1 |
dl |
1.1 |
/* |
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/licenses/publicdomain |
5 |
|
|
*/ |
6 |
|
|
|
7 |
|
|
//import jsr166y.*; |
8 |
|
|
import java.util.concurrent.*; |
9 |
|
|
import java.util.concurrent.TimeUnit; |
10 |
|
|
|
11 |
|
|
|
12 |
|
|
/** |
13 |
|
|
* LU matrix decomposition demo |
14 |
|
|
* Based on those in Cilk and Hood |
15 |
jsr166 |
1.3 |
*/ |
16 |
dl |
1.1 |
public final class LU { |
17 |
|
|
|
18 |
|
|
/** for time conversion */ |
19 |
|
|
static final long NPS = (1000L * 1000 * 1000); |
20 |
|
|
|
21 |
|
|
// granularity is hard-wired as compile-time constant here |
22 |
jsr166 |
1.2 |
static final int BLOCK_SIZE = 16; |
23 |
dl |
1.1 |
static final boolean CHECK = false; // set true to check answer |
24 |
|
|
|
25 |
|
|
public static void main(String[] args) throws Exception { |
26 |
|
|
|
27 |
|
|
final String usage = "Usage: java LU <threads> <matrix size (must be a power of two)> [runs] \n For example, try java LU 2 512"; |
28 |
|
|
|
29 |
|
|
int procs = 0; |
30 |
|
|
int n = 2048; |
31 |
|
|
int runs = 5; |
32 |
|
|
try { |
33 |
|
|
if (args.length > 0) |
34 |
|
|
procs = Integer.parseInt(args[0]); |
35 |
|
|
if (args.length > 1) |
36 |
|
|
n = Integer.parseInt(args[1]); |
37 |
jsr166 |
1.2 |
if (args.length > 2) |
38 |
dl |
1.1 |
runs = Integer.parseInt(args[2]); |
39 |
|
|
} catch (Exception e) { |
40 |
|
|
System.out.println(usage); |
41 |
|
|
return; |
42 |
|
|
} |
43 |
|
|
|
44 |
|
|
if ( ((n & (n - 1)) != 0)) { |
45 |
|
|
System.out.println(usage); |
46 |
|
|
return; |
47 |
|
|
} |
48 |
jsr166 |
1.4 |
ForkJoinPool pool = (procs == 0) ? new ForkJoinPool() : |
49 |
dl |
1.1 |
new ForkJoinPool(procs); |
50 |
jsr166 |
1.2 |
System.out.println("procs: " + pool.getParallelism() + |
51 |
dl |
1.1 |
" n: " + n + " runs: " + runs); |
52 |
|
|
for (int run = 0; run < runs; ++run) { |
53 |
|
|
double[][] m = new double[n][n]; |
54 |
|
|
randomInit(m, n); |
55 |
|
|
double[][] copy = null; |
56 |
|
|
if (CHECK) { |
57 |
|
|
copy = new double[n][n]; |
58 |
|
|
for (int i = 0; i < n; ++i) { |
59 |
|
|
for (int j = 0; j < n; ++j) { |
60 |
|
|
copy[i][j] = m[i][j]; |
61 |
|
|
} |
62 |
|
|
} |
63 |
|
|
} |
64 |
|
|
Block M = new Block(m, 0, 0); |
65 |
|
|
long start = System.nanoTime(); |
66 |
|
|
pool.invoke(new LowerUpper(n, M)); |
67 |
|
|
long time = System.nanoTime() - start; |
68 |
|
|
double secs = ((double)time) / NPS; |
69 |
|
|
System.out.printf("\tTime: %7.3f\n", secs); |
70 |
|
|
|
71 |
|
|
if (CHECK) check(m, copy, n); |
72 |
|
|
} |
73 |
|
|
System.out.println(pool.toString()); |
74 |
|
|
pool.shutdown(); |
75 |
|
|
} |
76 |
|
|
|
77 |
|
|
|
78 |
|
|
static void randomInit(double[][] M, int n) { |
79 |
|
|
java.util.Random rng = new java.util.Random(); |
80 |
|
|
for (int i = 0; i < n; ++i) |
81 |
|
|
for (int j = 0; j < n; ++j) |
82 |
|
|
M[i][j] = rng.nextDouble(); |
83 |
|
|
// for compatibility with hood demo, force larger diagonals |
84 |
|
|
for (int k = 0; k < n; ++k) |
85 |
|
|
M[k][k] *= 10.0; |
86 |
|
|
} |
87 |
|
|
|
88 |
|
|
static void check(double[][] LU, double[][] M, int n) { |
89 |
|
|
double maxDiff = 0.0; // track max difference |
90 |
|
|
for (int i = 0; i < n; ++i) { |
91 |
|
|
for (int j = 0; j < n; ++j) { |
92 |
|
|
double v = 0.0; |
93 |
|
|
int k; |
94 |
|
|
for (k = 0; k < i && k <= j; k++ ) v += LU[i][k] * LU[k][j]; |
95 |
|
|
if (k == i && k <= j ) v += LU[k][j]; |
96 |
|
|
double diff = M[i][j] - v; |
97 |
|
|
if (diff < 0) diff = -diff; |
98 |
|
|
if (diff > 0.001) { |
99 |
|
|
System.out.println("large diff at[" + i + "," + j + "]: " + M[i][j] + " vs " + v); |
100 |
|
|
} |
101 |
|
|
if (diff > maxDiff) maxDiff = diff; |
102 |
|
|
} |
103 |
|
|
} |
104 |
|
|
|
105 |
|
|
System.out.println("Max difference = " + maxDiff); |
106 |
|
|
} |
107 |
|
|
|
108 |
|
|
|
109 |
|
|
// Blocks record underlying matrix, and offsets into current block |
110 |
|
|
static final class Block { |
111 |
|
|
final double[][] m; |
112 |
|
|
final int loRow; |
113 |
|
|
final int loCol; |
114 |
|
|
|
115 |
|
|
Block(double[][] mat, int lr, int lc) { |
116 |
|
|
m = mat; loRow = lr; loCol = lc; |
117 |
|
|
} |
118 |
|
|
} |
119 |
|
|
|
120 |
|
|
static final class Schur extends RecursiveAction { |
121 |
|
|
final int size; |
122 |
|
|
final Block V; |
123 |
|
|
final Block W; |
124 |
|
|
final Block M; |
125 |
|
|
|
126 |
|
|
Schur(int size, Block V, Block W, Block M) { |
127 |
|
|
this.size = size; this.V = V; this.W = W; this.M = M; |
128 |
|
|
} |
129 |
|
|
|
130 |
|
|
void schur() { // base case |
131 |
|
|
for (int j = 0; j < BLOCK_SIZE; ++j) { |
132 |
|
|
for (int i = 0; i < BLOCK_SIZE; ++i) { |
133 |
|
|
double s = M.m[i+M.loRow][j+M.loCol]; |
134 |
|
|
for (int k = 0; k < BLOCK_SIZE; ++k) { |
135 |
|
|
s -= V.m[i+V.loRow][k+V.loCol] * W.m[k+W.loRow][j+W.loCol]; |
136 |
|
|
} |
137 |
|
|
M.m[i+M.loRow][j+M.loCol] = s; |
138 |
|
|
} |
139 |
|
|
} |
140 |
|
|
} |
141 |
|
|
|
142 |
|
|
public void compute() { |
143 |
|
|
if (size == BLOCK_SIZE) { |
144 |
|
|
schur(); |
145 |
|
|
} |
146 |
|
|
else { |
147 |
|
|
int h = size / 2; |
148 |
|
|
|
149 |
|
|
Block M00 = new Block(M.m, M.loRow, M.loCol); |
150 |
|
|
Block M01 = new Block(M.m, M.loRow, M.loCol+h); |
151 |
|
|
Block M10 = new Block(M.m, M.loRow+h, M.loCol); |
152 |
|
|
Block M11 = new Block(M.m, M.loRow+h, M.loCol+h); |
153 |
|
|
|
154 |
|
|
Block V00 = new Block(V.m, V.loRow, V.loCol); |
155 |
|
|
Block V01 = new Block(V.m, V.loRow, V.loCol+h); |
156 |
|
|
Block V10 = new Block(V.m, V.loRow+h, V.loCol); |
157 |
|
|
Block V11 = new Block(V.m, V.loRow+h, V.loCol+h); |
158 |
|
|
|
159 |
|
|
Block W00 = new Block(W.m, W.loRow, W.loCol); |
160 |
|
|
Block W01 = new Block(W.m, W.loRow, W.loCol+h); |
161 |
|
|
Block W10 = new Block(W.m, W.loRow+h, W.loCol); |
162 |
|
|
Block W11 = new Block(W.m, W.loRow+h, W.loCol+h); |
163 |
jsr166 |
1.2 |
|
164 |
dl |
1.1 |
Seq2 s3 = seq(new Schur(h, V10, W01, M11), |
165 |
|
|
new Schur(h, V11, W11, M11)); |
166 |
|
|
s3.fork(); |
167 |
|
|
Seq2 s2 = seq(new Schur(h, V10, W00, M10), |
168 |
|
|
new Schur(h, V11, W10, M10)); |
169 |
|
|
s2.fork(); |
170 |
|
|
Seq2 s1 = seq(new Schur(h, V00, W01, M01), |
171 |
|
|
new Schur(h, V01, W11, M01)); |
172 |
|
|
s1.fork(); |
173 |
|
|
new Schur(h, V00, W00, M00).compute(); |
174 |
|
|
new Schur(h, V01, W10, M00).compute(); |
175 |
|
|
if (s1.tryUnfork()) s1.compute(); else s1.join(); |
176 |
|
|
if (s2.tryUnfork()) s2.compute(); else s2.join(); |
177 |
|
|
if (s3.tryUnfork()) s3.compute(); else s3.join(); |
178 |
|
|
} |
179 |
|
|
} |
180 |
|
|
} |
181 |
|
|
|
182 |
|
|
static final class Lower extends RecursiveAction { |
183 |
|
|
final int size; |
184 |
|
|
final Block L; |
185 |
|
|
final Block M; |
186 |
|
|
Lower(int size, Block L, Block M) { |
187 |
|
|
this.size = size; this.L = L; this.M = M; |
188 |
|
|
} |
189 |
|
|
|
190 |
|
|
void lower() { // base case |
191 |
|
|
for (int i = 1; i < BLOCK_SIZE; ++i) { |
192 |
|
|
for (int k = 0; k < i; ++k) { |
193 |
|
|
double a = L.m[i+L.loRow][k+L.loCol]; |
194 |
|
|
double[] x = M.m[k+M.loRow]; |
195 |
|
|
double[] y = M.m[i+M.loRow]; |
196 |
|
|
int n = BLOCK_SIZE; |
197 |
|
|
for (int p = n-1; p >= 0; --p) { |
198 |
|
|
y[p+M.loCol] -= a * x[p+M.loCol]; |
199 |
|
|
} |
200 |
|
|
} |
201 |
|
|
} |
202 |
|
|
} |
203 |
|
|
|
204 |
|
|
public void compute() { |
205 |
|
|
if (size == BLOCK_SIZE) { |
206 |
|
|
lower(); |
207 |
|
|
} |
208 |
|
|
else { |
209 |
|
|
int h = size / 2; |
210 |
|
|
|
211 |
|
|
Block M00 = new Block(M.m, M.loRow, M.loCol); |
212 |
|
|
Block M01 = new Block(M.m, M.loRow, M.loCol+h); |
213 |
|
|
Block M10 = new Block(M.m, M.loRow+h, M.loCol); |
214 |
|
|
Block M11 = new Block(M.m, M.loRow+h, M.loCol+h); |
215 |
|
|
|
216 |
|
|
Block L00 = new Block(L.m, L.loRow, L.loCol); |
217 |
|
|
Block L01 = new Block(L.m, L.loRow, L.loCol+h); |
218 |
|
|
Block L10 = new Block(L.m, L.loRow+h, L.loCol); |
219 |
|
|
Block L11 = new Block(L.m, L.loRow+h, L.loCol+h); |
220 |
|
|
|
221 |
jsr166 |
1.2 |
|
222 |
|
|
Seq3 s1 = |
223 |
dl |
1.1 |
seq(new Lower(h, L00, M00), |
224 |
|
|
new Schur(h, L10, M00, M10), |
225 |
|
|
new Lower(h, L11, M10)); |
226 |
jsr166 |
1.2 |
Seq3 s2 = |
227 |
dl |
1.1 |
seq(new Lower(h, L00, M01), |
228 |
|
|
new Schur(h, L10, M01, M11), |
229 |
|
|
new Lower(h, L11, M11)); |
230 |
|
|
s2.fork(); |
231 |
|
|
s1.compute(); |
232 |
|
|
if (s2.tryUnfork()) s2.compute(); else s2.join(); |
233 |
|
|
} |
234 |
|
|
} |
235 |
|
|
} |
236 |
|
|
|
237 |
|
|
|
238 |
|
|
static final class Upper extends RecursiveAction { |
239 |
|
|
final int size; |
240 |
|
|
final Block U; |
241 |
|
|
final Block M; |
242 |
|
|
Upper(int size, Block U, Block M) { |
243 |
|
|
this.size = size; this.U = U; this.M = M; |
244 |
|
|
} |
245 |
|
|
|
246 |
|
|
void upper() { // base case |
247 |
|
|
for (int i = 0; i < BLOCK_SIZE; ++i) { |
248 |
|
|
for (int k = 0; k < BLOCK_SIZE; ++k) { |
249 |
|
|
double a = M.m[i+M.loRow][k+M.loCol] / U.m[k+U.loRow][k+U.loCol]; |
250 |
|
|
M.m[i+M.loRow][k+M.loCol] = a; |
251 |
|
|
double[] x = U.m[k+U.loRow]; |
252 |
|
|
double[] y = M.m[i+M.loRow]; |
253 |
|
|
int n = BLOCK_SIZE - k - 1; |
254 |
|
|
for (int p = n - 1; p >= 0; --p) { |
255 |
|
|
y[p+k+1+M.loCol] -= a * x[p+k+1+U.loCol]; |
256 |
|
|
} |
257 |
|
|
} |
258 |
|
|
} |
259 |
|
|
} |
260 |
|
|
|
261 |
|
|
|
262 |
|
|
public void compute() { |
263 |
|
|
if (size == BLOCK_SIZE) { |
264 |
|
|
upper(); |
265 |
|
|
} |
266 |
|
|
else { |
267 |
|
|
int h = size / 2; |
268 |
|
|
|
269 |
|
|
Block M00 = new Block(M.m, M.loRow, M.loCol); |
270 |
|
|
Block M01 = new Block(M.m, M.loRow, M.loCol+h); |
271 |
|
|
Block M10 = new Block(M.m, M.loRow+h, M.loCol); |
272 |
|
|
Block M11 = new Block(M.m, M.loRow+h, M.loCol+h); |
273 |
|
|
|
274 |
|
|
Block U00 = new Block(U.m, U.loRow, U.loCol); |
275 |
|
|
Block U01 = new Block(U.m, U.loRow, U.loCol+h); |
276 |
|
|
Block U10 = new Block(U.m, U.loRow+h, U.loCol); |
277 |
|
|
Block U11 = new Block(U.m, U.loRow+h, U.loCol+h); |
278 |
|
|
|
279 |
jsr166 |
1.2 |
Seq3 s1 = |
280 |
dl |
1.1 |
seq(new Upper(h, U00, M00), |
281 |
|
|
new Schur(h, M00, U01, M01), |
282 |
|
|
new Upper(h, U11, M01)); |
283 |
jsr166 |
1.2 |
Seq3 s2 = |
284 |
dl |
1.1 |
seq(new Upper(h, U00, M10), |
285 |
|
|
new Schur(h, M10, U01, M11), |
286 |
|
|
new Upper(h, U11, M11)); |
287 |
|
|
s2.fork(); |
288 |
|
|
s1.compute(); |
289 |
|
|
if (s2.tryUnfork()) s2.compute(); else s2.join(); |
290 |
|
|
} |
291 |
|
|
} |
292 |
|
|
} |
293 |
jsr166 |
1.2 |
|
294 |
dl |
1.1 |
|
295 |
|
|
static final class LowerUpper extends RecursiveAction { |
296 |
|
|
final int size; |
297 |
|
|
final Block M; |
298 |
|
|
LowerUpper(int size, Block M) { |
299 |
|
|
this.size = size; this.M = M; |
300 |
|
|
} |
301 |
|
|
|
302 |
|
|
void lu() { // base case |
303 |
|
|
for (int k = 0; k < BLOCK_SIZE; ++k) { |
304 |
|
|
for (int i = k+1; i < BLOCK_SIZE; ++i) { |
305 |
|
|
double b = M.m[k+M.loRow][k+M.loCol]; |
306 |
|
|
double a = M.m[i+M.loRow][k+M.loCol] / b; |
307 |
|
|
M.m[i+M.loRow][k+M.loCol] = a; |
308 |
|
|
double[] x = M.m[k+M.loRow]; |
309 |
|
|
double[] y = M.m[i+M.loRow]; |
310 |
|
|
int n = BLOCK_SIZE-k-1; |
311 |
|
|
for (int p = n-1; p >= 0; --p) { |
312 |
|
|
y[k+1+p+M.loCol] -= a * x[k+1+p+M.loCol]; |
313 |
|
|
} |
314 |
|
|
} |
315 |
|
|
} |
316 |
|
|
} |
317 |
|
|
|
318 |
|
|
public void compute() { |
319 |
|
|
if (size == BLOCK_SIZE) { |
320 |
|
|
lu(); |
321 |
|
|
} |
322 |
|
|
else { |
323 |
|
|
int h = size / 2; |
324 |
|
|
Block M00 = new Block(M.m, M.loRow, M.loCol); |
325 |
|
|
Block M01 = new Block(M.m, M.loRow, M.loCol+h); |
326 |
|
|
Block M10 = new Block(M.m, M.loRow+h, M.loCol); |
327 |
|
|
Block M11 = new Block(M.m, M.loRow+h, M.loCol+h); |
328 |
|
|
|
329 |
|
|
new LowerUpper(h, M00).compute(); |
330 |
|
|
Lower sl = new Lower(h, M00, M01); |
331 |
|
|
Upper su = new Upper(h, M00, M10); |
332 |
|
|
su.fork(); |
333 |
|
|
sl.compute(); |
334 |
|
|
if (su.tryUnfork()) su.compute(); else su.join(); |
335 |
|
|
new Schur(h, M10, M01, M11).compute(); |
336 |
|
|
new LowerUpper(h, M11).compute(); |
337 |
|
|
} |
338 |
|
|
} |
339 |
|
|
} |
340 |
|
|
|
341 |
jsr166 |
1.2 |
static Seq2 seq(RecursiveAction task1, |
342 |
|
|
RecursiveAction task2) { |
343 |
|
|
return new Seq2(task1, task2); |
344 |
dl |
1.1 |
} |
345 |
|
|
|
346 |
|
|
static final class Seq2 extends RecursiveAction { |
347 |
|
|
final RecursiveAction fst; |
348 |
|
|
final RecursiveAction snd; |
349 |
|
|
public Seq2(RecursiveAction task1, RecursiveAction task2) { |
350 |
|
|
fst = task1; |
351 |
|
|
snd = task2; |
352 |
|
|
} |
353 |
|
|
public void compute() { |
354 |
|
|
fst.invoke(); |
355 |
|
|
snd.invoke(); |
356 |
|
|
} |
357 |
|
|
} |
358 |
|
|
|
359 |
jsr166 |
1.2 |
static Seq3 seq(RecursiveAction task1, |
360 |
dl |
1.1 |
RecursiveAction task2, |
361 |
jsr166 |
1.2 |
RecursiveAction task3) { |
362 |
|
|
return new Seq3(task1, task2, task3); |
363 |
dl |
1.1 |
} |
364 |
|
|
|
365 |
|
|
static final class Seq3 extends RecursiveAction { |
366 |
|
|
final RecursiveAction fst; |
367 |
|
|
final RecursiveAction snd; |
368 |
|
|
final RecursiveAction thr; |
369 |
jsr166 |
1.2 |
public Seq3(RecursiveAction task1, |
370 |
dl |
1.1 |
RecursiveAction task2, |
371 |
|
|
RecursiveAction task3) { |
372 |
|
|
fst = task1; |
373 |
|
|
snd = task2; |
374 |
|
|
thr = task3; |
375 |
|
|
} |
376 |
|
|
public void compute() { |
377 |
|
|
fst.invoke(); |
378 |
|
|
snd.invoke(); |
379 |
|
|
thr.invoke(); |
380 |
|
|
} |
381 |
|
|
} |
382 |
|
|
|
383 |
|
|
|
384 |
|
|
|
385 |
|
|
} |
386 |
jsr166 |
1.2 |
|