COP5618/CIS4930 Concurrent Programming Assignment 3

$35.00

Category: You will Instantly receive a download link for .zip solution file upon Payment || To Order Original Work Click Custom Order?

Description

5/5 - (2 votes)

This assignment involves
1. implementing parallelism using the Java fork/join framework
2. Java 8 stream processing
3. determining a bound on speedup using Ahmdahl’s law
4. Instrumenting code and performing experiments
The setting will be some transformations on images using Java’s BufferedImage and related classes.
For this assignment, the following facts about
To read an image from a file using javax.imageio.ImageIO, and java.io.File:
File sourceFile = new File(sourceImageFilename);
sourceImage = ImageIO.read(sourceFile);
To write an image to a file
File outputFile = new File(filename);
ImageIO.write(newImage, “jpg”, outputFile);
Images are viewed as a 2d array of pixels, with (0,0) in the top left corner. The horizontal axis is typically
referred to as x, the vertical by y.
In order to manipulate pixels, which can be stored in many ways, you want them in the default RGB
color model, where each pixel has four 8-bit components, representing alpha, red, green, and blue
components of the pixel’s color, packed into a single integer. Alpha is the transparency. In this
assignment, we won’t do anything with the alpha component, but if you need to specify it, use 255.
The getRGB and setRGB methods defined in BufferedImage get and set either a single pixel, or a 1d
array of pixels. If necessary, these methods transform pixels to/from the sRGB ColorModel.
A ColorModel instance encapsulates the pixel representation and provides methods to extract the
individual components. For pixels that have been obtained with the getRGB method, you can use the
ColorModel returned from the ColorModel.getRGBdefault() method. Then use the getRed, getGreen,
getBlue methods to obtain the values of those color components from the pixel. To pack individual
color components into an int, create an int array containing the colors and invoke getDataElement.
ColorModel colorModel = ColorModel.getRGBdefault();
int red = colorModel.getRed(pixel);

int[] components = {red,green,blue,alpha};
int pixel2 = colorModel.getDataElement(components,0);
Alternatively, since we know that we are in the default RGB Color Model, we can just create the pixel
directly and avoid having to create the int array:
static int makeRGBPixel(int red, int green, int blue) {
int pixel = ((255 & 0xFF) << 24) /* alpha component */ | ((red & 0xFF) << 16) /* red component */ | ((green & 0xFF) << 8) /* green component */ | ((blue & 0xFF) << 0) /* blue component */; return pixel; } In this assignment, we will generally read an image from a file, get its pixels, compute new pixel values, create a new image and set its pixels to the new values, then write the new image to a file. We will use Java 8 Streams to compute the new pixels. An example which converts a color image to grayscale is given below. public static void gray_SS(BufferedImage image, BufferedImage newImage) { ColorModel colorModel = ColorModel.getRGBdefault(); int w = image.getWidth(); int h = image.getHeight(); //Get pixels from source image int[] sourcePixelArray = image.getRGB(0, 0, w, h, new int[w*h], 0, w); int[] grayPixelArray = //convert array of pixels to a stream Arrays.stream(sourcePixelArray) //combine red, green, and blue values to obtain gray value .map(pixel -> (int) ((colorModel.getRed(pixel) * .299) +
(colorModel.getGreen(pixel) * .587) +
(colorModel.getBlue(pixel) * .114)))
//make new pixel where all three colors have the same gray value
.map(grayVal -> makeRGBPixel(grayVal, grayVal, grayVal))
//convert stream to array
.toArray();
// set pixels of newImage to the grayPixelArray
newImage.setRGB(0, 0, w, h, grayPixelArray, 0, w);
}
The caller read an image source from a file, and created a matching newImage:
BufferedImage newImage = new BufferedImage(source.getWidth(),
source.getHeight(), source.getType());
There are three significant parts to this routine: getRGB, the stream computation, and setRGB.
We want to instrument the code and use Ahmdahl’s law to determine how much speedup we could
expect by parallelizing any of these three parts of the code.
First, instrument the code to measure the time taken by each of these parts. For the sake of uniformity,
a class called Timer is provided for you. This takes a list of Strings that serve as labels for various
durations in the code. Then we insert calls to the now() method to record the time. If there are N
labels, then an array of N+1 longs holds results from System.nanoTime(). The toString method returns a
String with the total time between the first and last call to now. Also, label[0] and the elapsed time in
milliseconds between the first and second calls to now(), label[1] with the elapsed time between second
and third calls to now(), etc. are printed along with the percent of the total time for each duration.
The example shown earlier could be instrumented as follows:
public static String[] grayLabels = {“getRGB”, “toGrayPixaleArray”, “setRGB”};
public static Timer gray_SS(BufferedImage image, BufferedImage newImage) {
Timer timer = new Timer(grayLabels);
ColorModel colorModel = ColorModel.getRGBdefault();
int w = image.getWidth();
int h = image.getHeight();
timer.now();
int[] sourcePixelArray = image.getRGB(0, 0, w, h, new int[w*h], 0, w);
timer.now(); //label = getRGB
int[] grayPixelArray =
//get array of pixels from image and convert to stream
Arrays.stream(sourcePixelArray)
//combine red, green, and blue values to obtain gray value
.map(pixel -> (int) ((colorModel.getRed(pixel) * .299) +
(colorModel.getGreen(pixel) * .587) +
(colorModel.getBlue(pixel) * .114)))
//make new pixel, all components have the same gray value
.map(grayVal -> makeRGBPixel(grayVal, grayVal, grayVal))
//convert stream to array
.toArray();
timer.now(); // label = toGrayPixaleArray
//create a new Buffered image and set its pixels to the gray pixel array
newImage.setRGB(0, 0, w, h, grayPixelArray, 0, w);
timer.now(); //label = setRGB
return time;
}
1. Execute this method several times and print the results. What do you notice
about the times? What does this tell you about timing programs written in
Java? You can use the serial test case in the provided HW3TestGray class to
do this, just uncomment the print statement.
2. Run your test when with as little as possible happening on your computer.
Then try again while you are doing other activities, say watching a youtube
video. What happens to the timings?
3. Consider this loop where timerData is never again referenced.
for (int rep = 0; rep < WARMUPREPS; rep++) { Timer timerData = GrayScale.gray_SS(source, newImage); } 4. What concerns should we generally have about a loop like this in a benchmark? Is it a problem in this particular case? Why not? 5. Use Amdahl’s law to compute an upper bound on the speedup you could expect from a. parallelizing only the stream processing portion of the code b. parallelizing only the setRGB and getRGB portions of the code c. parallelizing both stream processing and setRGB and getRGB portions of the code. 6. Implement the method gray_PS. This should do the stream processing step in parallel. 7. Consider parallelizing setRGB and getRGB. This cannot be done using stream processing. Which restrictions imposed by the Stream processing framework are violated by these methods? 8. Parallelize the setRGB and getRGB methods using a divide and conquer approach and Java’s fork/join framework. To do this, fill in the missing methods in the given FJBufferedImage file. This class extends BufferedImage and overrides the setRGB and getRGB methods. Thus the only thing that needs to be done to change from serial to parallel is to create a FJBufferedImage instead of a BufferedImage. Some things to think about: a. How should the 2d image be partitioned in the divide phase? By rows? By columns? Something else? Does the presence of caches affect your answer? b. What would be a sensible default threshold? Should it be a constant number of pixels? A constant number of rows or columns? Be a function of the number of threads (which can be obtained from the ForkJoinPool pool.getParallelism() method)? Experiment and choose one. 9. Run tests to determine the speedup with parallel setRGB and getRGB and serial stream processing (gray_SS_FJ(…)) and parallel setRGB and getRGB, and parallel stream processing (gray_PS_FJ(…)). COP5618 Students only Implement a class similar to Gray.java except that it performs (almost) serial and parallel histogram equalization on an image. In some images, this improves the contrast. As an example, see LotsaGators3.jpg and colorHisEq_LotsaGators3.jpg. Fill in the missing methods in ColorHistEq.java. The transformation has the following steps: 1. Get RGB pixel array from the image 2. Convert the pixels into HSB format and save into a new array. In HSB format, each pixel is represented by a float array with 3 elements for hue, saturation, and brightness. Use the java.awt.Color.RGBtoHSB method. 3. Create a histogram of the brightness, which is a value between 0 and 1. Divide the space into a given number of bins, then count how many times a brightness value falls in each bin. For example, if you had 4 bins, and 6 pixels with brightness values {.1,.2,.27,.80,.90,.2} the four bins hold [3,1,0,2] respectively. 4. Compute cumulative probability of a value falling this or an earlier bin. You can do this by computing the prefix sum of the bins, and dividing by the number of samples. For the previous example, the prefix sum is [3,4,4,6] and the cumulative probability is [3,4,4,6] [.5 , .66 , .66 ,1]. 5. Create a new pixel array by replacing the brightness value in the HSB pixel you found earlier with the value in the cumulative probability array and convert to RGB The brightness values in our original pixel array (from step 3) become [.5, .5, .66, 1,1,.5]. Use the Color.HSBtoRGB method. 6. Create a new image with your new pixels. Give an (almost) serial version and a parallel version. Implementation notes:  Call your methods colorHistEq_serial and colorHistEq_parallel.  Use streams to convert to HSB format  To create the histogram, take a stream of HSB pixels, extract the brightness element, use collect and Collectors.groupingBy. When you apply the stream map function, note that if the type converts from a primitive type to an Object, you need to use mapToObj, or from and Object to a primitive type, the mapToInt, or mapToDouble, etc. (When I did this, I used integers from [0,numBins) and let the brightness value belong to bin (int)brightness*numBins. Then the brightness level belonging to a bin is binNum/numBins.  You can use Arrays.parallelPrefix to compute prefix sums in both your solutions.  Use a stream to create the equalized pixel array in step 5.  Set the pixels in a new image.  The provided Timer class helps capture timing information. Pass an array of Strings containing labels of things that will be measured. Then instrument the code with labels.length+1 calls to the Timer’s now() method. Other methods should be self explanatory. TURN IN: Gray.java, ColorHistEq.java, FJBufferedImage.java You do not need to turn in the answers to the questions. However, some of them WILL appear on the final, so make sure to take them seriously.