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 |
// Barrier version of Jacobi iteration |
8 |
|
9 |
import java.util.*; |
10 |
import java.util.concurrent.*; |
11 |
//import jsr166y.*; |
12 |
|
13 |
public class ThreadPhaserJacobi { |
14 |
|
15 |
static final int nprocs = Runtime.getRuntime().availableProcessors(); |
16 |
|
17 |
/** |
18 |
* The maximum submatrix length (both row-wise and column-wise) |
19 |
* for any Segment |
20 |
*/ |
21 |
static final double EPSILON = 0.0001; // convergence criterion |
22 |
|
23 |
static int dimGran; |
24 |
|
25 |
public static void main(String[] args) throws Exception { |
26 |
int n = 2048; |
27 |
int steps = 1000; |
28 |
try { |
29 |
if (args.length > 0) |
30 |
n = Integer.parseInt(args[0]); |
31 |
if (args.length > 1) |
32 |
steps = Integer.parseInt(args[1]); |
33 |
} |
34 |
|
35 |
catch (Exception e) { |
36 |
System.out.println("Usage: java ThreadPhaserJacobi <matrix size> <max steps>"); |
37 |
return; |
38 |
} |
39 |
|
40 |
int granularity = n * n / nprocs; |
41 |
dimGran = (int) Math.sqrt(granularity); |
42 |
|
43 |
// allocate enough space for edges |
44 |
int dim = n+2; |
45 |
int ncells = dim * dim; |
46 |
double[][] a = new double[dim][dim]; |
47 |
double[][] b = new double[dim][dim]; |
48 |
// Initialize interiors to small value |
49 |
double smallVal = 1.0/dim; |
50 |
for (int i = 1; i < dim-1; ++i) { |
51 |
for (int j = 1; j < dim-1; ++j) |
52 |
a[i][j] = smallVal; |
53 |
} |
54 |
|
55 |
int nreps = 3; |
56 |
for (int rep = 0; rep < nreps; ++rep) { |
57 |
// Fill all edges with 1's. |
58 |
for (int k = 0; k < dim; ++k) { |
59 |
a[k][0] += 1.0; |
60 |
a[k][n+1] += 1.0; |
61 |
a[0][k] += 1.0; |
62 |
a[n+1][k] += 1.0; |
63 |
} |
64 |
Driver driver = new Driver(a, b, 1, n, 1, n, steps); |
65 |
long startTime = System.currentTimeMillis(); |
66 |
driver.compute(); |
67 |
|
68 |
long time = System.currentTimeMillis() - startTime; |
69 |
double secs = (double) time / 1000.0; |
70 |
|
71 |
System.out.println("Compute Time: " + secs); |
72 |
} |
73 |
} |
74 |
|
75 |
static class Segment implements Runnable { |
76 |
|
77 |
double[][] A; // matrix to get old values from |
78 |
double[][] B; // matrix to put new values into |
79 |
|
80 |
// indices of current submatrix |
81 |
final int loRow; |
82 |
final int hiRow; |
83 |
final int loCol; |
84 |
final int hiCol; |
85 |
final int steps; |
86 |
final Phaser barrier; |
87 |
double maxDiff; // maximum difference between old and new values |
88 |
|
89 |
Segment(double[][] A, double[][] B, |
90 |
int loRow, int hiRow, |
91 |
int loCol, int hiCol, |
92 |
int steps, |
93 |
Phaser barrier) { |
94 |
this.A = A; this.B = B; |
95 |
this.loRow = loRow; this.hiRow = hiRow; |
96 |
this.loCol = loCol; this.hiCol = hiCol; |
97 |
this.steps = steps; |
98 |
this.barrier = barrier; |
99 |
barrier.register(); |
100 |
} |
101 |
|
102 |
public void run() { |
103 |
try { |
104 |
double[][] a = A; |
105 |
double[][] b = B; |
106 |
|
107 |
for (int i = 0; i < steps; ++i) { |
108 |
maxDiff = update(a, b); |
109 |
if (barrier.awaitAdvance(barrier.arrive()) < 0) |
110 |
break; |
111 |
double[][] tmp = a; a = b; b = tmp; |
112 |
} |
113 |
} |
114 |
catch (Exception ex) { |
115 |
ex.printStackTrace(); |
116 |
return; |
117 |
} |
118 |
} |
119 |
|
120 |
double update(double[][] a, double[][] b) { |
121 |
double md = 0.0; // local for computing max diff |
122 |
|
123 |
for (int i = loRow; i <= hiRow; ++i) { |
124 |
for (int j = loCol; j <= hiCol; ++j) { |
125 |
double v = 0.25 * (a[i-1][j] + a[i][j-1] + |
126 |
a[i+1][j] + a[i][j+1]); |
127 |
b[i][j] = v; |
128 |
|
129 |
double diff = v - a[i][j]; |
130 |
if (diff < 0) diff = -diff; |
131 |
if (diff > md) md = diff; |
132 |
} |
133 |
} |
134 |
|
135 |
return md; |
136 |
} |
137 |
|
138 |
} |
139 |
|
140 |
static class Driver { |
141 |
double[][] A; // matrix to get old values from |
142 |
double[][] B; // matrix to put new values into |
143 |
|
144 |
final int loRow; // indices of current submatrix |
145 |
final int hiRow; |
146 |
final int loCol; |
147 |
final int hiCol; |
148 |
final int steps; |
149 |
Segment[] allSegments; |
150 |
|
151 |
Driver(double[][] mat1, double[][] mat2, |
152 |
int firstRow, int lastRow, |
153 |
int firstCol, int lastCol, |
154 |
int steps) { |
155 |
|
156 |
this.A = mat1; this.B = mat2; |
157 |
this.loRow = firstRow; this.hiRow = lastRow; |
158 |
this.loCol = firstCol; this.hiCol = lastCol; |
159 |
this.steps = steps; |
160 |
|
161 |
int rows = hiRow - loRow + 1; |
162 |
int cols = hiCol - loCol + 1; |
163 |
int rblocks = Math.round((float) rows / dimGran); |
164 |
int cblocks = Math.round((float) cols / dimGran); |
165 |
|
166 |
int n = rblocks * cblocks; |
167 |
|
168 |
Segment[] segs = new Segment[n]; |
169 |
Phaser barrier = new Phaser(); |
170 |
int k = 0; |
171 |
for (int i = 0; i < rblocks; ++i) { |
172 |
int lr = loRow + i * dimGran; |
173 |
int hr = lr + dimGran; |
174 |
if (i == rblocks-1) hr = hiRow; |
175 |
|
176 |
for (int j = 0; j < cblocks; ++j) { |
177 |
int lc = loCol + j * dimGran; |
178 |
int hc = lc + dimGran; |
179 |
if (j == cblocks-1) hc = hiCol; |
180 |
|
181 |
segs[k] = new Segment(A, B, lr, hr, lc, hc, steps, barrier); |
182 |
++k; |
183 |
} |
184 |
} |
185 |
System.out.println("Using " + n + " segments (threads)"); |
186 |
allSegments = segs; |
187 |
} |
188 |
|
189 |
public void compute() throws InterruptedException { |
190 |
Segment[] segs = allSegments; |
191 |
int n = segs.length; |
192 |
Thread[] threads = new Thread[n]; |
193 |
|
194 |
for (int k = 0; k < n; ++k) threads[k] = new Thread(segs[k]); |
195 |
for (int k = 0; k < n; ++k) threads[k].start(); |
196 |
for (int k = 0; k < n; ++k) threads[k].join(); |
197 |
|
198 |
double maxd = 0; |
199 |
for (int k = 0; k < n; ++k) { |
200 |
double md = segs[k].maxDiff; |
201 |
if (md > maxd) maxd = md; |
202 |
} |
203 |
|
204 |
System.out.println("Max diff after " + steps + " steps = " + maxd); |
205 |
} |
206 |
} |
207 |
} |