I wrote a program to render a Julia Set. The single threaded code is pretty straightforward and is essentially like so:
private Image drawFractal() {
BufferedImage img = new BufferedImage(WIDTH, HEIGHT, BufferedImage.TYPE_INT_ARGB);
for (int x = 0; x < WIDTH; x++) {
for (int y = 0; y < HEIGHT; y++) {
double X = map(x,0,WIDTH,-2.0,2.0);
double Y = map(y,0,HEIGHT,-1.0,1.0);
int color = getPixelColor(X,Y);
img.setRGB(x,y,color);
}
}
return img;
}
private int getPixelColor(double x, double y) {
float hue;
float saturation = 1f;
float brightness;
ComplexNumber z = new ComplexNumber(x, y);
int i;
for (i = 0; i < maxiter; i++) {
z.square();
z.add(c);
if (z.mod() > blowup) {
break;
}
}
brightness = (i < maxiter) ? 1f : 0;
hue = (i%maxiter)/(float)maxiter;
int rgb = Color.getHSBColor(hue,saturation,brightness).getRGB();
return rgb;
}
As you can see it is highly inefficient. Thus I went for Parallelizing this code using the fork/join framework in Java and this is what I came up with:
private Image drawFractal() {
BufferedImage img = new BufferedImage(WIDTH, HEIGHT, BufferedImage.TYPE_INT_ARGB);
ForkCalculate fork = new ForkCalculate(img, 0, WIDTH, HEIGHT);
ForkJoinPool forkPool = new ForkJoinPool();
forkPool.invoke(fork);
return img;
}
//ForkCalculate.java
public class ForkCalculate extends RecursiveAction {
BufferedImage img;
int minWidth;
int maxWidth;
int height;
int threshold;
int numPixels;
ForkCalculate(BufferedImage b, int minW, int maxW, int h) {
img = b;
minWidth = minW;
maxWidth = maxW;
height = h;
threshold = 100000; //TODO : Experiment with this value.
numPixels = (maxWidth - minWidth) * height;
}
void computeDirectly() {
for (int x = minWidth; x < maxWidth; x++) {
for (int y = 0; y < height; y++) {
double X = map(x,0,Fractal.WIDTH,-2.0,2.0);
double Y = map(y,0,Fractal.HEIGHT,-1.0,1.0);
int color = getPixelColor(X,Y);
img.setRGB(x,y,color);
}
}
}
@Override
protected void compute() {
if(numPixels < threshold) {
computeDirectly();
return;
}
int split = (minWidth + maxWidth)/2;
invokeAll(new ForkCalculate(img, minWidth, split, height), new ForkCalculate(img, split, maxWidth, height));
}
private int getPixelColor(double x, double y) {
float hue;
float saturation = 1f;
float brightness;
ComplexNumber z = new ComplexNumber(x, y);
int i;
for (i = 0; i < Fractal.maxiter; i++) {
z.square();
z.add(Fractal.c);
if (z.mod() > Fractal.blowup) {
break;
}
}
brightness = (i < Fractal.maxiter) ? 1f : 0;
hue = (i%Fractal.maxiter)/(float)Fractal.maxiter;
int rgb = Color.getHSBColor(hue*5,saturation,brightness).getRGB();
return rgb;
}
private double map(double x, double in_min, double in_max, double out_min, double out_max) {
return (x-in_min)*(out_max-out_min)/(in_max-in_min) + out_min;
}
}
I tested with a range of values varying the maxiter, blowup and threshold.
I made the threshold such that the number of threads are around the same as the number of cores that I have (4).
I measured the runtimes in both cases and expected some optimization in parallelized code. However the code ran in the same time if not slower sometimes. This has me baffled. Is this happening because the problem size isn't big enough? I also tested with varying image sizes ranging from 640*400 to 1020*720.
Why is this happening? How can I run the code parallely so that it runs faster as it should?
Edit If you want to checkout the code in its entirety head over to my Github The master branch has the single threaded code. The branch with the name Multicore has the Parallelized code.
