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 jsr166y.*; |
8 |
import java.util.concurrent.*; |
9 |
|
10 |
/** |
11 |
* LU matrix decomposition demo |
12 |
* Based on those in Cilk and Hood |
13 |
*/ |
14 |
public final class LU { |
15 |
|
16 |
/** for time conversion */ |
17 |
static final long NPS = (1000L * 1000 * 1000); |
18 |
|
19 |
// granularity is hard-wired as compile-time constant here |
20 |
static final int BLOCK_SIZE = 16; |
21 |
static final boolean CHECK = false; // set true to check answer |
22 |
|
23 |
public static void main(String[] args) throws Exception { |
24 |
|
25 |
final String usage = "Usage: java LU <threads> <matrix size (must be a power of two)> [runs] \n For example, try java LU 2 512"; |
26 |
|
27 |
int procs = 0; |
28 |
int n = 2048; |
29 |
int runs = 5; |
30 |
try { |
31 |
if (args.length > 0) |
32 |
procs = Integer.parseInt(args[0]); |
33 |
if (args.length > 1) |
34 |
n = Integer.parseInt(args[1]); |
35 |
if (args.length > 2) |
36 |
runs = Integer.parseInt(args[2]); |
37 |
} catch (Exception e) { |
38 |
System.out.println(usage); |
39 |
return; |
40 |
} |
41 |
|
42 |
if ( ((n & (n - 1)) != 0)) { |
43 |
System.out.println(usage); |
44 |
return; |
45 |
} |
46 |
ForkJoinPool pool = (procs == 0) ? new ForkJoinPool() : |
47 |
new ForkJoinPool(procs); |
48 |
System.out.println("procs: " + pool.getParallelism() + |
49 |
" n: " + n + " runs: " + runs); |
50 |
for (int run = 0; run < runs; ++run) { |
51 |
double[][] m = new double[n][n]; |
52 |
randomInit(m, n); |
53 |
double[][] copy = null; |
54 |
if (CHECK) { |
55 |
copy = new double[n][n]; |
56 |
for (int i = 0; i < n; ++i) { |
57 |
for (int j = 0; j < n; ++j) { |
58 |
copy[i][j] = m[i][j]; |
59 |
} |
60 |
} |
61 |
} |
62 |
Block M = new Block(m, 0, 0); |
63 |
long start = System.nanoTime(); |
64 |
pool.invoke(new LowerUpper(n, M)); |
65 |
long time = System.nanoTime() - start; |
66 |
double secs = ((double)time) / NPS; |
67 |
System.out.printf("\tTime: %7.3f\n", secs); |
68 |
|
69 |
if (CHECK) check(m, copy, n); |
70 |
} |
71 |
System.out.println(pool.toString()); |
72 |
pool.shutdown(); |
73 |
} |
74 |
|
75 |
static void randomInit(double[][] M, int n) { |
76 |
java.util.Random rng = new java.util.Random(); |
77 |
for (int i = 0; i < n; ++i) |
78 |
for (int j = 0; j < n; ++j) |
79 |
M[i][j] = rng.nextDouble(); |
80 |
// for compatibility with hood demo, force larger diagonals |
81 |
for (int k = 0; k < n; ++k) |
82 |
M[k][k] *= 10.0; |
83 |
} |
84 |
|
85 |
static void check(double[][] LU, double[][] M, int n) { |
86 |
double maxDiff = 0.0; // track max difference |
87 |
for (int i = 0; i < n; ++i) { |
88 |
for (int j = 0; j < n; ++j) { |
89 |
double v = 0.0; |
90 |
int k; |
91 |
for (k = 0; k < i && k <= j; k++ ) v += LU[i][k] * LU[k][j]; |
92 |
if (k == i && k <= j ) v += LU[k][j]; |
93 |
double diff = M[i][j] - v; |
94 |
if (diff < 0) diff = -diff; |
95 |
if (diff > 0.001) { |
96 |
System.out.println("large diff at[" + i + "," + j + "]: " + M[i][j] + " vs " + v); |
97 |
} |
98 |
if (diff > maxDiff) maxDiff = diff; |
99 |
} |
100 |
} |
101 |
|
102 |
System.out.println("Max difference = " + maxDiff); |
103 |
} |
104 |
|
105 |
// Blocks record underlying matrix, and offsets into current block |
106 |
static final class Block { |
107 |
final double[][] m; |
108 |
final int loRow; |
109 |
final int loCol; |
110 |
|
111 |
Block(double[][] mat, int lr, int lc) { |
112 |
m = mat; loRow = lr; loCol = lc; |
113 |
} |
114 |
} |
115 |
|
116 |
static final class Schur extends RecursiveAction { |
117 |
final int size; |
118 |
final Block V; |
119 |
final Block W; |
120 |
final Block M; |
121 |
|
122 |
Schur(int size, Block V, Block W, Block M) { |
123 |
this.size = size; this.V = V; this.W = W; this.M = M; |
124 |
} |
125 |
|
126 |
void schur() { // base case |
127 |
for (int j = 0; j < BLOCK_SIZE; ++j) { |
128 |
for (int i = 0; i < BLOCK_SIZE; ++i) { |
129 |
double s = M.m[i+M.loRow][j+M.loCol]; |
130 |
for (int k = 0; k < BLOCK_SIZE; ++k) { |
131 |
s -= V.m[i+V.loRow][k+V.loCol] * W.m[k+W.loRow][j+W.loCol]; |
132 |
} |
133 |
M.m[i+M.loRow][j+M.loCol] = s; |
134 |
} |
135 |
} |
136 |
} |
137 |
|
138 |
public void compute() { |
139 |
if (size == BLOCK_SIZE) { |
140 |
schur(); |
141 |
} |
142 |
else { |
143 |
int h = size / 2; |
144 |
|
145 |
Block M00 = new Block(M.m, M.loRow, M.loCol); |
146 |
Block M01 = new Block(M.m, M.loRow, M.loCol+h); |
147 |
Block M10 = new Block(M.m, M.loRow+h, M.loCol); |
148 |
Block M11 = new Block(M.m, M.loRow+h, M.loCol+h); |
149 |
|
150 |
Block V00 = new Block(V.m, V.loRow, V.loCol); |
151 |
Block V01 = new Block(V.m, V.loRow, V.loCol+h); |
152 |
Block V10 = new Block(V.m, V.loRow+h, V.loCol); |
153 |
Block V11 = new Block(V.m, V.loRow+h, V.loCol+h); |
154 |
|
155 |
Block W00 = new Block(W.m, W.loRow, W.loCol); |
156 |
Block W01 = new Block(W.m, W.loRow, W.loCol+h); |
157 |
Block W10 = new Block(W.m, W.loRow+h, W.loCol); |
158 |
Block W11 = new Block(W.m, W.loRow+h, W.loCol+h); |
159 |
|
160 |
Seq2 s3 = seq(new Schur(h, V10, W01, M11), |
161 |
new Schur(h, V11, W11, M11)); |
162 |
s3.fork(); |
163 |
Seq2 s2 = seq(new Schur(h, V10, W00, M10), |
164 |
new Schur(h, V11, W10, M10)); |
165 |
s2.fork(); |
166 |
Seq2 s1 = seq(new Schur(h, V00, W01, M01), |
167 |
new Schur(h, V01, W11, M01)); |
168 |
s1.fork(); |
169 |
new Schur(h, V00, W00, M00).compute(); |
170 |
new Schur(h, V01, W10, M00).compute(); |
171 |
if (s1.tryUnfork()) s1.compute(); else s1.join(); |
172 |
if (s2.tryUnfork()) s2.compute(); else s2.join(); |
173 |
if (s3.tryUnfork()) s3.compute(); else s3.join(); |
174 |
} |
175 |
} |
176 |
} |
177 |
|
178 |
static final class Lower extends RecursiveAction { |
179 |
final int size; |
180 |
final Block L; |
181 |
final Block M; |
182 |
Lower(int size, Block L, Block M) { |
183 |
this.size = size; this.L = L; this.M = M; |
184 |
} |
185 |
|
186 |
void lower() { // base case |
187 |
for (int i = 1; i < BLOCK_SIZE; ++i) { |
188 |
for (int k = 0; k < i; ++k) { |
189 |
double a = L.m[i+L.loRow][k+L.loCol]; |
190 |
double[] x = M.m[k+M.loRow]; |
191 |
double[] y = M.m[i+M.loRow]; |
192 |
int n = BLOCK_SIZE; |
193 |
for (int p = n-1; p >= 0; --p) { |
194 |
y[p+M.loCol] -= a * x[p+M.loCol]; |
195 |
} |
196 |
} |
197 |
} |
198 |
} |
199 |
|
200 |
public void compute() { |
201 |
if (size == BLOCK_SIZE) { |
202 |
lower(); |
203 |
} |
204 |
else { |
205 |
int h = size / 2; |
206 |
|
207 |
Block M00 = new Block(M.m, M.loRow, M.loCol); |
208 |
Block M01 = new Block(M.m, M.loRow, M.loCol+h); |
209 |
Block M10 = new Block(M.m, M.loRow+h, M.loCol); |
210 |
Block M11 = new Block(M.m, M.loRow+h, M.loCol+h); |
211 |
|
212 |
Block L00 = new Block(L.m, L.loRow, L.loCol); |
213 |
Block L01 = new Block(L.m, L.loRow, L.loCol+h); |
214 |
Block L10 = new Block(L.m, L.loRow+h, L.loCol); |
215 |
Block L11 = new Block(L.m, L.loRow+h, L.loCol+h); |
216 |
|
217 |
Seq3 s1 = |
218 |
seq(new Lower(h, L00, M00), |
219 |
new Schur(h, L10, M00, M10), |
220 |
new Lower(h, L11, M10)); |
221 |
Seq3 s2 = |
222 |
seq(new Lower(h, L00, M01), |
223 |
new Schur(h, L10, M01, M11), |
224 |
new Lower(h, L11, M11)); |
225 |
s2.fork(); |
226 |
s1.compute(); |
227 |
if (s2.tryUnfork()) s2.compute(); else s2.join(); |
228 |
} |
229 |
} |
230 |
} |
231 |
|
232 |
static final class Upper extends RecursiveAction { |
233 |
final int size; |
234 |
final Block U; |
235 |
final Block M; |
236 |
Upper(int size, Block U, Block M) { |
237 |
this.size = size; this.U = U; this.M = M; |
238 |
} |
239 |
|
240 |
void upper() { // base case |
241 |
for (int i = 0; i < BLOCK_SIZE; ++i) { |
242 |
for (int k = 0; k < BLOCK_SIZE; ++k) { |
243 |
double a = M.m[i+M.loRow][k+M.loCol] / U.m[k+U.loRow][k+U.loCol]; |
244 |
M.m[i+M.loRow][k+M.loCol] = a; |
245 |
double[] x = U.m[k+U.loRow]; |
246 |
double[] y = M.m[i+M.loRow]; |
247 |
int n = BLOCK_SIZE - k - 1; |
248 |
for (int p = n - 1; p >= 0; --p) { |
249 |
y[p+k+1+M.loCol] -= a * x[p+k+1+U.loCol]; |
250 |
} |
251 |
} |
252 |
} |
253 |
} |
254 |
|
255 |
public void compute() { |
256 |
if (size == BLOCK_SIZE) { |
257 |
upper(); |
258 |
} |
259 |
else { |
260 |
int h = size / 2; |
261 |
|
262 |
Block M00 = new Block(M.m, M.loRow, M.loCol); |
263 |
Block M01 = new Block(M.m, M.loRow, M.loCol+h); |
264 |
Block M10 = new Block(M.m, M.loRow+h, M.loCol); |
265 |
Block M11 = new Block(M.m, M.loRow+h, M.loCol+h); |
266 |
|
267 |
Block U00 = new Block(U.m, U.loRow, U.loCol); |
268 |
Block U01 = new Block(U.m, U.loRow, U.loCol+h); |
269 |
Block U10 = new Block(U.m, U.loRow+h, U.loCol); |
270 |
Block U11 = new Block(U.m, U.loRow+h, U.loCol+h); |
271 |
|
272 |
Seq3 s1 = |
273 |
seq(new Upper(h, U00, M00), |
274 |
new Schur(h, M00, U01, M01), |
275 |
new Upper(h, U11, M01)); |
276 |
Seq3 s2 = |
277 |
seq(new Upper(h, U00, M10), |
278 |
new Schur(h, M10, U01, M11), |
279 |
new Upper(h, U11, M11)); |
280 |
s2.fork(); |
281 |
s1.compute(); |
282 |
if (s2.tryUnfork()) s2.compute(); else s2.join(); |
283 |
} |
284 |
} |
285 |
} |
286 |
|
287 |
static final class LowerUpper extends RecursiveAction { |
288 |
final int size; |
289 |
final Block M; |
290 |
LowerUpper(int size, Block M) { |
291 |
this.size = size; this.M = M; |
292 |
} |
293 |
|
294 |
void lu() { // base case |
295 |
for (int k = 0; k < BLOCK_SIZE; ++k) { |
296 |
for (int i = k+1; i < BLOCK_SIZE; ++i) { |
297 |
double b = M.m[k+M.loRow][k+M.loCol]; |
298 |
double a = M.m[i+M.loRow][k+M.loCol] / b; |
299 |
M.m[i+M.loRow][k+M.loCol] = a; |
300 |
double[] x = M.m[k+M.loRow]; |
301 |
double[] y = M.m[i+M.loRow]; |
302 |
int n = BLOCK_SIZE-k-1; |
303 |
for (int p = n-1; p >= 0; --p) { |
304 |
y[k+1+p+M.loCol] -= a * x[k+1+p+M.loCol]; |
305 |
} |
306 |
} |
307 |
} |
308 |
} |
309 |
|
310 |
public void compute() { |
311 |
if (size == BLOCK_SIZE) { |
312 |
lu(); |
313 |
} |
314 |
else { |
315 |
int h = size / 2; |
316 |
Block M00 = new Block(M.m, M.loRow, M.loCol); |
317 |
Block M01 = new Block(M.m, M.loRow, M.loCol+h); |
318 |
Block M10 = new Block(M.m, M.loRow+h, M.loCol); |
319 |
Block M11 = new Block(M.m, M.loRow+h, M.loCol+h); |
320 |
|
321 |
new LowerUpper(h, M00).compute(); |
322 |
Lower sl = new Lower(h, M00, M01); |
323 |
Upper su = new Upper(h, M00, M10); |
324 |
su.fork(); |
325 |
sl.compute(); |
326 |
if (su.tryUnfork()) su.compute(); else su.join(); |
327 |
new Schur(h, M10, M01, M11).compute(); |
328 |
new LowerUpper(h, M11).compute(); |
329 |
} |
330 |
} |
331 |
} |
332 |
|
333 |
static Seq2 seq(RecursiveAction task1, |
334 |
RecursiveAction task2) { |
335 |
return new Seq2(task1, task2); |
336 |
} |
337 |
|
338 |
static final class Seq2 extends RecursiveAction { |
339 |
final RecursiveAction fst; |
340 |
final RecursiveAction snd; |
341 |
public Seq2(RecursiveAction task1, RecursiveAction task2) { |
342 |
fst = task1; |
343 |
snd = task2; |
344 |
} |
345 |
public void compute() { |
346 |
fst.invoke(); |
347 |
snd.invoke(); |
348 |
} |
349 |
} |
350 |
|
351 |
static Seq3 seq(RecursiveAction task1, |
352 |
RecursiveAction task2, |
353 |
RecursiveAction task3) { |
354 |
return new Seq3(task1, task2, task3); |
355 |
} |
356 |
|
357 |
static final class Seq3 extends RecursiveAction { |
358 |
final RecursiveAction fst; |
359 |
final RecursiveAction snd; |
360 |
final RecursiveAction thr; |
361 |
public Seq3(RecursiveAction task1, |
362 |
RecursiveAction task2, |
363 |
RecursiveAction task3) { |
364 |
fst = task1; |
365 |
snd = task2; |
366 |
thr = task3; |
367 |
} |
368 |
public void compute() { |
369 |
fst.invoke(); |
370 |
snd.invoke(); |
371 |
thr.invoke(); |
372 |
} |
373 |
} |
374 |
} |