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 java.util.*; |
8 |
|
|
import java.util.concurrent.*; |
9 |
|
|
import java.util.concurrent.atomic.*; |
10 |
|
|
|
11 |
|
|
// parallel sums and cumulations |
12 |
|
|
|
13 |
|
|
public class FJSums { |
14 |
|
|
static final long NPS = (1000L * 1000 * 1000); |
15 |
|
|
static int THRESHOLD; |
16 |
|
|
|
17 |
|
|
public static void main (String[] args) throws Exception { |
18 |
|
|
int procs = 0; |
19 |
|
|
int n = 1 << 25; |
20 |
|
|
int reps = 10; |
21 |
|
|
try { |
22 |
|
|
if (args.length > 0) |
23 |
|
|
procs = Integer.parseInt(args[0]); |
24 |
|
|
if (args.length > 1) |
25 |
|
|
n = Integer.parseInt(args[1]); |
26 |
|
|
if (args.length > 2) |
27 |
|
|
reps = Integer.parseInt(args[1]); |
28 |
|
|
} |
29 |
|
|
catch (Exception e) { |
30 |
|
|
System.out.println("Usage: java FJSums threads n reps"); |
31 |
|
|
return; |
32 |
|
|
} |
33 |
jsr166 |
1.2 |
ForkJoinPool g = procs == 0? new ForkJoinPool() : |
34 |
dl |
1.1 |
new ForkJoinPool(procs); |
35 |
|
|
System.out.println("Number of procs=" + g.getParallelism()); |
36 |
|
|
// for now hardwire Cumulate threshold to 8 * #CPUs leaf tasks |
37 |
|
|
THRESHOLD = 1 + ((n + 7) >>> 3) / g.getParallelism(); |
38 |
jsr166 |
1.2 |
|
39 |
dl |
1.1 |
long[] a = new long[n]; |
40 |
|
|
for (int i = 0; i < n; ++i) |
41 |
|
|
a[i] = i; |
42 |
|
|
long expected = ((long)n * (long)(n - 1)) / 2; |
43 |
|
|
for (int i = 0; i < 2; ++i) { |
44 |
|
|
System.out.print("Seq: "); |
45 |
|
|
long last = System.nanoTime(); |
46 |
|
|
long ss = seqSum(a, 0, n); |
47 |
|
|
double elapsed = elapsedTime(last); |
48 |
|
|
System.out.printf("sum = %24d time: %7.3f\n", ss, elapsed); |
49 |
jsr166 |
1.2 |
if (ss != expected) |
50 |
dl |
1.1 |
throw new Error("expected " + expected + " != " + ss); |
51 |
|
|
} |
52 |
|
|
for (int i = 0; i < reps; ++i) { |
53 |
|
|
System.out.print("Par: "); |
54 |
|
|
long last = System.nanoTime(); |
55 |
|
|
Summer s = new Summer(a, 0, a.length, null); |
56 |
|
|
g.invoke(s); |
57 |
|
|
long ss = s.result; |
58 |
|
|
double elapsed = elapsedTime(last); |
59 |
|
|
System.out.printf("sum = %24d time: %7.3f\n", ss, elapsed); |
60 |
jsr166 |
1.2 |
if (i == 0 && ss != expected) |
61 |
dl |
1.1 |
throw new Error("expected " + expected + " != " + ss); |
62 |
|
|
System.out.print("Cum: "); |
63 |
|
|
last = System.nanoTime(); |
64 |
|
|
g.invoke(new Cumulater(null, a, 0, n)); |
65 |
|
|
long sc = a[n - 1]; |
66 |
|
|
elapsed = elapsedTime(last); |
67 |
|
|
System.out.printf("sum = %24d time: %7.3f\n", ss, elapsed); |
68 |
|
|
if (sc != ss) |
69 |
|
|
throw new Error("expected " + ss + " != " + sc); |
70 |
|
|
} |
71 |
|
|
System.out.println(g); |
72 |
|
|
g.shutdown(); |
73 |
|
|
} |
74 |
|
|
|
75 |
|
|
static double elapsedTime(long startTime) { |
76 |
|
|
return (double)(System.nanoTime() - startTime) / NPS; |
77 |
|
|
} |
78 |
|
|
|
79 |
|
|
static long seqSum(long[] array, int l, int h) { |
80 |
|
|
long sum = 0; |
81 |
jsr166 |
1.2 |
for (int i = l; i < h; ++i) |
82 |
dl |
1.1 |
sum += array[i]; |
83 |
|
|
return sum; |
84 |
|
|
} |
85 |
|
|
|
86 |
|
|
static long seqCumulate(long[] array, int lo, int hi, long base) { |
87 |
|
|
long sum = base; |
88 |
|
|
for (int i = lo; i < hi; ++i) |
89 |
|
|
array[i] = sum += array[i]; |
90 |
|
|
return sum; |
91 |
|
|
} |
92 |
jsr166 |
1.2 |
|
93 |
dl |
1.1 |
/** |
94 |
|
|
* Adapted from Applyer demo in RecursiveAction docs |
95 |
|
|
*/ |
96 |
|
|
static final class Summer extends RecursiveAction { |
97 |
|
|
final long[] array; |
98 |
|
|
final int lo, hi; |
99 |
|
|
long result; |
100 |
|
|
Summer next; // keeps track of right-hand-side tasks |
101 |
|
|
Summer(long[] array, int lo, int hi, Summer next) { |
102 |
|
|
this.array = array; this.lo = lo; this.hi = hi; |
103 |
|
|
this.next = next; |
104 |
|
|
} |
105 |
jsr166 |
1.2 |
|
106 |
dl |
1.1 |
protected void compute() { |
107 |
|
|
int l = lo; |
108 |
|
|
int h = hi; |
109 |
|
|
Summer right = null; |
110 |
|
|
while (h - l > 1 && getSurplusQueuedTaskCount() <= 3) { |
111 |
|
|
int mid = (l + h) >>> 1; |
112 |
|
|
right = new Summer(array, mid, h, right); |
113 |
|
|
right.fork(); |
114 |
|
|
h = mid; |
115 |
|
|
} |
116 |
|
|
long sum = seqSum(array, l, h); |
117 |
|
|
while (right != null) { |
118 |
|
|
if (right.tryUnfork()) // directly calculate if not stolen |
119 |
|
|
sum += seqSum(array, right.lo, right.hi); |
120 |
|
|
else { |
121 |
|
|
right.join(); |
122 |
|
|
sum += right.result; |
123 |
|
|
} |
124 |
|
|
right = right.next; |
125 |
|
|
} |
126 |
|
|
result = sum; |
127 |
|
|
} |
128 |
|
|
} |
129 |
|
|
|
130 |
|
|
/** |
131 |
|
|
* Cumulative scan, adapted from ParallelArray code |
132 |
|
|
* |
133 |
|
|
* A basic version of scan is straightforward. |
134 |
|
|
* Keep dividing by two to threshold segment size, and then: |
135 |
|
|
* Pass 1: Create tree of partial sums for each segment |
136 |
|
|
* Pass 2: For each segment, cumulate with offset of left sibling |
137 |
|
|
* See G. Blelloch's http://www.cs.cmu.edu/~scandal/alg/scan.html |
138 |
|
|
* |
139 |
|
|
* This version improves performance within FJ framework mainly by |
140 |
|
|
* allowing second pass of ready left-hand sides to proceed even |
141 |
|
|
* if some right-hand side first passes are still executing. It |
142 |
|
|
* also combines first and second pass for leftmost segment, and |
143 |
|
|
* for cumulate (not precumulate) also skips first pass for |
144 |
|
|
* rightmost segment (whose result is not needed for second pass). |
145 |
|
|
* |
146 |
|
|
* To manage this, it relies on "phase" phase/state control field |
147 |
|
|
* maintaining bits CUMULATE, SUMMED, and FINISHED. CUMULATE is |
148 |
|
|
* main phase bit. When false, segments compute only their sum. |
149 |
|
|
* When true, they cumulate array elements. CUMULATE is set at |
150 |
|
|
* root at beginning of second pass and then propagated down. But |
151 |
|
|
* it may also be set earlier for subtrees with lo==0 (the |
152 |
|
|
* left spine of tree). SUMMED is a one bit join count. For leafs, |
153 |
|
|
* set when summed. For internal nodes, becomes true when one |
154 |
|
|
* child is summed. When second child finishes summing, it then |
155 |
|
|
* moves up tree to trigger cumulate phase. FINISHED is also a one |
156 |
|
|
* bit join count. For leafs, it is set when cumulated. For |
157 |
|
|
* internal nodes, it becomes true when one child is cumulated. |
158 |
|
|
* When second child finishes cumulating, it then moves up tree, |
159 |
|
|
* executing complete() at the root. |
160 |
|
|
* |
161 |
|
|
*/ |
162 |
|
|
static final class Cumulater extends ForkJoinTask<Void> { |
163 |
|
|
static final short CUMULATE = (short)1; |
164 |
|
|
static final short SUMMED = (short)2; |
165 |
|
|
static final short FINISHED = (short)4; |
166 |
|
|
|
167 |
|
|
final Cumulater parent; |
168 |
|
|
final long[] array; |
169 |
|
|
Cumulater left, right; |
170 |
|
|
final int lo; |
171 |
|
|
final int hi; |
172 |
|
|
volatile int phase; // phase/state |
173 |
|
|
long in, out; // initially zero |
174 |
|
|
|
175 |
|
|
static final AtomicIntegerFieldUpdater<Cumulater> phaseUpdater = |
176 |
|
|
AtomicIntegerFieldUpdater.newUpdater(Cumulater.class, "phase"); |
177 |
|
|
|
178 |
|
|
Cumulater(Cumulater parent, long[] array, int lo, int hi) { |
179 |
|
|
this.parent = parent; |
180 |
|
|
this.array = array; |
181 |
|
|
this.lo = lo; |
182 |
|
|
this.hi = hi; |
183 |
|
|
} |
184 |
|
|
|
185 |
|
|
public final Void getRawResult() { return null; } |
186 |
|
|
protected final void setRawResult(Void mustBeNull) { } |
187 |
|
|
|
188 |
|
|
/** Returns true if can CAS CUMULATE bit true */ |
189 |
|
|
final boolean transitionToCumulate() { |
190 |
|
|
int c; |
191 |
|
|
while (((c = phase) & CUMULATE) == 0) |
192 |
|
|
if (phaseUpdater.compareAndSet(this, c, c | CUMULATE)) |
193 |
|
|
return true; |
194 |
|
|
return false; |
195 |
|
|
} |
196 |
|
|
|
197 |
|
|
public final boolean exec() { |
198 |
|
|
if (hi - lo > THRESHOLD) { |
199 |
|
|
if (left == null) { // first pass |
200 |
|
|
int mid = (lo + hi) >>> 1; |
201 |
|
|
left = new Cumulater(this, array, lo, mid); |
202 |
|
|
right = new Cumulater(this, array, mid, hi); |
203 |
|
|
} |
204 |
|
|
|
205 |
|
|
boolean cumulate = (phase & CUMULATE) != 0; |
206 |
|
|
if (cumulate) { |
207 |
|
|
long pin = in; |
208 |
|
|
left.in = pin; |
209 |
|
|
right.in = pin + left.out; |
210 |
|
|
} |
211 |
|
|
|
212 |
|
|
if (!cumulate || right.transitionToCumulate()) |
213 |
|
|
right.fork(); |
214 |
|
|
if (!cumulate || left.transitionToCumulate()) |
215 |
|
|
left.exec(); |
216 |
|
|
} |
217 |
|
|
else { |
218 |
|
|
int cb; |
219 |
|
|
for (;;) { // Establish action: sum, cumulate, or both |
220 |
|
|
int b = phase; |
221 |
|
|
if ((b & FINISHED) != 0) // already done |
222 |
|
|
return false; |
223 |
|
|
if ((b & CUMULATE) != 0) |
224 |
|
|
cb = FINISHED; |
225 |
|
|
else if (lo == 0) // combine leftmost |
226 |
|
|
cb = (SUMMED|FINISHED); |
227 |
|
|
else |
228 |
|
|
cb = SUMMED; |
229 |
|
|
if (phaseUpdater.compareAndSet(this, b, b|cb)) |
230 |
|
|
break; |
231 |
|
|
} |
232 |
|
|
|
233 |
|
|
if (cb == SUMMED) |
234 |
|
|
out = seqSum(array, lo, hi); |
235 |
|
|
else if (cb == FINISHED) |
236 |
|
|
seqCumulate(array, lo, hi, in); |
237 |
|
|
else if (cb == (SUMMED|FINISHED)) |
238 |
|
|
out = seqCumulate(array, lo, hi, 0L); |
239 |
|
|
|
240 |
|
|
// propagate up |
241 |
|
|
Cumulater ch = this; |
242 |
|
|
Cumulater par = parent; |
243 |
|
|
for (;;) { |
244 |
|
|
if (par == null) { |
245 |
|
|
if ((cb & FINISHED) != 0) |
246 |
|
|
ch.complete(null); |
247 |
|
|
break; |
248 |
|
|
} |
249 |
|
|
int pb = par.phase; |
250 |
|
|
if ((pb & cb & FINISHED) != 0) { // both finished |
251 |
|
|
ch = par; |
252 |
|
|
par = par.parent; |
253 |
|
|
} |
254 |
|
|
else if ((pb & cb & SUMMED) != 0) { // both summed |
255 |
|
|
par.out = par.left.out + par.right.out; |
256 |
|
|
int refork = |
257 |
|
|
((pb & CUMULATE) == 0 && |
258 |
|
|
par.lo == 0)? CUMULATE : 0; |
259 |
|
|
int nextPhase = pb|cb|refork; |
260 |
|
|
if (pb == nextPhase || |
261 |
|
|
phaseUpdater.compareAndSet(par, pb, nextPhase)) { |
262 |
|
|
if (refork != 0) |
263 |
|
|
par.fork(); |
264 |
|
|
cb = SUMMED; // drop finished bit |
265 |
|
|
ch = par; |
266 |
|
|
par = par.parent; |
267 |
|
|
} |
268 |
|
|
} |
269 |
|
|
else if (phaseUpdater.compareAndSet(par, pb, pb|cb)) |
270 |
|
|
break; |
271 |
|
|
} |
272 |
|
|
} |
273 |
|
|
return false; |
274 |
|
|
} |
275 |
|
|
|
276 |
|
|
} |
277 |
|
|
|
278 |
|
|
} |
279 |
|
|
|
280 |
|
|
|