1 |
mojays |
2 |
package appl.parallel.spmd.split; |
2 |
|
|
|
3 |
|
|
import java.awt.Point; |
4 |
|
|
import java.awt.Rectangle; |
5 |
|
|
import java.util.Vector; |
6 |
|
|
|
7 |
|
|
import appl.parallel.util.Helper; |
8 |
mojays |
114 |
import de.schmitzm.geotools.data.WritableGrid; |
9 |
mojays |
2 |
|
10 |
|
|
/** |
11 |
|
|
* Responsible for splitting a 2D Area (e.g a {@link WritableGrid}) in a 2D |
12 |
|
|
* fashion. The splitting is irregular, meaning that the rectangles are not |
13 |
|
|
* equally sized, but partitioned according to the given weights. |
14 |
|
|
* Splitting is as follows: <br><br> |
15 |
|
|
* 1 to 4 Partitions: 1 row <br> |
16 |
|
|
* 4 to 8 Partitions: 2 rows <br> |
17 |
|
|
* 9 to 15 Partitions: 3 rows <br> |
18 |
|
|
* and so on (depending on the square root of the partititons)<br> |
19 |
|
|
* <br> |
20 |
|
|
* For partitioning the this class strongly depends on {@link SplitMap1DHorizontal} |
21 |
|
|
* and {@link SplitMap1DVertical}. First the squareroot of the number of |
22 |
|
|
* participating resources is taken to determine how many rows should be |
23 |
|
|
* created. After this the height of the rows is weighted by adding the weight |
24 |
|
|
* of the resources in that horizontal partition. The Grid is then splitted |
25 |
|
|
* horizontal. Finally the horizontal partitions are splitted vertical according |
26 |
|
|
* to the relative weight to each other. This is only a a virtual split (a map |
27 |
|
|
* of a split). <br> |
28 |
|
|
* A neighborhood range can be specified to create partitions that overlap each |
29 |
|
|
* other. This overlapping can be done in two different ways: <br> |
30 |
|
|
* <br> |
31 |
|
|
* <b>Inboxing</b>(default):<br> |
32 |
|
|
* <br> |
33 |
|
|
* Inboxing means, that the neighborhood area is not part of the calculation |
34 |
|
|
* area. <br> |
35 |
|
|
* <br> |
36 |
|
|
* <b>Outboxing:</b><br> |
37 |
|
|
* Outboxing means that the neighborhood area is part of the calculation area. |
38 |
|
|
* <br> |
39 |
|
|
* <br> |
40 |
|
|
* Note that this may also be applied to three dimensional data structures (a 2D |
41 |
|
|
* splitting of a 3D datatype). |
42 |
|
|
* |
43 |
|
|
* @author Dominik Appl |
44 |
|
|
*/ |
45 |
|
|
|
46 |
|
|
public class SplitMap2D extends AbstractSplitMap implements SplitMap { |
47 |
|
|
|
48 |
|
|
private SplitMap1DHorizontal horizontalMap; |
49 |
|
|
|
50 |
|
|
private SplitMap1DVertical[] verticalMaps; |
51 |
|
|
|
52 |
|
|
public SplitMap2D(int width, int height, int neighborhoodRange, |
53 |
|
|
int noOfPartitions, NeighborhoodBoxingMode boxingMode) { |
54 |
|
|
super(width, height, neighborhoodRange, noOfPartitions, boxingMode); |
55 |
|
|
} |
56 |
|
|
|
57 |
|
|
/** |
58 |
|
|
* needed for serialization |
59 |
|
|
*/ |
60 |
|
|
public SplitMap2D() { |
61 |
|
|
} |
62 |
|
|
|
63 |
|
|
/* |
64 |
|
|
* (non-Javadoc) |
65 |
|
|
* |
66 |
|
|
* @see appl.parallel.spmd.split.AbstractSplitMap#makeMap() |
67 |
|
|
*/ |
68 |
|
|
public void makeMap() { |
69 |
|
|
// take the squareroot of the number of participating resources to |
70 |
|
|
// determine the number of rows: |
71 |
|
|
int noOfRows = (int) Math.sqrt(noOfPartitions); |
72 |
|
|
// calculate weights for the rows (sum of the weights of the partitions |
73 |
|
|
// inside that row) |
74 |
|
|
double rowWeights[] = new double[noOfRows]; |
75 |
|
|
int noOfCols[] = new int[noOfRows]; // noOfCols per row |
76 |
|
|
for (int row = 0; row < noOfRows; row++) { |
77 |
|
|
// calculate no. of cols in the row: |
78 |
|
|
noOfCols[row] = (noOfPartitions / noOfRows); |
79 |
|
|
// the last row may contain a greater number of cols |
80 |
|
|
if (row == noOfRows - 1) |
81 |
|
|
noOfCols[row] = noOfPartitions |
82 |
|
|
- ((noOfRows - 1) * noOfCols[row]); |
83 |
|
|
for (int col = 0; col < noOfCols[row]; col++) |
84 |
|
|
rowWeights[row] += weights[row * col + col]; |
85 |
|
|
} |
86 |
|
|
|
87 |
|
|
// create the horizontal Splitmap |
88 |
|
|
horizontalMap = new SplitMap1DHorizontal(globalWidth, globalHeight, |
89 |
|
|
neighborhoodRange, noOfRows, boxingMode); |
90 |
|
|
horizontalMap.setWeights(rowWeights); |
91 |
|
|
horizontalMap.makeMap(); |
92 |
|
|
|
93 |
|
|
// for each horizontal partition create a vertical split |
94 |
|
|
verticalMaps = new SplitMap1DVertical[noOfRows]; |
95 |
|
|
for (int row = 0; row < noOfRows; row++) { |
96 |
|
|
//notice: all splitmaps start at (0,0), later a conversion must be made |
97 |
|
|
verticalMaps[row] = new SplitMap1DVertical((int) horizontalMap |
98 |
|
|
.getGlobalBounds().getWidth(), (int) horizontalMap |
99 |
|
|
.getPartitionCalculationBounds(row).getHeight(), neighborhoodRange, |
100 |
|
|
noOfCols[row], boxingMode); |
101 |
|
|
int ratings[] = new int[noOfCols[row]]; |
102 |
|
|
for (int col = 0; col < noOfCols[row]; col++) { |
103 |
|
|
{ |
104 |
|
|
// the rating of the col is the weight of the col * 100000 |
105 |
|
|
// (ratings must be positive integers) |
106 |
|
|
ratings[col] = (int) (weights[row * col + col] * 1000000); |
107 |
|
|
|
108 |
|
|
} |
109 |
|
|
|
110 |
|
|
verticalMaps[row].setWeights(Helper.calculateWeights(ratings)); |
111 |
|
|
verticalMaps[row].makeMap(); |
112 |
|
|
} |
113 |
|
|
} |
114 |
|
|
// now we have partitioned the grid. Lets assign the calculation and |
115 |
|
|
// neighborhoodareas: |
116 |
|
|
for (int i = 0; i < noOfPartitions; i++) { |
117 |
|
|
int row = getRowForIdx(i); |
118 |
|
|
int col = getColForIdx(i); |
119 |
|
|
partitionCalculationBounds[i] = verticalMaps[row] |
120 |
|
|
.getPartitionCalculationBounds(col); |
121 |
|
|
//because all splitmaps start with (0,0) we must move |
122 |
|
|
//the rectangles to the right position |
123 |
|
|
int x = (int) partitionCalculationBounds[i].getX(); //the x value remains unchanged |
124 |
|
|
partitionCalculationBounds[i].setLocation( |
125 |
|
|
new Point(x,(int) horizontalMap.getPartitionCalculationBounds(row).getY())); |
126 |
|
|
// if there is only one partition there are no explicit neighborhood |
127 |
|
|
// bounds |
128 |
|
|
if (noOfPartitions == 1) |
129 |
|
|
partitionNeighborhoodBounds[i] = partitionCalculationBounds[i]; |
130 |
|
|
// else simply extend the Rectangle with the Neighborhoodbounds |
131 |
|
|
// created calculation Area (@see AbstractSplitMap) |
132 |
|
|
else |
133 |
|
|
partitionNeighborhoodBounds[i] = new Rectangle( |
134 |
|
|
(int) partitionCalculationBounds[i].getX() |
135 |
|
|
- neighborhoodRange, |
136 |
|
|
(int) partitionCalculationBounds[i].getY() |
137 |
|
|
- neighborhoodRange, |
138 |
|
|
(int) (partitionCalculationBounds[i].getWidth() + 2 * neighborhoodRange), |
139 |
|
|
(int) (partitionCalculationBounds[i].getHeight() + 2 * neighborhoodRange)); |
140 |
|
|
|
141 |
|
|
// cut of the obverlapping sections which are out of the grid: |
142 |
|
|
partitionNeighborhoodBounds[i] = partitionNeighborhoodBounds[i] |
143 |
|
|
.intersection(globalBounds); |
144 |
|
|
|
145 |
|
|
// the partitionCalculation bounds were created for inboxing: |
146 |
|
|
if (boxingMode == NeighborhoodBoxingMode.outBoxing) |
147 |
|
|
partitionCalculationBounds[i] = partitionNeighborhoodBounds[i]; |
148 |
|
|
} |
149 |
|
|
} |
150 |
|
|
|
151 |
|
|
/** |
152 |
|
|
* @param i |
153 |
|
|
* @return |
154 |
|
|
*/ |
155 |
|
|
private int getColForIdx(int index) { |
156 |
|
|
int noOfRows = (int) Math.sqrt(noOfPartitions); |
157 |
|
|
int avgNoOfCols = noOfPartitions / noOfRows; |
158 |
|
|
int currentRow = getRowForIdx(index); |
159 |
|
|
int returnValue = (index - currentRow*avgNoOfCols); |
160 |
|
|
return returnValue; |
161 |
|
|
} |
162 |
|
|
|
163 |
|
|
private int getRowForIdx(int index) { |
164 |
|
|
int noOfRows = (int) Math.sqrt(noOfPartitions); |
165 |
|
|
int avgNoOfCols = noOfPartitions / noOfRows; |
166 |
|
|
int returnValue = index / avgNoOfCols; |
167 |
|
|
//special handling for the last row (contains more partitions) |
168 |
|
|
if(noOfRows==returnValue) |
169 |
|
|
return noOfRows-1; |
170 |
|
|
else return returnValue; |
171 |
|
|
} |
172 |
|
|
|
173 |
|
|
/* |
174 |
|
|
* (non-Javadoc) |
175 |
|
|
* |
176 |
|
|
* @see appl.parallel.spmd.split.SplitMap#getNeighborsForPosition(int) |
177 |
|
|
*/ |
178 |
|
|
public int[] getNeighborsForPosition(int pos) { |
179 |
|
|
Vector neighbors = new Vector(); |
180 |
|
|
// simply intersect with all partitions to find the neighbors: |
181 |
|
|
for (int i = 0; i < noOfPartitions; i++) { |
182 |
|
|
if (i == pos) |
183 |
|
|
continue; |
184 |
|
|
if (partitionNeighborhoodBounds[i] |
185 |
|
|
.intersects(partitionNeighborhoodBounds[pos])) |
186 |
|
|
neighbors.add(pos); |
187 |
|
|
} |
188 |
|
|
int results[] = new int[neighbors.size()]; |
189 |
|
|
// copy into array |
190 |
|
|
for (int i = 0; i < results.length; i++) { |
191 |
|
|
results[i] = (Integer) neighbors.get(i); |
192 |
|
|
} |
193 |
|
|
return results; |
194 |
|
|
} |
195 |
|
|
|
196 |
|
|
/* |
197 |
|
|
* (non-Javadoc) |
198 |
|
|
* |
199 |
|
|
* @see appl.parallel.spmd.split.AbstractSplitMap#getDescription() |
200 |
|
|
*/ |
201 |
|
|
@Override |
202 |
|
|
public String getDescription() { |
203 |
|
|
return "2D-irregular Splitmap"; |
204 |
|
|
} |
205 |
|
|
|
206 |
|
|
} |