufferedImage(width, height, type, colorModel)

BufferStrategy:
	new BufferStrategy()
	(Note: This method cannot be called directly)

Canvas:
	createBufferStrategy(numBuffers)
	createBufferStrategy(numBuffers, capabilities)
	
Component:
	createImage(imageProducer)
	createImage(width, height)
	createVolatileImage(width, height)
	createVolatileImage(width, height, capabilities)

GraphicsConfiguration:
	createCompatibleImage(width, height)
	createCompatibleImage(width, height, transparency)
	createCompatibleVolatileImage(width, height)
	createCompatibleVolatileImage(width, height, capabilities)

Image:
	new Image()
	(Note: This method cannot be called directly)

ImageIO:
	read(file)
	read(imageInputStream)
	read(inputStream)
	read(url)

ImageIcon:
	new ImageIcon(imageData[]).getImage()
	new ImageIcon(imageData[], description).getImage()
	new ImageIcon(image).getImage()
	new ImageIcon(image, description).getImage()
	new ImageIcon(filename).getImage()
	new ImageIcon(filename, description).getImage()
	new ImageIcon(url).getImage()
	new ImageIcon(url, description).getImage()

Toolkit:
	createImage(imagedata[])
	createImage(imagedata[], offset, length)
	createImage(producer)
	createImage(filename)
	createImage(url)
	getImage(filename)
	getImage(url)

Window:
	createBufferStrategy(numBuffers)
	createBufferStrategy(numBuffers, capabilities)

I'm sure there's more out there, especially using things like ImageIO (which is all about reading and writing images, as you might guess from the name...). But this list will do for now.

So it's a wrap. This article's pretty much finished; just use the above API calls to create your images. Left as an exercise to the reader. Q.E.D. It's obvious, isn't it?

Okay, so maybe it isn't obvious; there are a lot of methods above that all seem to need different parameters or that create different types of images.

Here's the trick: All of the above image creation methods (and any others that are not on the list) can be broken down into just a few categories. Then the plethora of ways of creating an image in one of those categories can just be seen as utility methods; different ways of getting the same result. The convenience methods may be because of logic (why do I have to get the GraphicsConfig to create an image associated with a Component? Why not use the Component directly?), or convenience (instead of using some InputStream mechanism for all image readers, we provide several ways to read the image directly including from filenames, URLs, and streams; just call the method appropriate for your situation).

So the real work in this article is to break down the categories of image types and describe which types of images and methods you may want to use in which situations. Once you get that down, the rest, as they say, is just implementation details.

3) Image Loading or Creation?

First of all, are you loading existing image data? Or are you creating an image buffer in memory? Image loading means that you have image data (either locally or across the network) that you want to load into your application, possibly to copy that image onto the screen or to read and operate on the data. Image creation means that you want some arbitrary image memory created for your application; perhaps you want to create a buffer for double-buffered animations, or you want a place to cache intermediate filtering results.

3.1) Image Loading

In the above method list, all of the methods that take filenames, urls, streams, producers, and data arrays are those intended for loading existing images. In particular, all of the methods listed above for Applet, ImageIO, ImageIcon, and Toolkit are intended for image loading:

Applet:
	getImage(url)
	getImage(url, name)

ImageIO:
	read(file)
	read(imageInputStream)
	read(inputStream)
	read(url)

ImageIcon:
	new ImageIcon(imageData[]).getImage()
	new ImageIcon(imageData[], description).getImage()
	new ImageIcon(filename).getImage()
	new ImageIcon(filename, description).getImage()
	new ImageIcon(url).getImage()
	new ImageIcon(url, description).getImage()

Toolkit:
	createImage(imageData[])
	createImage(imageData[], offset, length)
	createImage(producer)
	createImage(filename)
	createImage(url)
	getImage(filename)
	getImage(url)

There are at least four major things that differentiate these methods:

• the location of the image data
• the format of the image data
• the synchronous or asynchronous behavior of each method
• the type of image that is created (Image or BufferedImage)

When I'm talking about location, I'm mainly concerned with whether the file is local or across a network. Also, if it's packed into some resource file, such as a jar file, that also comes into play here.

Loading across the network

If you are accessing the data across a network, it's probably easiest to use the URL variations:

	Applet.getImage(url)
	Applet.getImage(url, name)
	ImageIO.read(url)
	new ImageIcon(url).getImage()
	new ImageIcon(url, description).getImage()
	Toolkit.createImage(url)
	Toolkit.getImage(url)

	URL dukeURL = new URL("http://java.sun.com/duke.gif");
	BufferedImage img = ImageIO.read(dukeURL);

Now suppose you have another image that you have saved locally in a file; just use the filename variation of the above. For example, let's say you loved one particular instantiation of the lovable-yet-quirky duke.gif file above so much that you downloaded and saved it for use in your application (see the above note on scary lawyers). Then you could use the following code to load that file from the directory where the program was launched:

	BufferedImage img = ImageIO.read(new File("duke.gif"));

	URL dukeURL = getClass().getResource("media/" + "duke.gif");
	BufferedImage img = ImageIO.read(dukeURL);

Another consideration is the format of your stored image. The old Toolkit/Applet loaders only understand GIF, JPEG, and PNG format files. (Okay, they also understand XBM and XPM2, old X11 image formats, but those are probably not formats you are terribly concerned about). These loaders works well for most web applications since these image types are traditional web image formats. But what if you have an image in some other format that the Toolkit/Applet loaders do not understand?

ImageIO currently has built-in readers for GIF, JEG, and PNG. In addition, it will have BMP and WBMP capability in the jdk1.5 release. Moreover, there will be more image readers/writers for ImageIO going forward, whereas there are no specific plans to support more formats for the old Toolkit/Applet loaders. And finally, ImageIO has a pluggable reader API, so if you have a custom image format, or some other format not yet supported by the core library, you can write your own loader for that format within ImageIO.

In fact, the JAI team has made available a package with additional ImageIO readers/writers at http://java.sun.com/products/java-media/jai/downloads/download-iio.html if you have requirements beyond the current ImageIO defaults.

So ImageIO could also be the right choice if you need to deal with formats beyond the basic web image formats.

3.1.3) Synchronicity

The Applet and Toolkit image loading methods came from the old days of Java 1.0, when Java was seen primarily as a networked application API and image data might come from any source, potentially one on an unreliable or slow network connection. To make networked applications more robust, it is reasonable to put network-dependent operations in separate threads to ensure that an application's main or GUI threads do not hang while waiting for a slow download. Because this was a common pattern for Java GUI applications at that time, the image loading operations were all created to run on a separate image loading thread. Thus when an application calls:

	Applet.getImage(url)

Note that this model of asynchronous loading does not apply solely to networked applications, or even to image loading specifically; any operation that takes a significant amount of time should not be done on the GUI thread, lest you run the chance of making your application appear hung while the operation is taking place. So, for example, if you are loading in a huge image from a local file, you may want that non-networked operation to happen in a separate worker thread to ensure that your GUI has no pauses during image loading.

This model works well enough for applications that create their images early for later use. The application simply may need to check whether the image has been loaded whenever it is required in the application.

When applications do need the data (for example, if they need image sizes in order to determine layout correctly, or if they need to display images in their final form), they may need to synchronize on the image loader and wait until the image loading is done. For example, an application may want to load local image data and be willing to wait for that data to load before proceeding (knowing that a local load will usually not take very long). In that case, the application might do something similar to the following:

	public Image loadImageSynchronously(Component comp, String filename) {
	    URL url = getClass().getResource("media/" + filename);
	    Image image = Toolkit.getDefaultToolkit().getImage(url);
	    MediaTracker tracker = new MediaTracker(comp);
	    int id = 0;
	    tracker.addImage(image, id);
	    try {
		tracker.waitForID(id);
	    } catch (InterruptedException e) {
		System.out.println("INTERRUPTED while loading Image");
	    }
	    return image;
	}

	public Image loadImageSynchronously(String filename) {
	    URL url = getClass().getResource("media/" + filename);
	    Image image = new ImageIcon(url).getImage();
	    return image;
	}

Meanwhile, ImageIO has synchronous loading methods that do not return until the image has been loaded and is ready to go. Note that some applications and situations may still need asynchronous loading behavior (for long image loads or to more efficiently multitask). For example, it does not take a huge amount of time to affect perceived GUI performance, so if an image load will take even as long as a tenth of a second, you may want to avoid loading that image synchronously on the Event Dispatch Thread (so don't load it in your paint() method). You can always spawn a new Thread yourself to call the ImageIO loading methods if necessary.

3.1.4) Resulting Java Image Type

Part of the decision over which image creation API you is in which image type you want to get back from the creation method. In particular, do you want a Toolkit Image or a BufferedImage?

Toolkit Images are created by the Applet, Toolkit, and ImageIcon methods listed above. The resulting images are easy to use for display purposes (just call drawImage(...) from this image into a Graphics destination and the image will be copied appropriately), but lack the power of BufferedImage objects for manipulation of the data.

BufferedImage objects are created by the ImageIO methods listed above. These objects offer a more powerful API, albeit with potentially more work involved to do some operations (such as displaying an animating GIF image).

Image or BufferedImage: What's in a Name?

Although both Image and BufferedImage have similar properties in terms of being displayable, BufferedImage has many more capabilities. For one thing, the Image objects created by the Toolkit, Applet, and ImageIcon load methods are read-only; you cannot get the Graphics of those Images and render to them. So if you want to modify the image data, you will need to do more work (such as creating another image that is modifiable and copying the loaded Image into that new image). Image has some very simple methods and is mostly intended to be a simple object that holds image data. But BufferedImage has many methods for modifying and extracting all kinds of data from an image; color models, pixel data, and more. Given a choice between the two, I would always opt for the one that gave me more power and flexibility.

But doesn't that increased capability mean increased overhead? Not at all; there is no extra processing involved in BufferedImages when these other powerful methods are not used. If all you do is load and image and display it, BufferedImage can do this just as easily as the more streamlined Image object.

So go ahead and use BufferedImage. It is, after all, better than butter.

Dirty Laundry

One good (and not entirely obvious) reason for using the ImageIO API for loading images is the unfortunate reality that the code is simply newer, cleaner, and more maintained (both now and in the future). Much of the old Applet and Toolkit image code was written years ago and has many assumptions and situations that it must account for and is therefore tricky to maintain and upgrade.

Our future image reading/writing direction is with ImageIO; yours should be too, because that's where the focus of our efforts will be in the future.

Having said all that wonderful stuff about ImageIO, there could be situations in which the old Toolkit/Applet/ImageIcon approach makes more sense for your particular application, including:

• Performance: Ideally, ImageIO would be more performant than the old code. But some code paths in ImageIO have not yet been optimized to the extent that the old code was, so you may find some situations (such as loading images in a particular format) which perform better using the old APIs than they do with ImageIO. The problems are (mostly) known, the bugs have (mostly) been filed, and these performance gaps will (definitely) be fixed. But if performance in these situations is important for your application, by all means use the old methods in the meantime if they work better for you.
• Ease of Use: One of the goals of ImageIO was to expose more capabilities to the developer. There is much information, such as metadata, in an image that is not exposed through the old image loading methods or the old Image class. The flipside of this approach (at least in the current API) is that it can be more work to do trivial tasks. A good example of this is loading and displaying an animated GIF.
  In the old image loading APIs, if you loaded an animated GIF file, you did not need to worry about the details of animating that image. Instead, you could just load the image and display it and we would automatically schedule the animations of that image appropriately. You could not access the information about the animation (how many frames? what animation rate? loop? halt?), but making it animate was quite simple.
  ImageIO, on the other hand, exposes all of that animation information to you through the API, but you must actually use that information to manage the image animation yourself; you cannot just tell us to draw the image and have us manage it for you. This is a known issue and we will do something about it Real Soon Now; look for an upcoming article about doing this, and look for future API changes to make this simpler in the future.
  
• Compatibility: Sometimes applications need to be as backward-compatible as possible across most or all jdk releases. In particular, some applications need to be able to run on old jdk1.1 or earlier releases. For example, applet games tend to need ubiquity of the runtime environment and cannot count on browsers having any later version of Java than the 1.1 APIs shipped with the MSVM. In this case, you may need to use the older APIs simply so that they will run across all of these platforms.

Note also that if you need to use the old APIs for some reason but you still want the power and flexibility of BufferedImage, it is easy enough to load the images in through whatever methods are appropriate, create a new BufferedImage object, and then simply copy the loaded images into the BufferedImage. For example:

	Image img = new ImageIcon(filename).getImage();
	BufferedImage bImg = new BufferedImage(img.getWidth(null),
					       img.getHeight(null).
					       BufferedImage.TYPE_INT_RGB);
	Graphics g = bImg.getGraphics();
	g.drawImage(img, 0, 0, null);

3.1.5) Hey! What about the other loading methods above?

The approaches above cover most of the loading methods I listed, but some are notably skipped. The *Stream methods of ImageIO are simply variations on a theme; if you happen to have your data in that format (versus a URL or file), go for it; it's just a convenience to use these alternatives.

As for the other skipped methods (one using an ImageProducer and some using data arrays), I hoped you wouldn't notice....
These other loading approaches are somewhat dated and come from the old days of Java image processing. There are still some situations which might require the ImageProducer/ImageConsumer approach, but in general it should be easier, more straightforward, and more performant to use the newer APIs such as ImageIO and BufferedImage. The uses and misuses of the older Consumer/Producer/PixelGrabber APIs could stand a separate article all by themselves. I'll try to tackle this one in the future.

As far as reading the image data from an array of data (see the methods above with the imageData[] parameter), this is really only appropriate if you've already read the data into the array to begin with. This could be necessary if you have some custom image storage mechanism, such as a database. But if the image existing in a regular file/URL/stream format, you should probably be using one of the other loading methods instead.

3.2) Image Creation

What if you do not have an existing image on the network or file system? What if you just want a buffer of pixel data that you can use in your application? This could be for creating sprites or icons with rendering calls instead of loaded image information (perhaps you've found this to be faster in your situation than reading image files). Or it could be a buffer that you can use for caching intermediate results or for providing double-buffered rendering for an animation.

For the purposes of this discussion, I'll break down this category of images into three types:

• Static Images: These are images that you render to infrequently (perhaps just once, when you create the image). The images are mainly used to copy from.
• Dynamic Images: These are images that you render to often, like an animating sprite.
• Back Buffers: These are like dynamic images, in that they are rendered to frequently (usually at least once per frame), but they are specifically intended to provide buffering support for the onscreen window, and are thus usually copied from often as well, like once per frame.

3.2.1) Static Images

Static images are ones that are created and rendered to once (or infrequently) but probably copied from often. Examples of this type of image include icons for a GUI or sprites for a game.

The best approach for this type of image is to create an image that is in the same format as the image or window that the image will be copied to; this ensures the most straightforward copy mechanism since the underlying software will not have to perform a conversion on the image data while copying to the destination.

You could, of course, create a BufferedImage object manually through one of its constructors; you could query the GraphicsDevice for its display information and then create a BufferedImage of the appropriate type:

	new BufferedImage(colorModel, raster, premultiplied, properties)
	new BufferedImage(width, height, type)
	new BufferedImage(width, height, type, colorModel)

But why go to the hassle of all of that when there are convenience mechanisms that do all of this for you? Specifically, take a look at:

	Component.createImage(width, height)
	GraphicsConfiguration.createCompatibleImage(width, height)
	GraphicsConfiguration.createCompatibleImage(width, height, transparency)

The best part about static images is that you can use very simple means to create the images and then we will try very hard internally to see that you get any available hardware acceleration for these images when they get copied around. We call these "managed images", because we manage the acceleration details for you. For more information on managed images, please see my blog on BufferedImage performance.

Note that we currently (in all jdk1.4.* releases) manage images that are created with the above APIs and some of the Toolkit image loading methods described previously, but in jdk 5.0 (available in Beta form now, and full release soon) we manage nearly all types of images and take advantage of hardware acceleration if it exists. So go ahead and create the type of image that is most convenient for you and we'll try to do the right thing under the hood.

3.2.2) Dynamic Images

This kind of image may be rendered to quite often, as in an animating icon, or a sprite that is modified on a frequent basis. You could certainly use the same image-creation APIs listed above for static images; these will work fine in most situations and are certainly the easiest way to go in general.

However, some developers interested in maximizing performance may want to know more about image management and how dynamic images can affect it.

We manage images by detecting when the application is copying from an image to a destination (either another image or an onscreen window) that lives in accelerated memory. If this copy is done successively when the source image has not changed, then we may decide to cache a copy of that image in accelerated memory and perform future copies from this cached version.

In the case of a dynamic image, if that image is being updated one or more times for every copy to the destination, then we will never create an accelerated version of it, and thus the image will never benefit from any hardware acceleration that we could otherwise provide.

(Aside: For the insatiably curious, the reason for this oddity in acceleration comes from "surface loss", where an accelerated version of an image may simple go away at any time due to operating-system or user-caused situations. To keep the original image data intact, we store the main image data (that which is modified by the application) in an unaccelerated location, and only accelerate a mirror copy of that image. That way, if the accelerated version gets wiped out, we still have the original data from which we can create a new accelerated copy. The problem here, in terms of performance, is that an "unaccelerated image" means that all rendering to and from that image is unaccelerated. And if an application is constantly modifying the image, all of that rendering will be unaccelerated and it is never appropriate for us to create and use an accelerated version of that image.)

Developers that care about top performance for these types of images may want to look into using VolatileImages instead. These images store the data in accelerated memory (when possible) and thus rendering to and from that image may be accelerated automatically. The downside is that these images require a bit more care and feeding, due to "surface loss" issues that arise with current video memory architectures and operating systems. Note that not all types of rendering to these images is accelerated, either, but simple types of rendering like lines and rectangular fills and copies can usually be accelerated, depending on the platform configuration.

I've already written about VolatileImages in past blogs (Part I and Part II), so I will not go into the details of their usage here; please check out those other articles for more information. But it is worth covering the APIs used to create the images, just for consistency's sake in this article:

	Component.createVolatileImage(width, height)
	Component.createVolatileImage(width, height, capabilities)
	GraphicsConfiguration.createCompatibleVolatileImage(width, height)
	GraphicsConfiguration.createCompatibleVolatileImage(width, height, capabilities)

	GraphicsConfiguration.createCompatibleVolatileImage(width, height, transparency)
	GraphicsConfiguration.createCompatibleVolatileImage(width, height, capabilities, transparency)

The use of the ImageCapabilities object in these methods gives you the ability to require certain attributes (such as hardware acceleration) from any image created with that method. In general, you probably will not need to use that variation, although as we enable more hardware acceleration features in our platform, we may expand the ImageCapabilities API to be more powerful and useful. (Note, too, that ImageCapabilities can be used effectively as a means of inquiring what capabilities an existing image has).

3.2.3) Back Buffers

By "back buffer" I mean an arbitrary offscreen image that is created for use in a double-buffering situation. Typically, an application that wishes to have smooth graphics, especially animations, will draw to a back buffer and then copy that buffer onto the screen instead of drawing directly to the screen. Swing does this by default, so that you do not see the various GUI elements in a Swing app flash as they are drawn to the screen. The buffer copy in these applications typically happens so fast that the graphics in the application are perceptibly smoother than if they were drawn one-by-one directly to the screen.

A developer could use any of the above static or dynamic image APIs that I listed for creating a back buffer. However, the following things should be taken into account when doing so:

• Use of managed images for buffers has the same performance implications as dynamic images; we will be unable to manage the images effectively, thus you will not be able to take advantage of any hardware acceleration either for rendering to or copying from the buffer. Since the copy from the back buffer to the front buffer involves a lot of pixels, being able to accelerate that operation can have a huge performance impact on your application.
• Use of VolatileImages for buffers is a better way to go in terms of hardware acceleration potential. But why go to all the bother of managing your VolatileImages when there is an easier route?

BufferStrategy: the preferred way of buffering in Java

In jdk1.4, we introduced the BufferStrategy API, which is a wrapper around VolatileImages. This API allows you to ask for an accelerated back buffer and avoid having to manage the details of surface loss associated with VolatileImages. It also ensures that you will get a buffer of the optimal type for your application. In particular, you will get either a FlipBuffer (which can only be used in fullscreen-exclusive mode on Windows) or a BltBuffer (which is used by default for windowed applications). A FlipBuffer performs a swap of the front and back buffers in video memory when you request BufferStrategy.show(). A BltBuffer copies the contents from the back buffer to the front (just as you would if you called drawImage() from a VolatileImage back buffer to the front buffer).

With this API, there is little need to create and manage VolatileImages directly; just let us manage the details for you inside the BufferStrategy implementation.

For more information on BufferStrategy, check out the javadocs; they're pretty clear on how the system works.

The APIs you will need when creating a BufferStrategy are:

	Canvas.createBufferStrategy(numBuffers)
	Canvas.createBufferStrategy(numBuffers, capabilities)
	Frame.createBufferStrategy(numBuffers)
	Frame.createBufferStrategy(numBuffers, capabilities)

4) Wrap-Up

So that's pretty much it. You have image loading methods and image (or buffer) creation methods. And in each category, you have various flavors depending on the location and type of the data, and the type of image you want returned to you. So even though there are a lot of methods listed at the top of this article, they all break down into just a few comprehensible categories and can be used effectively, once you understand the implications of each variation.

Although there is certainly more complexity here than we can cover with a simple table, it might help to break down some of the basic attributes of the image types we have talked about and the reasons to consider one type over another when writing your application:

• Toolkit Image:
  • Creation methods: Applet.getImage(...), Toolkit.createImage(...), Toolkit.getImage(...), ImageIcon.getImage()
  • Most useful for: Simple loading and display of image data, applications that need to use 1.x APIs, asynchronous usage models (except for ImageIcon), loading performance (versus some ImageIO loading operations in the current releases).
  • Less useful for: Image data manipulation, reading/writing individual pixel data within the image, synchronous loading operations (except for ImageIcon), more varied or custom source image formats..
• BufferedImage:
  • Creation methods: new BufferedImage(...), Component.createImage(...), GraphicsConfiguration.createCompatibleImage(...), ImageIO.read(...)
  • Most useful for: Image data manipulation, read/write access to individual pixel data, synchronous loading or creation, various image data formats (both source data that ImageIO can read and Java format (i.e., TYPE_INT_RGB, TYPE_INT_RGBA, etc.).
  • Less useful for: Applications that require jdk 1.x compatibility, easy display of animating GIF images.
• VolatileImage:
  • Creation methods: Component.createVolatileImage(...), GraphicsConfiguration.createVolatileImage(...)
  • Most useful for: Hardware acceleration of dynamic images, more fine-grained control over acceleration properties.
  • Less useful for: Simple image creation and display (applications must use the VolatileImage API to handle surface loss issues)
• BufferStrategy:
  • Creation methods: Canvas.createBufferStrategy(), Window.createBufferStrategy()
  • Most useful for: Easy creation and management of window buffering schemes
  • Less useful for: Access to actual image data of buffer

4.1) Hey! You forgot some methods!

There are still a couple of the creation methods up top that I have not covered yet:

BufferStrategy:
	new BufferStrategy()
Image:
	new Image()

4.1) What About Performance?

It is difficult or impossible for me to write a long block of text or code without thinking about performance. And since some of the users of the APIs above, and image operations in general, care a great deal about performance, I should spend a few words discussing some performance issues to be aware of.

Again, check out my blogs on BufferedImage ( Part I and Part II); I go into much more detail on image management there. Some important things to keep in mind with respect to managed images (and making sure they are benefiting from available acceleration):

• If you want copies from an image to be accelerated but you will be modifying the image frequently, think about VolatileImage instead (see the above discussion on Dynamic Images).
• Careful with that Raster, Eugene! This is a hidden gotcha in the current implementation. Part of the BufferedImage API is a method that allows you to get the Raster of the image (getRaster()). From that Raster, you can create a writable child raster (WritableRaster.createWritableChild()) or get the internal data array for the raster (Raster.getDataBuffer()). Both of these methods end up giving you a copy of the data that we have no control over. Specifically, we cannot tell when you have updated the contents of the image, so we cannot tell whether our cached/accelerated version of the image needs to be updated. So, when you call these Raster methods, we throw up our hands and give up on acceleration entirely.
  If you need to access the pixel data but you still want us to accelerate image copies, call methods which manipulate the data for you without giving you a reference to the data; e.g., BufferedImage.setRGB() and WritableRaster.setPixels() for setting arbitrary pixel values, or any of the Graphics.draw*() or Graphics.fill* methods for setting blocks of pixel values.
• If you are transforming the image while copying it around (say, scaling it to the destination, or rotating it), we will not be able to do that using hardware acceleration with the current implementation. If you anticipate using a particular transformation of an image (like a zoomed-in view or several rotations of a sprite) multiple times, then it might be worth creating a temporary version of that transformed image and letting us manage (accelerate) that transformed version for you.

4.2) What about 5.0?

Most of the APIs I discussed above are pre-5.0, so you can use everything above (except where noted) in the current releases available for download. If you are looking forward to using 5.0 (available in beta form today, in full release Real Soon Now), then I'll mention a couple of tweaks to the above:

• VolatileImage APIs: It will be possible in 1.5 to create a VolatileImage with a transparency parameter. Developers have been asking for that capability for some time; the original createVolatileImage() methods only allowed creation of opaque VolatileImages.
  That's the good news: the API is there. The bad news is that there is no hardware acceleration for these non-opaque images by default because our mechanisms for accelerating these types of images are still not ready for prime time. Translucent volatile images can achieve hardware acceleration by using the handy -Dsun.java2d.translaccel=true flag.
• Managed images everywhere! You will now get managed images by calling most of our image creation and loading mechanisms. In jdk1.4.*, only some of the image creation methods (such as Component.createImage() and GraphicsConfiguration.createCompatibleImage()) returned managed images. But in 1.5, you can even call new BufferedImage(...) with a type that is not in the screen bit depth and end up with an accelerated managed image.
• OpenGL Acceleration: The 5.0 release has an OpenGL-based rendering pipeline available on all platforms. This pipeline is not be enabled by default (due to various concerns including graphics hardware and driver issues), but there is a flag that you can use to experiment with it (sun.java2d.opengl=true). This pipeline provides our most comprehensive hardware acceleration so far on all our supported platforms, so it should be pretty exciting to see this one in action.

This work is licensed under a Creative Commons License.