Use bloom filter in multi bigram map.

Evaluated with previous word "this".
without bloom filter (use only hash_map):
Total 147792.34 (sum of others 147771.57)
with bloom filter:
Total 145900.64 (sum of others 145874.30)
always read binary dictionary:
Total 148603.14 (sum of others 148579.90)

Bug: 8592527
Change-Id: I821dc39454543826adb73b9eeeef6408fad8ae28
This commit is contained in:
Keisuke Kuroynagi 2013-06-14 20:35:41 +09:00
parent 4f19193560
commit 1ff81e8890
6 changed files with 121 additions and 46 deletions

View File

@ -57,9 +57,11 @@ LATIN_IME_CORE_SRC_FILES := \
binary_dictionary_format_utils.cpp \
binary_dictionary_header.cpp \
binary_dictionary_header_reading_utils.cpp \
bloom_filter.cpp \
byte_array_utils.cpp \
dictionary.cpp \
digraph_utils.cpp) \
digraph_utils.cpp \
multi_bigram_map.cpp) \
$(addprefix suggest/core/layout/, \
additional_proximity_chars.cpp \
proximity_info.cpp \

View File

@ -300,33 +300,6 @@ static inline void prof_out(void) {
#define DIC_NODES_CACHE_INITIAL_QUEUE_ID_CACHE_FOR_CONTINUOUS_SUGGESTION 3
#define DIC_NODES_CACHE_PRIORITY_QUEUES_SIZE 4
// 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
// Max number of bigram maps (previous word contexts) to be cached. Increasing this number could
// improve bigram lookup speed for multi-word suggestions, but at the cost of more memory usage.
// Also, there are diminishing returns since the most frequently used bigrams are typically near
// the beginning of the input and are thus the first ones to be cached. Note that these bigrams
// are reset for each new composing word.
#define MAX_CACHED_PREV_WORDS_IN_BIGRAM_MAP 25
// Most common previous word contexts currently have 100 bigrams
#define DEFAULT_HASH_MAP_SIZE_FOR_EACH_BIGRAM_MAP 100
template<typename T> AK_FORCE_INLINE const T &min(const T &a, const T &b) { return a < b ? a : b; }
template<typename T> AK_FORCE_INLINE const T &max(const T &a, const T &b) { return a > b ? a : b; }

View File

@ -0,0 +1,25 @@
/*
* Copyright (C) 2013, The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "suggest/core/dictionary/bloom_filter.h"
namespace latinime {
// Must be smaller than BIGRAM_FILTER_BYTE_SIZE * 8, and preferably prime. 1021 is the largest
// prime under 128 * 8.
const int BloomFilter::BIGRAM_FILTER_MODULO = 1021;
} // namespace latinime

View File

@ -23,16 +23,48 @@
namespace latinime {
// TODO: uint32_t position
static inline void setInFilter(uint8_t *filter, const int32_t position) {
const uint32_t bucket = static_cast<uint32_t>(position % BIGRAM_FILTER_MODULO);
filter[bucket >> 3] |= static_cast<uint8_t>(1 << (bucket & 0x7));
}
// This bloom filter is used for optimizing bigram retrieval.
// Execution times with previous word "this" are as follows:
// without bloom filter (use only hash_map):
// Total 147792.34 (sum of others 147771.57)
// with bloom filter:
// Total 145900.64 (sum of others 145874.30)
// always read binary dictionary:
// Total 148603.14 (sum of others 148579.90)
class BloomFilter {
public:
BloomFilter() {
ASSERT(BIGRAM_FILTER_BYTE_SIZE * 8 >= BIGRAM_FILTER_MODULO);
}
// TODO: uint32_t position
static inline bool isInFilter(const uint8_t *filter, const int32_t position) {
const uint32_t bucket = static_cast<uint32_t>(position % BIGRAM_FILTER_MODULO);
return filter[bucket >> 3] & static_cast<uint8_t>(1 << (bucket & 0x7));
}
// TODO: uint32_t position
AK_FORCE_INLINE void setInFilter(const int32_t position) {
const uint32_t bucket = static_cast<uint32_t>(position % BIGRAM_FILTER_MODULO);
mFilter[bucket >> 3] |= static_cast<uint8_t>(1 << (bucket & 0x7));
}
// TODO: uint32_t position
AK_FORCE_INLINE bool isInFilter(const int32_t position) const {
const uint32_t bucket = static_cast<uint32_t>(position % BIGRAM_FILTER_MODULO);
return (mFilter[bucket >> 3] & static_cast<uint8_t>(1 << (bucket & 0x7))) != 0;
}
private:
// 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%.
// This is assigned here because it is used for array size.
static const int BIGRAM_FILTER_BYTE_SIZE = 128;
static const int BIGRAM_FILTER_MODULO;
uint8_t mFilter[BIGRAM_FILTER_BYTE_SIZE];
};
} // namespace latinime
#endif // LATINIME_BLOOM_FILTER_H

View File

@ -0,0 +1,33 @@
/*
* Copyright (C) 2013, The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "suggest/core/dictionary/multi_bigram_map.h"
#include <cstddef>
namespace latinime {
// Max number of bigram maps (previous word contexts) to be cached. Increasing this number
// could improve bigram lookup speed for multi-word suggestions, but at the cost of more memory
// usage. Also, there are diminishing returns since the most frequently used bigrams are
// typically near the beginning of the input and are thus the first ones to be cached. Note
// that these bigrams are reset for each new composing word.
const size_t MultiBigramMap::MAX_CACHED_PREV_WORDS_IN_BIGRAM_MAP = 25;
// Most common previous word contexts currently have 100 bigrams
const int MultiBigramMap::BigramMap::DEFAULT_HASH_MAP_SIZE_FOR_EACH_BIGRAM_MAP = 100;
} // namespace latinime

View File

@ -17,10 +17,13 @@
#ifndef LATINIME_MULTI_BIGRAM_MAP_H
#define LATINIME_MULTI_BIGRAM_MAP_H
#include <cstddef>
#include "defines.h"
#include "suggest/core/dictionary/binary_dictionary_bigrams_iterator.h"
#include "suggest/core/dictionary/binary_dictionary_info.h"
#include "suggest/core/dictionary/binary_format.h"
#include "suggest/core/dictionary/bloom_filter.h"
#include "utils/hash_map_compat.h"
namespace latinime {
@ -60,7 +63,7 @@ class MultiBigramMap {
class BigramMap {
public:
BigramMap() : mBigramMap(DEFAULT_HASH_MAP_SIZE_FOR_EACH_BIGRAM_MAP) {}
BigramMap() : mBigramMap(DEFAULT_HASH_MAP_SIZE_FOR_EACH_BIGRAM_MAP), mBloomFilter() {}
~BigramMap() {}
void init(const BinaryDictionaryInfo *const binaryDictionaryInfo, const int nodePos) {
@ -73,24 +76,30 @@ class MultiBigramMap {
bigramsIt.hasNext(); /* no-op */) {
bigramsIt.next();
mBigramMap[bigramsIt.getBigramPos()] = bigramsIt.getProbability();
mBloomFilter.setInFilter(bigramsIt.getBigramPos());
}
}
AK_FORCE_INLINE int getBigramProbability(
const int nextWordPosition, const int unigramProbability) const {
const hash_map_compat<int, int>::const_iterator bigramProbabilityIt =
mBigramMap.find(nextWordPosition);
if (bigramProbabilityIt != mBigramMap.end()) {
const int bigramProbability = bigramProbabilityIt->second;
return ProbabilityUtils::computeProbabilityForBigram(
unigramProbability, bigramProbability);
if (mBloomFilter.isInFilter(nextWordPosition)) {
const hash_map_compat<int, int>::const_iterator bigramProbabilityIt =
mBigramMap.find(nextWordPosition);
if (bigramProbabilityIt != mBigramMap.end()) {
const int bigramProbability = bigramProbabilityIt->second;
return ProbabilityUtils::computeProbabilityForBigram(
unigramProbability, bigramProbability);
}
}
return ProbabilityUtils::backoff(unigramProbability);
}
private:
// Note: Default copy constructor needed for use in hash_map.
// NOTE: The BigramMap class doesn't use DISALLOW_COPY_AND_ASSIGN() because its default
// copy constructor is needed for use in hash_map.
static const int DEFAULT_HASH_MAP_SIZE_FOR_EACH_BIGRAM_MAP;
hash_map_compat<int, int> mBigramMap;
BloomFilter mBloomFilter;
};
AK_FORCE_INLINE void addBigramsForWordPosition(
@ -117,6 +126,7 @@ class MultiBigramMap {
return ProbabilityUtils::backoff(unigramProbability);
}
static const size_t MAX_CACHED_PREV_WORDS_IN_BIGRAM_MAP;
hash_map_compat<int, BigramMap> mBigramMaps;
};
} // namespace latinime