michael@0: // Copyright 2013 Google Inc. All Rights Reserved.
michael@0: //
michael@0: // Licensed under the Apache License, Version 2.0 (the "License");
michael@0: // you may not use this file except in compliance with the License.
michael@0: // You may obtain a copy of the License at
michael@0: //
michael@0: //     http://www.apache.org/licenses/LICENSE-2.0
michael@0: //
michael@0: // Unless required by applicable law or agreed to in writing, software
michael@0: // distributed under the License is distributed on an "AS IS" BASIS,
michael@0: // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
michael@0: // See the License for the specific language governing permissions and
michael@0: // limitations under the License.
michael@0: 
michael@0: //
michael@0: // Author: dsites@google.com (Dick Sites)
michael@0: //
michael@0: 
michael@0: #include "cldutil_shared.h"
michael@0: #include <string>
michael@0: 
michael@0: #include "cld2tablesummary.h"
michael@0: #include "integral_types.h"
michael@0: #include "port.h"
michael@0: #include "utf8statetable.h"
michael@0: 
michael@0: namespace CLD2 {
michael@0: 
michael@0: // Runtime routines for hashing, looking up, and scoring
michael@0: // unigrams (CJK), bigrams (CJK), quadgrams, and octagrams.
michael@0: // Unigrams and bigrams are for CJK languages only, including simplified/
michael@0: // traditional Chinese, Japanese, Korean, Vietnamese Han characters, and
michael@0: // Zhuang Han characters. Surrounding spaces are not considered.
michael@0: // Quadgrams and octagrams for for non-CJK and include two bits indicating
michael@0: // preceding and trailing spaces (word boundaries).
michael@0: 
michael@0: 
michael@0: // Indicator bits for leading/trailing space around quad/octagram
michael@0: // NOTE: 4444 bits are chosen to flip constant bits in hash of four chars of
michael@0: // 1-, 2-, or 3-bytes each.
michael@0: static const uint32 kPreSpaceIndicator =  0x00004444;
michael@0: static const uint32 kPostSpaceIndicator = 0x44440000;
michael@0: 
michael@0: // Little-endian masks for 0..24 bytes picked up as uint32's
michael@0: static const uint32 kWordMask0[4] = {
michael@0:   0xFFFFFFFF, 0x000000FF, 0x0000FFFF, 0x00FFFFFF
michael@0: };
michael@0: 
michael@0: static const int kMinCJKUTF8CharBytes = 3;
michael@0: 
michael@0: static const int kMinGramCount = 3;
michael@0: static const int kMaxGramCount = 16;
michael@0: 
michael@0: static const int UTFmax = 4;        // Max number of bytes in a UTF-8 character
michael@0: 
michael@0: 
michael@0: // Routines to access a hash table of <key:wordhash, value:probs> pairs
michael@0: // Buckets have 4-byte wordhash for sizes < 32K buckets, but only
michael@0: // 2-byte wordhash for sizes >= 32K buckets, with other wordhash bits used as
michael@0: // bucket subscript.
michael@0: // Probs is a packed: three languages plus a subscript for probability table
michael@0: // Buckets have all the keys together, then all the values.Key array never
michael@0: // crosses a cache-line boundary, so no-match case takes exactly one cache miss.
michael@0: // Match case may sometimes take an additional cache miss on value access.
michael@0: //
michael@0: // Other possibilites include 5 or 10 6-byte entries plus pad to make 32 or 64
michael@0: // byte buckets with single cache miss.
michael@0: // Or 2-byte key and 6-byte value, allowing 5 languages instead  of three.
michael@0: 
michael@0: 
michael@0: //----------------------------------------------------------------------------//
michael@0: // Hashing groups of 1/2/4/8 letters, perhaps with spaces or underscores      //
michael@0: //----------------------------------------------------------------------------//
michael@0: 
michael@0: // Design principles for these hash functions
michael@0: // - Few operations
michael@0: // - Handle 1-, 2-, and 3-byte UTF-8 scripts, ignoring intermixing except in
michael@0: //   Latin script expect 1- and 2-byte mixtures.
michael@0: // - Last byte of each character has about 5 bits of information
michael@0: // - Spread good bits around so they can interact in at least two ways
michael@0: //   with other characters
michael@0: // - Use add for additional mixing thorugh carries
michael@0: 
michael@0: // CJK Three-byte bigram
michael@0: //   ....dddd..cccccc..bbbbbb....aaaa
michael@0: //   ..................ffffff..eeeeee
michael@0: // make
michael@0: //   ....dddd..cccccc..bbbbbb....aaaa
michael@0: //   000....dddd..cccccc..bbbbbb....a
michael@0: //   ..................ffffff..eeeeee
michael@0: //   ffffff..eeeeee000000000000000000
michael@0: //
michael@0: // CJK Four-byte bigram
michael@0: //   ..dddddd..cccccc....bbbb....aaaa
michael@0: //   ..hhhhhh..gggggg....ffff....eeee
michael@0: // make
michael@0: //   ..dddddd..cccccc....bbbb....aaaa
michael@0: //   000..dddddd..cccccc....bbbb....a
michael@0: //   ..hhhhhh..gggggg....ffff....eeee
michael@0: //   ..ffff....eeee000000000000000000
michael@0: 
michael@0: // BIGRAM
michael@0: // Pick up 1..8 bytes and hash them via mask/shift/add. NO pre/post
michael@0: // OVERSHOOTS up to 3 bytes
michael@0: // For runtime use of tables
michael@0: // Does X86 unaligned loads
michael@0: uint32 BiHashV2(const char* word_ptr, int bytecount) {
michael@0:   if (bytecount == 0) {return 0;}
michael@0:   const uint32* word_ptr32 = reinterpret_cast<const uint32*>(word_ptr);
michael@0:   uint32 word0, word1;
michael@0:   if (bytecount <= 4) {
michael@0:     word0 = UNALIGNED_LOAD32(word_ptr32) & kWordMask0[bytecount & 3];
michael@0:     word0 = word0 ^ (word0 >> 3);
michael@0:     return word0;
michael@0:   }
michael@0:   // Else do 8 bytes
michael@0:   word0 = UNALIGNED_LOAD32(word_ptr32);
michael@0:   word0 = word0 ^ (word0 >> 3);
michael@0:   word1 = UNALIGNED_LOAD32(word_ptr32 + 1) & kWordMask0[bytecount & 3];
michael@0:   word1 = word1 ^ (word1 << 18);
michael@0:   return word0 + word1;
michael@0: }
michael@0: 
michael@0: //
michael@0: // Ascii-7 One-byte chars
michael@0: //   ...ddddd...ccccc...bbbbb...aaaaa
michael@0: // make
michael@0: //   ...ddddd...ccccc...bbbbb...aaaaa
michael@0: //   000...ddddd...ccccc...bbbbb...aa
michael@0: //
michael@0: // Latin 1- and 2-byte chars
michael@0: //   ...ddddd...ccccc...bbbbb...aaaaa
michael@0: //   ...................fffff...eeeee
michael@0: // make
michael@0: //   ...ddddd...ccccc...bbbbb...aaaaa
michael@0: //   000...ddddd...ccccc...bbbbb...aa
michael@0: //   ...................fffff...eeeee
michael@0: //   ...............fffff...eeeee0000
michael@0: //
michael@0: // Non-CJK Two-byte chars
michael@0: //   ...ddddd...........bbbbb........
michael@0: //   ...hhhhh...........fffff........
michael@0: // make
michael@0: //   ...ddddd...........bbbbb........
michael@0: //   000...ddddd...........bbbbb.....
michael@0: //   ...hhhhh...........fffff........
michael@0: //   hhhh...........fffff........0000
michael@0: //
michael@0: // Non-CJK Three-byte chars
michael@0: //   ...........ccccc................
michael@0: //   ...................fffff........
michael@0: //   ...lllll...................iiiii
michael@0: // make
michael@0: //   ...........ccccc................
michael@0: //   000...........ccccc.............
michael@0: //   ...................fffff........
michael@0: //   ...............fffff........0000
michael@0: //   ...lllll...................iiiii
michael@0: //   .lllll...................iiiii00
michael@0: //
michael@0: 
michael@0: // QUADGRAM
michael@0: // Pick up 1..12 bytes plus pre/post space and hash them via mask/shift/add
michael@0: // OVERSHOOTS up to 3 bytes
michael@0: // For runtime use of tables
michael@0: // Does X86 unaligned loads
michael@0: uint32 QuadHashV2Mix(const char* word_ptr, int bytecount, uint32 prepost) {
michael@0:   const uint32* word_ptr32 = reinterpret_cast<const uint32*>(word_ptr);
michael@0:   uint32 word0, word1, word2;
michael@0:   if (bytecount <= 4) {
michael@0:     word0 = UNALIGNED_LOAD32(word_ptr32) & kWordMask0[bytecount & 3];
michael@0:     word0 = word0 ^ (word0 >> 3);
michael@0:     return word0 ^ prepost;
michael@0:   } else if (bytecount <= 8) {
michael@0:     word0 = UNALIGNED_LOAD32(word_ptr32);
michael@0:     word0 = word0 ^ (word0 >> 3);
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 1) & kWordMask0[bytecount & 3];
michael@0:     word1 = word1 ^ (word1 << 4);
michael@0:     return (word0 ^ prepost) + word1;
michael@0:   }
michael@0:   // else do 12 bytes
michael@0:   word0 = UNALIGNED_LOAD32(word_ptr32);
michael@0:   word0 = word0 ^ (word0 >> 3);
michael@0:   word1 = UNALIGNED_LOAD32(word_ptr32 + 1);
michael@0:   word1 = word1 ^ (word1 << 4);
michael@0:   word2 = UNALIGNED_LOAD32(word_ptr32 + 2) & kWordMask0[bytecount & 3];
michael@0:   word2 = word2 ^ (word2 << 2);
michael@0:   return (word0 ^ prepost) + word1 + word2;
michael@0: }
michael@0: 
michael@0: 
michael@0: // QUADGRAM wrapper with surrounding spaces
michael@0: // Pick up 1..12 bytes plus pre/post space and hash them via mask/shift/add
michael@0: // UNDERSHOOTS 1 byte, OVERSHOOTS up to 3 bytes
michael@0: // For runtime use of tables
michael@0: uint32 QuadHashV2(const char* word_ptr, int bytecount) {
michael@0:   if (bytecount == 0) {return 0;}
michael@0:   uint32 prepost = 0;
michael@0:   if (word_ptr[-1] == ' ') {prepost |= kPreSpaceIndicator;}
michael@0:   if (word_ptr[bytecount] == ' ') {prepost |= kPostSpaceIndicator;}
michael@0:   return QuadHashV2Mix(word_ptr, bytecount, prepost);
michael@0: }
michael@0: 
michael@0: // QUADGRAM wrapper with surrounding underscores (offline use)
michael@0: // Pick up 1..12 bytes plus pre/post '_' and hash them via mask/shift/add
michael@0: // OVERSHOOTS up to 3 bytes
michael@0: // For offline construction of tables
michael@0: uint32 QuadHashV2Underscore(const char* word_ptr, int bytecount) {
michael@0:   if (bytecount == 0) {return 0;}
michael@0:   const char* local_word_ptr = word_ptr;
michael@0:   int local_bytecount = bytecount;
michael@0:   uint32 prepost = 0;
michael@0:   if (local_word_ptr[0] == '_') {
michael@0:     prepost |= kPreSpaceIndicator;
michael@0:     ++local_word_ptr;
michael@0:     --local_bytecount;
michael@0:   }
michael@0:   if (local_word_ptr[local_bytecount - 1] == '_') {
michael@0:     prepost |= kPostSpaceIndicator;
michael@0:     --local_bytecount;
michael@0:   }
michael@0:   return QuadHashV2Mix(local_word_ptr, local_bytecount, prepost);
michael@0: }
michael@0: 
michael@0: 
michael@0: // OCTAGRAM
michael@0: // Pick up 1..24 bytes plus pre/post space and hash them via mask/shift/add
michael@0: // UNDERSHOOTS 1 byte, OVERSHOOTS up to 3 bytes
michael@0: //
michael@0: // The low 32 bits follow the pattern from above, tuned to different scripts
michael@0: // The high 8 bits are a simple sum of all bytes, shifted by 0/1/2/3 bits each
michael@0: // For runtime use of tables V3
michael@0: // Does X86 unaligned loads
michael@0: uint64 OctaHash40Mix(const char* word_ptr, int bytecount, uint64 prepost) {
michael@0:   const uint32* word_ptr32 = reinterpret_cast<const uint32*>(word_ptr);
michael@0:   uint64 word0;
michael@0:   uint64 word1;
michael@0:   uint64 sum;
michael@0: 
michael@0:   if (word_ptr[-1] == ' ') {prepost |= kPreSpaceIndicator;}
michael@0:   if (word_ptr[bytecount] == ' ') {prepost |= kPostSpaceIndicator;}
michael@0:   switch ((bytecount - 1) >> 2) {
michael@0:   case 0:       // 1..4 bytes
michael@0:     word0 = UNALIGNED_LOAD32(word_ptr32) & kWordMask0[bytecount & 3];
michael@0:     sum = word0;
michael@0:     word0 = word0 ^ (word0 >> 3);
michael@0:     break;
michael@0:   case 1:       // 5..8 bytes
michael@0:     word0 = UNALIGNED_LOAD32(word_ptr32);
michael@0:     sum = word0;
michael@0:     word0 = word0 ^ (word0 >> 3);
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 1) & kWordMask0[bytecount & 3];
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 << 4);
michael@0:     word0 += word1;
michael@0:     break;
michael@0:   case 2:       // 9..12 bytes
michael@0:     word0 = UNALIGNED_LOAD32(word_ptr32);
michael@0:     sum = word0;
michael@0:     word0 = word0 ^ (word0 >> 3);
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 1);
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 << 4);
michael@0:     word0 += word1;
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 2) & kWordMask0[bytecount & 3];
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 << 2);
michael@0:     word0 += word1;
michael@0:     break;
michael@0:   case 3:       // 13..16 bytes
michael@0:     word0 =UNALIGNED_LOAD32(word_ptr32);
michael@0:     sum = word0;
michael@0:     word0 = word0 ^ (word0 >> 3);
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 1);
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 << 4);
michael@0:     word0 += word1;
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 2);
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 << 2);
michael@0:     word0 += word1;
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 3) & kWordMask0[bytecount & 3];
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 >> 8);
michael@0:     word0 += word1;
michael@0:     break;
michael@0:   case 4:       // 17..20 bytes
michael@0:     word0 = UNALIGNED_LOAD32(word_ptr32);
michael@0:     sum = word0;
michael@0:     word0 = word0 ^ (word0 >> 3);
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 1);
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 << 4);
michael@0:     word0 += word1;
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 2);
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 << 2);
michael@0:     word0 += word1;
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 3);
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 >> 8);
michael@0:     word0 += word1;
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 4) & kWordMask0[bytecount & 3];
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 >> 4);
michael@0:     word0 += word1;
michael@0:     break;
michael@0:   default:      // 21..24 bytes and higher (ignores beyond 24)
michael@0:     word0 = UNALIGNED_LOAD32(word_ptr32);
michael@0:     sum = word0;
michael@0:     word0 = word0 ^ (word0 >> 3);
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 1);
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 << 4);
michael@0:     word0 += word1;
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 2);
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 << 2);
michael@0:     word0 += word1;
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 3);
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 >> 8);
michael@0:     word0 += word1;
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 4);
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 >> 4);
michael@0:     word0 += word1;
michael@0:     word1 = UNALIGNED_LOAD32(word_ptr32 + 5) & kWordMask0[bytecount & 3];
michael@0:     sum += word1;
michael@0:     word1 = word1 ^ (word1 >> 6);
michael@0:     word0 += word1;
michael@0:     break;
michael@0:   }
michael@0: 
michael@0:   sum += (sum >> 17);             // extra 1-bit shift for bytes 2 & 3
michael@0:   sum += (sum >> 9);              // extra 1-bit shift for bytes 1 & 3
michael@0:   sum = (sum & 0xff) << 32;
michael@0:   return (word0 ^ prepost) + sum;
michael@0: }
michael@0: 
michael@0: // OCTAGRAM wrapper with surrounding spaces
michael@0: // Pick up 1..24 bytes plus pre/post space and hash them via mask/shift/add
michael@0: // UNDERSHOOTS 1 byte, OVERSHOOTS up to 3 bytes
michael@0: //
michael@0: // The low 32 bits follow the pattern from above, tuned to different scripts
michael@0: // The high 8 bits are a simple sum of all bytes, shifted by 0/1/2/3 bits each
michael@0: // For runtime use of tables V3
michael@0: uint64 OctaHash40(const char* word_ptr, int bytecount) {
michael@0:   if (bytecount == 0) {return 0;}
michael@0:   uint64 prepost = 0;
michael@0:   if (word_ptr[-1] == ' ') {prepost |= kPreSpaceIndicator;}
michael@0:   if (word_ptr[bytecount] == ' ') {prepost |= kPostSpaceIndicator;}
michael@0:   return OctaHash40Mix(word_ptr, bytecount, prepost);
michael@0: }
michael@0: 
michael@0: 
michael@0: // OCTAGRAM wrapper with surrounding underscores (offline use)
michael@0: // Pick up 1..24 bytes plus pre/post space and hash them via mask/shift/add
michael@0: // UNDERSHOOTS 1 byte, OVERSHOOTS up to 3 bytes
michael@0: //
michael@0: // The low 32 bits follow the pattern from above, tuned to different scripts
michael@0: // The high 8 bits are a simple sum of all bytes, shifted by 0/1/2/3 bits each
michael@0: // For offline construction of tables
michael@0: uint64 OctaHash40underscore(const char* word_ptr, int bytecount) {
michael@0:   if (bytecount == 0) {return 0;}
michael@0:   const char* local_word_ptr = word_ptr;
michael@0:   int local_bytecount = bytecount;
michael@0:   uint64 prepost = 0;
michael@0:   if (local_word_ptr[0] == '_') {
michael@0:     prepost |= kPreSpaceIndicator;
michael@0:     ++local_word_ptr;
michael@0:     --local_bytecount;
michael@0:   }
michael@0:   if (local_word_ptr[local_bytecount - 1] == '_') {
michael@0:     prepost |= kPostSpaceIndicator;
michael@0:     --local_bytecount;
michael@0:   }
michael@0:   return OctaHash40Mix(local_word_ptr, local_bytecount, prepost);
michael@0: }
michael@0: 
michael@0: // Hash a consecutive pair of tokens/words A B
michael@0: // Old: hash is B - A, which gives too many false hits on one-char diffs
michael@0: // Now: rotate(A,13) + B
michael@0: uint64 PairHash(uint64 worda_hash, uint64 wordb_hash) {
michael@0:    return ((worda_hash >> 13) | (worda_hash << (64 - 13))) + wordb_hash;
michael@0: }
michael@0: 
michael@0: 
michael@0: 
michael@0: 
michael@0: //----------------------------------------------------------------------------//
michael@0: // Finding groups of 1/2/4/8 letters                                          //
michael@0: //----------------------------------------------------------------------------//
michael@0: 
michael@0: // src points to a letter. Find the byte length of a unigram starting there.
michael@0: int UniLen(const char* src) {
michael@0:   const char* src_end = src;
michael@0:   src_end += kAdvanceOneCharButSpace[(uint8)src_end[0]];
michael@0:   return src_end - src;
michael@0: }
michael@0: 
michael@0: // src points to a letter. Find the byte length of a bigram starting there.
michael@0: int BiLen(const char* src) {
michael@0:   const char* src_end = src;
michael@0:   src_end += kAdvanceOneCharButSpace[(uint8)src_end[0]];
michael@0:   src_end += kAdvanceOneCharButSpace[(uint8)src_end[0]];
michael@0:   return src_end - src;
michael@0: }
michael@0: 
michael@0: // src points to a letter. Find the byte length of a quadgram starting there.
michael@0: int QuadLen(const char* src) {
michael@0:   const char* src_end = src;
michael@0:   src_end += kAdvanceOneCharButSpace[(uint8)src_end[0]];
michael@0:   src_end += kAdvanceOneCharButSpace[(uint8)src_end[0]];
michael@0:   src_end += kAdvanceOneCharButSpace[(uint8)src_end[0]];
michael@0:   src_end += kAdvanceOneCharButSpace[(uint8)src_end[0]];
michael@0:   return src_end - src;
michael@0: }
michael@0: 
michael@0: // src points to a letter. Find the byte length of an octagram starting there.
michael@0: int OctaLen(const char* src) {
michael@0:   const char* src_end = src;
michael@0:   int charcount = 0;
michael@0:   while (src_end[0] != ' ') {
michael@0:     src_end += UTF8OneCharLen(src);
michael@0:     ++charcount;
michael@0:     if (charcount == 8) {break;}
michael@0:   }
michael@0:   return src_end - src;
michael@0: }
michael@0: 
michael@0: }       // End namespace CLD2
michael@0: 
michael@0: 
michael@0: 
michael@0: 
michael@0: