1 /* =============================================================================
4 * -- Implementation of normal k-means clustering algorithm
6 * =============================================================================
11 * ECE Department, Northwestern University
12 * email: wkliao@ece.northwestern.edu
18 * Northwestern University.
24 * University of California, Irvine
27 * =============================================================================
29 * For the license of bayes/sort.h and bayes/sort.c, please see the header
32 * ------------------------------------------------------------------------
34 * For the license of kmeans, please see kmeans/LICENSE.kmeans
36 * ------------------------------------------------------------------------
38 * For the license of ssca2, please see ssca2/COPYRIGHT
40 * ------------------------------------------------------------------------
42 * For the license of lib/mt19937ar.c and lib/mt19937ar.h, please see the
43 * header of the files.
45 * ------------------------------------------------------------------------
47 * For the license of lib/rbtree.h and lib/rbtree.c, please see
48 * lib/LEGALNOTICE.rbtree and lib/LICENSE.rbtree
50 * ------------------------------------------------------------------------
52 * Unless otherwise noted, the following license applies to STAMP files:
54 * Copyright (c) 2007, Stanford University
55 * All rights reserved.
57 * Redistribution and use in source and binary forms, with or without
58 * modification, are permitted provided that the following conditions are
61 * * Redistributions of source code must retain the above copyright
62 * notice, this list of conditions and the following disclaimer.
64 * * Redistributions in binary form must reproduce the above copyright
65 * notice, this list of conditions and the following disclaimer in
66 * the documentation and/or other materials provided with the
69 * * Neither the name of Stanford University nor the names of its
70 * contributors may be used to endorse or promote products derived
71 * from this software without specific prior written permission.
73 * THIS SOFTWARE IS PROVIDED BY STANFORD UNIVERSITY ``AS IS'' AND ANY
74 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
75 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
76 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL STANFORD UNIVERSITY BE LIABLE
77 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
78 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
79 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
80 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
81 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
82 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
83 * THE POSSIBILITY OF SUCH DAMAGE.
85 * =============================================================================
95 /* =============================================================================
97 * =============================================================================
99 public static void work(int myId, GlobalArgs args) {
101 float[][] feature = args.feature;
102 int nfeatures = args.nfeatures;
103 int npoints = args.npoints;
104 int nclusters = args.nclusters;
105 int[] membership = args.membership;
106 float[][] clusters = args.clusters;
107 intwrapper[] new_centers_len = args.new_centers_len;
108 float[][] new_centers = args.new_centers;
110 int index, start, stop;
112 start = myId * CHUNK;
114 //System.out.println("myId= " + myId + " start= " + start + " npoints= " + npoints);
115 while (start < npoints) {
116 stop = (((start + CHUNK) < npoints) ? (start + CHUNK) : npoints);
118 for (int i = start; i < stop; i++) {
119 index = Common.common_findNearestPoint(feature[i],
124 * If membership changes, increase delta by 1.
125 * membership[i] cannot be changed by other threads
127 if (membership[i] != index) {
131 /* Assign the membership to object i */
132 /* membership[i] can't be changed by other thread */
133 membership[i] = index;
135 /* Update new cluster centers : sum of objects located within */
137 new_centers_len[index] = new_centers_len[index] + 1;
138 for (int j = 0; j < nfeatures; j++) {
139 new_centers[index][j] = new_centers[index][j] + feature[i][j];
144 /* Update task queue */
145 if (start + CHUNK < npoints) {
147 start = (int) args.global_i;
148 args.global_i = start + CHUNK;
156 args.global_delta = args.global_delta + delta;
160 /* =============================================================================
162 * =============================================================================
164 public float[][] normal_exec (
166 float[][] feature, /* in: [npoints][nfeatures] */
172 Random randomPtr, /* out: [npoints] */
176 float[][] clusters; /* out: [nclusters][nfeatures] */
178 /* Allocate space for returning variable clusters[] */
179 clusters = new float[nclusters][nfeatures];
181 /* Randomly pick cluster centers */
182 for (int i = 0; i < nclusters; i++) {
183 int n = (int)(randomPtr.random_generate() % npoints);
184 for (int j = 0; j < nfeatures; j++) {
185 clusters[i][j] = feature[n][j];
189 for (int i = 0; i < npoints; i++) {
194 * Need to initialize new_centers_len and new_centers[0] to all 0.
195 * Allocate clusters on different cache lines to reduce false sharing.
197 intwrapper[] new_centers_len = new intwrapper[nclusters];
199 float[][] new_centers = new float[nclusters][nfeatures];
205 args.feature = feature;
206 args.nfeatures = nfeatures;
207 args.npoints = npoints;
208 args.nclusters = nclusters;
209 args.membership = membership;
210 args.clusters = clusters;
211 args.new_centers_len = new_centers_len;
212 args.new_centers = new_centers;
214 args.global_i = nthreads * CHUNK;
215 args.global_delta = delta;
217 //Work in parallel with other threads
220 delta = args.global_delta;
222 /* Replace old cluster centers with new_centers */
223 for (int i = 0; i < nclusters; i++) {
224 for (int j = 0; j < nfeatures; j++) {
225 if (new_centers_len[i] > 0) {
226 clusters[i][j] = new_centers[i][j] / new_centers_len[i];
228 new_centers[i][j] = (float)0.0; /* set back to 0 */
230 new_centers_len[i] = 0; /* set back to 0 */
235 //System.out.println("delta= " + delta + " loop= " + loop + " threshold= " + threshold);
237 } while ((delta > threshold) && (loop++ < 500));
243 * Work done by primary thread in parallel with other threads
245 void thread_work(GlobalArgs args) {
246 Barrier.enterBarrier();
247 Normal.work(0, args); //threadId = 0 because primary thread
248 Barrier.enterBarrier();
252 /* =============================================================================
256 * =============================================================================