Optimize RDRPosTagger to use integer comparisons instead of string co…

…mparisons.

Also reduce the cache-thrashing by deconstructing the tree's nodes into arrays.

third-party/rdrpostagger/build.gradle

@@ -9,6 +9,7 @@ java {
 }
 
 dependencies {
+    implementation libs.trove
 }
 
 test {

third-party/rdrpostagger/src/main/java/com/github/datquocnguyen/FWObject.java

@@ -13,6 +13,9 @@
 public class FWObject
 {
 	public String[] context;
+
+	int[] objectCtxI = new int[13];
+
 	private final static String[] contextPrototype;
 	static {
 		contextPrototype = new String[13];

third-party/rdrpostagger/src/main/java/com/github/datquocnguyen/Node.java

@@ -40,26 +40,12 @@ public void setFatherNode(Node node)
 		this.fatherNode = node;
 	}
 
-	public int countNodes()
-	{
-		int count = 1;
-		if (exceptNode != null) {
-			count += exceptNode.countNodes();
-		}
-		if (ifnotNode != null) {
-			count += ifnotNode.countNodes();
-		}
-		return count;
-	}
-
 	public boolean satisfy(FWObject object)
 	{
 		for (int i = 0; i < 13; i++) {
 			String key = condition.context[i];
-			if (key != null) {
-				if (!key.equals(object.context[i])) {
-					return false;
-				}
+			if (key != null && !key.equals(object.context[i])) { // this is not equivalent to Objects.equals(a,b)
+				return false;
 			}
 		}
 		return true;

third-party/rdrpostagger/src/main/java/com/github/datquocnguyen/RDRPOSTagger.java

@@ -1,8 +1,11 @@
 package com.github.datquocnguyen;
 
+import gnu.trove.map.hash.TObjectIntHashMap;
+
 import java.io.*;
 import java.nio.charset.StandardCharsets;
 import java.nio.file.Path;
+import java.util.Arrays;
 import java.util.HashMap;
 
 /**
@@ -12,22 +15,59 @@
 public class RDRPOSTagger
 {
 	private final HashMap<String, String> FREQDICT;
-	public final Node root;
+	final int OUGHT_TO_BE_ENOUGH = 5000;
+	final int CONTEXT_SIZE = 13;
+
+	// Use dense array representation to reduce the level of indirection
+	// and improve the performance of the tagger
+	int[] conditions = new int[OUGHT_TO_BE_ENOUGH * CONTEXT_SIZE];
+	String[] conclusions = new String[OUGHT_TO_BE_ENOUGH];
+	short[] exceptIdx = new short[OUGHT_TO_BE_ENOUGH];
+	short[] ifNotIdx = new short[OUGHT_TO_BE_ENOUGH];
+	short[] fatherIdx = new short[OUGHT_TO_BE_ENOUGH];
+	byte[] depthL = new byte[OUGHT_TO_BE_ENOUGH];
+
+	short size = 0;
+
+	private final TObjectIntHashMap<String> tagDict = new TObjectIntHashMap<>(10000, 0.75f, -1);
+
+	private short addNode(FWObject condition, String conclusion, byte d) {
+		short idx = size++;
+
+		for (int i = 0; i < CONTEXT_SIZE; i++) {
+			String context = condition.context[i];
+			if (context != null) {
+				tagDict.putIfAbsent(context, tagDict.size());
+
+				conditions[idx * CONTEXT_SIZE + i] = tagDict.get(context);
+			}
+			else {
+				conditions[idx * CONTEXT_SIZE + i] = -1;
+			}
+		}
+
+		conclusions[idx] = conclusion;
+		exceptIdx[idx] = -1;
+		ifNotIdx[idx] = -1;
+		fatherIdx[idx] = -1;
+		depthL[idx] = d;
+
+		return idx;
+	}
 
 	public RDRPOSTagger(Path dictPath, Path rulesFilePath) throws IOException {
 		this.FREQDICT = Utils.getDictionary(dictPath.toString());
+		Arrays.fill(conditions, -1);
 
 		BufferedReader buffer = new BufferedReader(new InputStreamReader(
 				new FileInputStream(rulesFilePath.toFile()), StandardCharsets.UTF_8));
 		String line = buffer.readLine();
 
-		this.root = new Node(new FWObject(false), "NN", null, null, null, 0);
-
-		Node currentNode = this.root;
-		int currentDepth = 0;
+		short currentIdx = addNode(new FWObject(false), "NN", (byte) 0);
+		byte currentDepth = 0;
 
 		while ((line = buffer.readLine()) != null) {
-			int depth = 0;
+			byte depth = 0;
 			for (int i = 0; i <= 6; i++) { // Supposed that the maximum
 				// exception level is up to 6.
 				if (line.charAt(i) == '\t')
@@ -48,53 +88,72 @@ public RDRPOSTagger(Path dictPath, Path rulesFilePath) throws IOException {
 			String conclusion = Utils.getConcreteValue(line.split(" : ")[1]
 					.trim());
 
-			Node node = new Node(condition, conclusion, null, null, null, depth);
+			short newIdx = addNode(condition, conclusion, depth);
 
 			if (depth > currentDepth) {
-				currentNode.setExceptNode(node);
+				exceptIdx[currentIdx] = newIdx;
 			}
 			else if (depth == currentDepth) {
-				currentNode.setIfnotNode(node);
+				ifNotIdx[currentIdx] = newIdx;
 			}
 			else {
-				while (currentNode.depth != depth)
-					currentNode = currentNode.fatherNode;
-				currentNode.setIfnotNode(node);
+				while (depthL[currentIdx] != depth) {
+					currentIdx = fatherIdx[currentIdx];
+				}
+				ifNotIdx[currentIdx] = newIdx;
 			}
-			node.setFatherNode(currentNode);
 
-			currentNode = node;
+			fatherIdx[newIdx] = currentIdx;
+
+			currentIdx = newIdx;
 			currentDepth = depth;
 		}
 		buffer.close();
 	}
 
-	public Node findFiredNode(FWObject object)
+	public String findFiredNode(FWObject object)
 	{
-		Node currentN = root;
-		Node firedN = null;
-		while (true) {
-			if (currentN.satisfy(object)) {
-				firedN = currentN;
-				if (currentN.exceptNode == null) {
-					break;
-				}
-				else {
-					currentN = currentN.exceptNode;
-				}
+		int currentIdx = 0;
+		int firedIdx = -1;
+
+		int[] objCtxI = object.objectCtxI;
+
+		for (int i = 0; i < CONTEXT_SIZE; i++) {
+			objCtxI[i] = tagDict.get(object.context[i]);
+		}
+
+		int[] conditionsL = conditions;
+		short[] exceptIdxL = exceptIdx;
+		short[] ifNotIdxL = ifNotIdx;
+
+		while (currentIdx >= 0) {
+			if (satisfy(objCtxI, conditionsL, currentIdx)) {
+				firedIdx = currentIdx;
+				currentIdx = exceptIdxL[currentIdx];
 			}
 			else {
-				if (currentN.ifnotNode == null) {
-					break;
-				}
-				else {
-					currentN = currentN.ifnotNode;
-				}
+				currentIdx = ifNotIdxL[currentIdx];
 			}
+		}
 
+		if (firedIdx >= 0) {
+			return conclusions[firedIdx];
+		}
+		else {
+			return "";
 		}
+	}
 
-		return firedN;
+	public boolean satisfy(int[] objectCtxI, int[] conditions, int contextIdx)
+	{
+		// This is a good candidate for a vector operation
+		for (int i = 0; i < CONTEXT_SIZE; i++) {
+			int key = conditions[CONTEXT_SIZE *contextIdx + i];
+			if (key >= 0 && key != objectCtxI[i]) {
+				return false;
+			}
+		}
+		return true;
 	}
 
 	public String[] tagsForEnSentence(String[] sentence)
@@ -107,7 +166,7 @@ public String[] tagsForEnSentence(String[] sentence)
 
 		for (int i = 0; i < initialTags.length; i++) {
 			Utils.getObject(object, sentence, initialTags, initialTags.length, i);
-			tags[i] = findFiredNode(object).conclusion;
+			tags[i] = findFiredNode(object);
 		}
 
 		return tags;

third-party/rdrpostagger/src/main/java/com/github/datquocnguyen/Utils.java

@@ -9,6 +9,8 @@
 import java.util.ArrayList;
 import java.util.HashMap;
 import java.util.List;
+import java.util.Map;
+import java.util.function.Function;
 
 /**
  * @author DatQuocNguyen
@@ -69,6 +71,7 @@ public static boolean isAbbre(String word)
 		return true;
 	}
 
+	static Map<String, String> conditionInstancePool = new HashMap<>();
 	public static FWObject getCondition(String strCondition)
 	{
 		FWObject condition = new FWObject(false);
@@ -120,6 +123,16 @@ else if (key.equals("suffixL4")) {
 			}
 		}
 
+		// pool the conditions to increase the chances the data is in cache
+		// when comparing later
+
+		for (var i = 0; i < condition.context.length; i++) {
+			if (condition.context[i] != null) {
+				condition.context[i] = conditionInstancePool
+						.computeIfAbsent(condition.context[i], Function.identity());
+			}
+		}
+
 		return condition;
 	}