1 |
// Barrier version of Jacobi iteration |
2 |
|
3 |
import java.util.*; |
4 |
import java.util.concurrent.*; |
5 |
//import jsr166y.*; |
6 |
|
7 |
public class ThreadPhaserJacobi { |
8 |
|
9 |
static final int nprocs = Runtime.getRuntime().availableProcessors(); |
10 |
|
11 |
/** |
12 |
* The maximum submatrix length (both row-wise and column-wise) |
13 |
* for any Segment |
14 |
**/ |
15 |
|
16 |
static final double EPSILON = 0.0001; // convergence criterion |
17 |
|
18 |
static int dimGran; |
19 |
|
20 |
public static void main(String[] args) throws Exception { |
21 |
int n = 2048; |
22 |
int steps = 1000; |
23 |
try { |
24 |
if (args.length > 0) |
25 |
n = Integer.parseInt(args[0]); |
26 |
if (args.length > 1) |
27 |
steps = Integer.parseInt(args[1]); |
28 |
} |
29 |
|
30 |
catch (Exception e) { |
31 |
System.out.println("Usage: java ThreadPhaserJacobi <matrix size> <max steps>"); |
32 |
return; |
33 |
} |
34 |
|
35 |
int granularity = n * n / nprocs; |
36 |
dimGran = (int)(Math.sqrt(granularity)); |
37 |
|
38 |
// allocate enough space for edges |
39 |
int dim = n+2; |
40 |
int ncells = dim * dim; |
41 |
double[][] a = new double[dim][dim]; |
42 |
double[][] b = new double[dim][dim]; |
43 |
// Initialize interiors to small value |
44 |
double smallVal = 1.0/dim; |
45 |
for (int i = 1; i < dim-1; ++i) { |
46 |
for (int j = 1; j < dim-1; ++j) |
47 |
a[i][j] = smallVal; |
48 |
} |
49 |
|
50 |
int nreps = 3; |
51 |
for (int rep = 0; rep < nreps; ++rep) { |
52 |
// Fill all edges with 1's. |
53 |
for (int k = 0; k < dim; ++k) { |
54 |
a[k][0] += 1.0; |
55 |
a[k][n+1] += 1.0; |
56 |
a[0][k] += 1.0; |
57 |
a[n+1][k] += 1.0; |
58 |
} |
59 |
Driver driver = new Driver(a, b, 1, n, 1, n, steps); |
60 |
long startTime = System.currentTimeMillis(); |
61 |
driver.compute(); |
62 |
|
63 |
long time = System.currentTimeMillis() - startTime; |
64 |
double secs = ((double)time) / 1000.0; |
65 |
|
66 |
System.out.println("Compute Time: " + secs); |
67 |
} |
68 |
} |
69 |
|
70 |
static class Segment implements Runnable { |
71 |
|
72 |
double[][] A; // matrix to get old values from |
73 |
double[][] B; // matrix to put new values into |
74 |
|
75 |
// indices of current submatrix |
76 |
final int loRow; |
77 |
final int hiRow; |
78 |
final int loCol; |
79 |
final int hiCol; |
80 |
final int steps; |
81 |
final Phaser barrier; |
82 |
double maxDiff; // maximum difference between old and new values |
83 |
|
84 |
Segment(double[][] A, double[][] B, |
85 |
int loRow, int hiRow, |
86 |
int loCol, int hiCol, |
87 |
int steps, |
88 |
Phaser barrier) { |
89 |
this.A = A; this.B = B; |
90 |
this.loRow = loRow; this.hiRow = hiRow; |
91 |
this.loCol = loCol; this.hiCol = hiCol; |
92 |
this.steps = steps; |
93 |
this.barrier = barrier; |
94 |
barrier.register(); |
95 |
} |
96 |
|
97 |
|
98 |
public void run() { |
99 |
try { |
100 |
double[][] a = A; |
101 |
double[][] b = B; |
102 |
|
103 |
for (int i = 0; i < steps; ++i) { |
104 |
maxDiff = update(a, b); |
105 |
if (barrier.awaitAdvance(barrier.arrive()) < 0) |
106 |
break; |
107 |
double[][] tmp = a; a = b; b = tmp; |
108 |
} |
109 |
} |
110 |
catch(Exception ex) { |
111 |
ex.printStackTrace(); |
112 |
return; |
113 |
} |
114 |
} |
115 |
|
116 |
double update(double[][] a, double[][] b) { |
117 |
double md = 0.0; // local for computing max diff |
118 |
|
119 |
for (int i = loRow; i <= hiRow; ++i) { |
120 |
for (int j = loCol; j <= hiCol; ++j) { |
121 |
double v = 0.25 * (a[i-1][j] + a[i][j-1] + |
122 |
a[i+1][j] + a[i][j+1]); |
123 |
b[i][j] = v; |
124 |
|
125 |
double diff = v - a[i][j]; |
126 |
if (diff < 0) diff = -diff; |
127 |
if (diff > md) md = diff; |
128 |
} |
129 |
} |
130 |
|
131 |
return md; |
132 |
} |
133 |
|
134 |
} |
135 |
|
136 |
|
137 |
static class Driver { |
138 |
double[][] A; // matrix to get old values from |
139 |
double[][] B; // matrix to put new values into |
140 |
|
141 |
final int loRow; // indices of current submatrix |
142 |
final int hiRow; |
143 |
final int loCol; |
144 |
final int hiCol; |
145 |
final int steps; |
146 |
Segment[] allSegments; |
147 |
|
148 |
Driver(double[][] mat1, double[][] mat2, |
149 |
int firstRow, int lastRow, |
150 |
int firstCol, int lastCol, |
151 |
int steps) { |
152 |
|
153 |
this.A = mat1; this.B = mat2; |
154 |
this.loRow = firstRow; this.hiRow = lastRow; |
155 |
this.loCol = firstCol; this.hiCol = lastCol; |
156 |
this.steps = steps; |
157 |
|
158 |
int rows = hiRow - loRow + 1; |
159 |
int cols = hiCol - loCol + 1; |
160 |
int rblocks = (int)(Math.round((float)rows / dimGran)); |
161 |
int cblocks = (int)(Math.round((float)cols / dimGran)); |
162 |
|
163 |
int n = rblocks * cblocks; |
164 |
|
165 |
Segment[] segs = new Segment[n]; |
166 |
Phaser barrier = new Phaser(); |
167 |
int k = 0; |
168 |
for (int i = 0; i < rblocks; ++i) { |
169 |
int lr = loRow + i * dimGran; |
170 |
int hr = lr + dimGran; |
171 |
if (i == rblocks-1) hr = hiRow; |
172 |
|
173 |
for (int j = 0; j < cblocks; ++j) { |
174 |
int lc = loCol + j * dimGran; |
175 |
int hc = lc + dimGran; |
176 |
if (j == cblocks-1) hc = hiCol; |
177 |
|
178 |
segs[k] = new Segment(A, B, lr, hr, lc, hc, steps, barrier); |
179 |
++k; |
180 |
} |
181 |
} |
182 |
System.out.println("Using " + n + " segments (threads)"); |
183 |
allSegments = segs; |
184 |
} |
185 |
|
186 |
public void compute() throws InterruptedException { |
187 |
Segment[] segs = allSegments; |
188 |
int n = segs.length; |
189 |
Thread[] threads = new Thread[n]; |
190 |
|
191 |
for (int k = 0; k < n; ++k) threads[k] = new Thread(segs[k]); |
192 |
for (int k = 0; k < n; ++k) threads[k].start(); |
193 |
for (int k = 0; k < n; ++k) threads[k].join(); |
194 |
|
195 |
double maxd = 0; |
196 |
for (int k = 0; k < n; ++k) { |
197 |
double md = segs[k].maxDiff; |
198 |
if (md > maxd) maxd = md; |
199 |
} |
200 |
|
201 |
System.out.println("Max diff after " + steps + " steps = " + maxd); |
202 |
} |
203 |
} |
204 |
|
205 |
|
206 |
} |