Fill in the bloom filter for bigram lookup.

Bug: 6313806
Change-Id: Ib79e14f6f8b241f053da6069c15f19c71084317e

4 files changed, 34 insertions, 6 deletions

diff --git a/native/jni/src/bigram_dictionary.cpp b/native/jni/src/bigram_dictionary.cpp
index 67f96281d..d120261ae 100644
--- a/native/jni/src/bigram_dictionary.cpp
+++ b/native/jni/src/bigram_dictionary.cpp
@@ -153,8 +153,14 @@ int BigramDictionary::getBigramListPositionForWord(const int32_t *prevWord,
     return pos;
 }
 
-void BigramDictionary::fillBigramAddressToFrequencyMap(const int32_t *prevWord,
-        const int prevWordLength, std::map<int, int> *map) {
+static inline void setInFilter(uint8_t *filter, const int position) {
+    const unsigned int bucket = position % BIGRAM_FILTER_MODULO;
+    filter[bucket >> 3] |= (1 << (bucket & 0x7));
+}
+
+void BigramDictionary::fillBigramAddressToFrequencyMapAndFilter(const int32_t *prevWord,
+        const int prevWordLength, std::map<int, int> *map, uint8_t *filter) {
+    memset(filter, 0, BIGRAM_FILTER_BYTE_SIZE);
     const uint8_t* const root = DICT;
     int pos = getBigramListPositionForWord(prevWord, prevWordLength);
     if (0 == pos) return;
@@ -166,6 +172,7 @@ void BigramDictionary::fillBigramAddressToFrequencyMap(const int32_t *prevWord,
         const int bigramPos = BinaryFormat::getAttributeAddressAndForwardPointer(root, bigramFlags,
                 &pos);
         (*map)[bigramPos] = frequency;
+        setInFilter(filter, bigramPos);
     } while (0 != (UnigramDictionary::FLAG_ATTRIBUTE_HAS_NEXT & bigramFlags));
 }
 
diff --git a/native/jni/src/bigram_dictionary.h b/native/jni/src/bigram_dictionary.h
index b1233215b..7328d5828 100644
--- a/native/jni/src/bigram_dictionary.h
+++ b/native/jni/src/bigram_dictionary.h
@@ -20,6 +20,8 @@
 #include <map>
 #include <stdint.h>
 
+#include "defines.h"
+
 namespace latinime {
 
 class Dictionary;
@@ -29,8 +31,8 @@ class BigramDictionary {
     int getBigrams(const int32_t *word, int length, int *codes, int codesSize,
             unsigned short *outWords, int *frequencies, int maxWordLength, int maxBigrams);
     int getBigramListPositionForWord(const int32_t *prevWord, const int prevWordLength);
-    void fillBigramAddressToFrequencyMap(const int32_t *prevWord, const int prevWordLength,
-            std::map<int, int> *map);
+    void fillBigramAddressToFrequencyMapAndFilter(const int32_t *prevWord, const int prevWordLength,
+            std::map<int, int> *map, uint8_t *filter);
     ~BigramDictionary();
  private:
     bool addWordBigram(unsigned short *word, int length, int frequency);
diff --git a/native/jni/src/defines.h b/native/jni/src/defines.h
index c99f8a8b2..cb3dbb115 100644
--- a/native/jni/src/defines.h
+++ b/native/jni/src/defines.h
@@ -241,6 +241,24 @@ static inline void prof_out(void) {
 #define MIN_USER_TYPED_LENGTH_FOR_MULTIPLE_WORD_SUGGESTION 3
 #define MIN_USER_TYPED_LENGTH_FOR_EXCESSIVE_CHARACTER_SUGGESTION 3
 
+// Size, in bytes, of the bloom filter index for bigrams
+// 128 gives us 1024 buckets. The probability of false positive is (1 - e ** (-kn/m))**k,
+// where k is the number of hash functions, n the number of bigrams, and m the number of
+// bits we can test.
+// At the moment 100 is the maximum number of bigrams for a word with the current
+// dictionaries, so n = 100. 1024 buckets give us m = 1024.
+// With 1 hash function, our false positive rate is about 9.3%, which should be enough for
+// our uses since we are only using this to increase average performance. For the record,
+// k = 2 gives 3.1% and k = 3 gives 1.6%. With k = 1, making m = 2048 gives 4.8%,
+// and m = 4096 gives 2.4%.
+#define BIGRAM_FILTER_BYTE_SIZE 128
+// Must be smaller than BIGRAM_FILTER_BYTE_SIZE * 8, and preferably prime. 1021 is the largest
+// prime under 128 * 8.
+#define BIGRAM_FILTER_MODULO 1021
+#if BIGRAM_FILTER_BYTE_SIZE * 8 < BIGRAM_FILTER_MODULO
+#error "BIGRAM_FILTER_MODULO is larger than BIGRAM_FILTER_BYTE_SIZE"
+#endif
+
 template<typename T> inline T min(T a, T b) { return a < b ? a : b; }
 template<typename T> inline T max(T a, T b) { return a > b ? a : b; }
 
diff --git a/native/jni/src/dictionary.h b/native/jni/src/dictionary.h
index a2b0491c5..8bdd77165 100644
--- a/native/jni/src/dictionary.h
+++ b/native/jni/src/dictionary.h
@@ -42,8 +42,9 @@ class Dictionary {
         const int bigramListPosition = !prevWordChars ? 0
                 : mBigramDictionary->getBigramListPositionForWord(prevWordChars, prevWordLength);
         std::map<int, int> bigramMap;
-        mBigramDictionary->fillBigramAddressToFrequencyMap(prevWordChars, prevWordLength,
-                &bigramMap);
+        uint8_t bigramFilter[BIGRAM_FILTER_BYTE_SIZE];
+        mBigramDictionary->fillBigramAddressToFrequencyMapAndFilter(prevWordChars,
+                prevWordLength, &bigramMap, bigramFilter);
         return mUnigramDictionary->getSuggestions(proximityInfo, mWordsPriorityQueuePool,
                 mCorrection, xcoordinates, ycoordinates, codes, codesSize, bigramListPosition,
                 useFullEditDistance, outWords, frequencies);