Replace multiplication with addition

kgryte · kgryte · commit 3609273204ef · 2018-07-11T04:30:23.000+01:00
diff --git a/lib/node_modules/@stdlib/ml/incr/kmeans/lib/init_kmeansplusplus.js b/lib/node_modules/@stdlib/ml/incr/kmeans/lib/init_kmeansplusplus.js
@@ -195,9 +195,11 @@ function kmeansplusplus( out, buffer, metric, normalize, trials, seed ) {
 
 	// Create a strided array for storing closest centroid results:
 	dhash = new Array( npts*2 );
+	ind = 0;
 	for ( i = 0; i < npts; i++ ) {
-		dhash[ 2*i ] = PINF; // squared distance
-		dhash[ (2*i)+1 ] = 0; // index of the closest centroid
+		dhash[ ind ] = PINF; // squared distance
+		dhash[ ind+1 ] = 0; // index of the closest centroid
+		ind += 2; // +stride
 	}
 	// Create a scratch array for storing cumulative probabilities:
 	probs = new Array( npts );
@@ -207,20 +209,23 @@ function kmeansplusplus( out, buffer, metric, normalize, trials, seed ) {
 		// Note: instead of repeatedly computing centroid distances for each data point, we only need to compute the distances for the most recent centroid and to maintain a hash of closest distance results...
 		dapply( d2, dist, npts, ndims, buffer, centroids[ j-1 ] );
 		csum = 0.0; // total cumulative distance
+		ind = 0;
 		for ( i = 0; i < npts; i++ ) {
-			ind = 2 * i;
 			if ( d2[ i ] < dhash[ ind ] ) {
 				dhash[ ind ] = d2[ i ];
 				dhash[ ind+1 ] = j - 1;
 				csum += d2[ i ];
 			} else {
 				csum += dhash[ ind ];
 			}
+			ind += 2; // +stride
 		}
 		// Compute the cumulative probabilities...
 		probs[ 0 ] = dhash[ 0 ] / csum;
+		ind = 2;
 		for ( i = 1; i < npts; i++ ) {
-			probs[ i ] = probs[ i-1 ] + ( dhash[ 2*i ] / csum );
+			probs[ i ] = probs[ i-1 ] + ( dhash[ ind ] / csum );
+			ind += 2; // +stride
 		}
 		// Based Arthur's and Vassilvitskii's paper "kmeans++: The Advantages of Careful Seeding" (see conclusion), randomly select candidate centroids and pick the candidate which minimizes the total squared distance...
 		bsum = PINF; // best sum
@@ -242,14 +247,15 @@ function kmeansplusplus( out, buffer, metric, normalize, trials, seed ) {
 			// Compute the sum of squared distances were we to include the candidate centroid...
 			csum = 0.0;
 			offsetC = sb1 * c;
+			ind = 0;
 			for ( i = 0; i < npts; i++ ) {
 				d = dist( ndims, buf, 1, sb1*i, buf, 1, offsetC ); // Magic number `1` for stride as matrix is row-major single-segment contiguous
-				ind = 2 * i;
 				if ( d < dhash[ ind ] ) {
 					csum += d;
 				} else {
 					csum += dhash[ ind ];
 				}
+				ind += 2; // +stride
 			}
 			// Determine if the candidate is the best candidate we have seen thus far...
 			if ( csum < bsum ) {
