| 1 |
/* |
| 2 |
* Written by Doug Lea with assistance from members of JCP JSR-166 |
| 3 |
* Expert Group and released to the public domain, as explained at |
| 4 |
* http://creativecommons.org/publicdomain/zero/1.0/ |
| 5 |
*/ |
| 6 |
|
| 7 |
import java.util.concurrent.*; |
| 8 |
|
| 9 |
/** Barrier version of Jacobi iteration */ |
| 10 |
public class FJPhaserJacobi { |
| 11 |
|
| 12 |
static int dimGran; |
| 13 |
|
| 14 |
static final double EPSILON = 0.0001; // convergence criterion |
| 15 |
|
| 16 |
public static void main(String[] args) { |
| 17 |
int n = 2048; |
| 18 |
int steps = 1000; |
| 19 |
try { |
| 20 |
if (args.length > 0) |
| 21 |
n = Integer.parseInt(args[0]); |
| 22 |
if (args.length > 1) |
| 23 |
steps = Integer.parseInt(args[1]); |
| 24 |
} |
| 25 |
|
| 26 |
catch (Exception e) { |
| 27 |
System.out.println("Usage: java ThreadPhaserJacobi <matrix size> <max steps>"); |
| 28 |
return; |
| 29 |
} |
| 30 |
|
| 31 |
ForkJoinPool fjp = new ForkJoinPool(); |
| 32 |
// int granularity = (n * n / fjp.getParallelism()) / 2; |
| 33 |
int granularity = n * n / fjp.getParallelism(); |
| 34 |
dimGran = (int)(Math.sqrt(granularity)); |
| 35 |
|
| 36 |
// allocate enough space for edges |
| 37 |
int dim = n+2; |
| 38 |
int ncells = dim * dim; |
| 39 |
double[][] a = new double[dim][dim]; |
| 40 |
double[][] b = new double[dim][dim]; |
| 41 |
// Initialize interiors to small value |
| 42 |
double smallVal = 1.0/dim; |
| 43 |
for (int i = 1; i < dim-1; ++i) { |
| 44 |
for (int j = 1; j < dim-1; ++j) |
| 45 |
a[i][j] = smallVal; |
| 46 |
} |
| 47 |
int nreps = 3; |
| 48 |
for (int rep = 0; rep < nreps; ++rep) { |
| 49 |
// Fill all edges with 1's. |
| 50 |
for (int k = 0; k < dim; ++k) { |
| 51 |
a[k][0] += 1.0; |
| 52 |
a[k][n+1] += 1.0; |
| 53 |
a[0][k] += 1.0; |
| 54 |
a[n+1][k] += 1.0; |
| 55 |
} |
| 56 |
Driver driver = new Driver(a, b, 1, n, 1, n, steps); |
| 57 |
long startTime = System.currentTimeMillis(); |
| 58 |
fjp.invoke(driver); |
| 59 |
|
| 60 |
long time = System.currentTimeMillis() - startTime; |
| 61 |
double secs = ((double)time) / 1000.0; |
| 62 |
|
| 63 |
System.out.println("Compute Time: " + secs); |
| 64 |
System.out.println(fjp); |
| 65 |
|
| 66 |
} |
| 67 |
|
| 68 |
} |
| 69 |
|
| 70 |
static class Segment extends CyclicAction { |
| 71 |
double[][] A; // matrix to get old values from |
| 72 |
double[][] B; // matrix to put new values into |
| 73 |
// indices of current submatrix |
| 74 |
final int loRow; |
| 75 |
final int hiRow; |
| 76 |
final int loCol; |
| 77 |
final int hiCol; |
| 78 |
volatile double maxDiff; // maximum difference between old and new values |
| 79 |
|
| 80 |
Segment(double[][] A, double[][] B, |
| 81 |
int loRow, int hiRow, |
| 82 |
int loCol, int hiCol, |
| 83 |
Phaser br) { |
| 84 |
super(br); |
| 85 |
this.A = A; this.B = B; |
| 86 |
this.loRow = loRow; this.hiRow = hiRow; |
| 87 |
this.loCol = loCol; this.hiCol = hiCol; |
| 88 |
} |
| 89 |
|
| 90 |
public void step() { |
| 91 |
maxDiff = update(A, B); |
| 92 |
double[][] tmp = A; A = B; B = tmp; |
| 93 |
} |
| 94 |
|
| 95 |
double update(double[][] a, double[][] b) { |
| 96 |
double md = 0.0; // local for computing max diff |
| 97 |
|
| 98 |
for (int i = loRow; i <= hiRow; ++i) { |
| 99 |
for (int j = loCol; j <= hiCol; ++j) { |
| 100 |
double v = 0.25 * (a[i-1][j] + a[i][j-1] + |
| 101 |
a[i+1][j] + a[i][j+1]); |
| 102 |
b[i][j] = v; |
| 103 |
|
| 104 |
double diff = v - a[i][j]; |
| 105 |
if (diff < 0) diff = -diff; |
| 106 |
if (diff > md) md = diff; |
| 107 |
} |
| 108 |
} |
| 109 |
|
| 110 |
return md; |
| 111 |
} |
| 112 |
|
| 113 |
} |
| 114 |
|
| 115 |
static class MyPhaser extends Phaser { |
| 116 |
final int max; |
| 117 |
MyPhaser(int steps) { this.max = steps - 1; } |
| 118 |
public boolean onAdvance(int phase, int registeredParties) { |
| 119 |
return phase >= max || registeredParties <= 0; |
| 120 |
} |
| 121 |
} |
| 122 |
|
| 123 |
static class Driver extends RecursiveAction { |
| 124 |
double[][] A; // matrix to get old values from |
| 125 |
double[][] B; // matrix to put new values into |
| 126 |
final int loRow; // indices of current submatrix |
| 127 |
final int hiRow; |
| 128 |
final int loCol; |
| 129 |
final int hiCol; |
| 130 |
final int steps; |
| 131 |
Driver(double[][] mat1, double[][] mat2, |
| 132 |
int firstRow, int lastRow, |
| 133 |
int firstCol, int lastCol, |
| 134 |
int steps) { |
| 135 |
|
| 136 |
this.A = mat1; this.B = mat2; |
| 137 |
this.loRow = firstRow; this.hiRow = lastRow; |
| 138 |
this.loCol = firstCol; this.hiCol = lastCol; |
| 139 |
this.steps = steps; |
| 140 |
} |
| 141 |
|
| 142 |
public void compute() { |
| 143 |
int rows = hiRow - loRow + 1; |
| 144 |
int cols = hiCol - loCol + 1; |
| 145 |
int rblocks = (int)(Math.round((float)rows / dimGran)); |
| 146 |
int cblocks = (int)(Math.round((float)cols / dimGran)); |
| 147 |
|
| 148 |
int n = rblocks * cblocks; |
| 149 |
|
| 150 |
System.out.println("Using " + n + " segments"); |
| 151 |
|
| 152 |
Segment[] segs = new Segment[n]; |
| 153 |
Phaser barrier = new MyPhaser(steps); |
| 154 |
int k = 0; |
| 155 |
for (int i = 0; i < rblocks; ++i) { |
| 156 |
int lr = loRow + i * dimGran; |
| 157 |
int hr = lr + dimGran; |
| 158 |
if (i == rblocks-1) hr = hiRow; |
| 159 |
|
| 160 |
for (int j = 0; j < cblocks; ++j) { |
| 161 |
int lc = loCol + j * dimGran; |
| 162 |
int hc = lc + dimGran; |
| 163 |
if (j == cblocks-1) hc = hiCol; |
| 164 |
|
| 165 |
segs[k] = new Segment(A, B, lr, hr, lc, hc, barrier); |
| 166 |
++k; |
| 167 |
} |
| 168 |
} |
| 169 |
invokeAll(segs); |
| 170 |
double maxd = 0; |
| 171 |
for (k = 0; k < n; ++k) { |
| 172 |
double md = segs[k].maxDiff; |
| 173 |
if (md > maxd) maxd = md; |
| 174 |
} |
| 175 |
System.out.println("Max diff after " + steps + " steps = " + maxd); |
| 176 |
} |
| 177 |
} |
| 178 |
} |
