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: // Updated 2014.01 for dual table lookup
michael@0: //
michael@0: 
michael@0: #include "scoreonescriptspan.h"
michael@0: 
michael@0: #include "cldutil.h"
michael@0: #include "debug.h"
michael@0: #include "lang_script.h"
michael@0: 
michael@0: #include <stdio.h>
michael@0: 
michael@0: using namespace std;
michael@0: 
michael@0: namespace CLD2 {
michael@0: 
michael@0: static const int kUnreliablePercentThreshold = 75;
michael@0: 
michael@0: void AddLangProb(uint32 langprob, Tote* chunk_tote) {
michael@0:   ProcessProbV2Tote(langprob, chunk_tote);
michael@0: }
michael@0: 
michael@0: void ZeroPSLang(uint32 langprob, Tote* chunk_tote) {
michael@0:   uint8 top1 = (langprob >> 8) & 0xff;
michael@0:   chunk_tote->SetScore(top1, 0);
michael@0: }
michael@0: 
michael@0: bool SameCloseSet(uint16 lang1, uint16 lang2) {
michael@0:   int lang1_close_set = LanguageCloseSet(static_cast<Language>(lang1));
michael@0:   if (lang1_close_set == 0) {return false;}
michael@0:   int lang2_close_set = LanguageCloseSet(static_cast<Language>(lang2));
michael@0:   return (lang1_close_set == lang2_close_set);
michael@0: }
michael@0: 
michael@0: bool SameCloseSet(Language lang1, Language lang2) {
michael@0:   int lang1_close_set = LanguageCloseSet(lang1);
michael@0:   if (lang1_close_set == 0) {return false;}
michael@0:   int lang2_close_set = LanguageCloseSet(lang2);
michael@0:   return (lang1_close_set == lang2_close_set);
michael@0: }
michael@0: 
michael@0: 
michael@0: // Needs expected score per 1KB in scoring context
michael@0: void SetChunkSummary(ULScript ulscript, int first_linear_in_chunk,
michael@0:                      int offset, int len,
michael@0:                      const ScoringContext* scoringcontext,
michael@0:                      const Tote* chunk_tote,
michael@0:                      ChunkSummary* chunksummary) {
michael@0:   int key3[3];
michael@0:   chunk_tote->CurrentTopThreeKeys(key3);
michael@0:   Language lang1 = FromPerScriptNumber(ulscript, key3[0]);
michael@0:   Language lang2 = FromPerScriptNumber(ulscript, key3[1]);
michael@0: 
michael@0:   int actual_score_per_kb = 0;
michael@0:   if (len > 0) {
michael@0:     actual_score_per_kb = (chunk_tote->GetScore(key3[0]) << 10) / len;
michael@0:   }
michael@0:   int expected_subscr = lang1 * 4 + LScript4(ulscript);
michael@0:   int expected_score_per_kb =
michael@0:      scoringcontext->scoringtables->kExpectedScore[expected_subscr];
michael@0: 
michael@0:   chunksummary->offset = offset;
michael@0:   chunksummary->chunk_start = first_linear_in_chunk;
michael@0:   chunksummary->lang1 = lang1;
michael@0:   chunksummary->lang2 = lang2;
michael@0:   chunksummary->score1 = chunk_tote->GetScore(key3[0]);
michael@0:   chunksummary->score2 = chunk_tote->GetScore(key3[1]);
michael@0:   chunksummary->bytes = len;
michael@0:   chunksummary->grams = chunk_tote->GetScoreCount();
michael@0:   chunksummary->ulscript = ulscript;
michael@0:   chunksummary->reliability_delta = ReliabilityDelta(chunksummary->score1,
michael@0:                                                      chunksummary->score2,
michael@0:                                                      chunksummary->grams);
michael@0:   // If lang1/lang2 in same close set, set delta reliability to 100%
michael@0:   if (SameCloseSet(lang1, lang2)) {
michael@0:     chunksummary->reliability_delta = 100;
michael@0:   }
michael@0:   chunksummary->reliability_score =
michael@0:      ReliabilityExpected(actual_score_per_kb, expected_score_per_kb);
michael@0: }
michael@0: 
michael@0: // Return true if just lang1 is there: lang2=0 and lang3=0
michael@0: bool IsSingleLang(uint32 langprob) {
michael@0:   // Probably a bug -- which end is lang1? But only used to call empty Boost1
michael@0:   return ((langprob & 0x00ffff00) == 0);
michael@0: }
michael@0: 
michael@0: // Update scoring context distinct_boost for single language quad
michael@0: void AddDistinctBoost1(uint32 langprob, ScoringContext* scoringcontext) {
michael@0:   // Probably keep this empty -- not a good enough signal
michael@0: }
michael@0: 
michael@0: // Update scoring context distinct_boost for distinct octagram
michael@0: // Keep last 4 used. Since these are mostly (except at splices) in
michael@0: // hitbuffer, we might be able to just use a subscript and splice
michael@0: void AddDistinctBoost2(uint32 langprob, ScoringContext* scoringcontext) {
michael@0: // this is called 0..n times per chunk with decoded hitbuffer->distinct...
michael@0:   LangBoosts* distinct_boost = &scoringcontext->distinct_boost.latn;
michael@0:   if (scoringcontext->ulscript != ULScript_Latin) {
michael@0:     distinct_boost = &scoringcontext->distinct_boost.othr;
michael@0:   }
michael@0:   int n = distinct_boost->n;
michael@0:   distinct_boost->langprob[n] = langprob;
michael@0:   distinct_boost->n = distinct_boost->wrap(n + 1);
michael@0: }
michael@0: 
michael@0: // For each chunk, add extra weight for language priors (from content-lang and
michael@0: // meta lang=xx) and distinctive tokens
michael@0: void ScoreBoosts(const ScoringContext* scoringcontext, Tote* chunk_tote) {
michael@0:   // Get boosts for current script
michael@0:   const LangBoosts* langprior_boost = &scoringcontext->langprior_boost.latn;
michael@0:   const LangBoosts* langprior_whack = &scoringcontext->langprior_whack.latn;
michael@0:   const LangBoosts* distinct_boost = &scoringcontext->distinct_boost.latn;
michael@0:   if (scoringcontext->ulscript != ULScript_Latin) {
michael@0:     langprior_boost = &scoringcontext->langprior_boost.othr;
michael@0:     langprior_whack = &scoringcontext->langprior_whack.othr;
michael@0:     distinct_boost = &scoringcontext->distinct_boost.othr;
michael@0:   }
michael@0: 
michael@0:   for (int k = 0; k < kMaxBoosts; ++k) {
michael@0:     uint32 langprob = langprior_boost->langprob[k];
michael@0:     if (langprob > 0) {AddLangProb(langprob, chunk_tote);}
michael@0:   }
michael@0:   for (int k = 0; k < kMaxBoosts; ++k) {
michael@0:     uint32 langprob = distinct_boost->langprob[k];
michael@0:     if (langprob > 0) {AddLangProb(langprob, chunk_tote);}
michael@0:   }
michael@0:   // boost has a packed set of per-script langs and probabilites
michael@0:   // whack has a packed set of per-script lang to be suppressed (zeroed)
michael@0:   // When a language in a close set is given as an explicit hint, others in
michael@0:   //  that set will be whacked here.
michael@0:   for (int k = 0; k < kMaxBoosts; ++k) {
michael@0:     uint32 langprob = langprior_whack->langprob[k];
michael@0:     if (langprob > 0) {ZeroPSLang(langprob, chunk_tote);}
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: 
michael@0: // At this point, The chunk is described by
michael@0: //  hitbuffer->base[cspan->chunk_base .. cspan->chunk_base + cspan->base_len)
michael@0: //  hitbuffer->delta[cspan->chunk_delta ... )
michael@0: //  hitbuffer->distinct[cspan->chunk_distinct ... )
michael@0: // Scored text is in text[lo..hi) where
michael@0: //  lo is 0 or the min of first base/delta/distinct hitbuffer offset and
michael@0: //  hi is the min of next base/delta/distinct hitbuffer offset after
michael@0: //  base_len, etc.
michael@0: void GetTextSpanOffsets(const ScoringHitBuffer* hitbuffer,
michael@0:                         const ChunkSpan* cspan, int* lo, int* hi) {
michael@0:   // Front of this span
michael@0:   int lo_base = hitbuffer->base[cspan->chunk_base].offset;
michael@0:   int lo_delta = hitbuffer->delta[cspan->chunk_delta].offset;
michael@0:   int lo_distinct = hitbuffer->distinct[cspan->chunk_distinct].offset;
michael@0:   // Front of next span
michael@0:   int hi_base = hitbuffer->base[cspan->chunk_base +
michael@0:     cspan->base_len].offset;
michael@0:   int hi_delta = hitbuffer->delta[cspan->chunk_delta +
michael@0:     cspan->delta_len].offset;
michael@0:   int hi_distinct = hitbuffer->distinct[cspan->chunk_distinct +
michael@0:     cspan->distinct_len].offset;
michael@0: 
michael@0:   *lo = 0;
michael@0: //  if (cspan->chunk_base > 0) {
michael@0: //    *lo = minint(minint(lo_base, lo_delta), lo_distinct);
michael@0: //  }
michael@0:   *lo = minint(minint(lo_base, lo_delta), lo_distinct);
michael@0:   *hi = minint(minint(hi_base, hi_delta), hi_distinct);
michael@0: }
michael@0: 
michael@0: 
michael@0: int DiffScore(const CLD2TableSummary* obj, int indirect,
michael@0:               uint16 lang1, uint16 lang2) {
michael@0:   if (indirect < static_cast<int>(obj->kCLDTableSizeOne)) {
michael@0:     // Up to three languages at indirect
michael@0:     uint32 langprob = obj->kCLDTableInd[indirect];
michael@0:     return GetLangScore(langprob, lang1) - GetLangScore(langprob, lang2);
michael@0:   } else {
michael@0:     // Up to six languages at start + 2 * (indirect - start)
michael@0:     indirect += (indirect - obj->kCLDTableSizeOne);
michael@0:     uint32 langprob = obj->kCLDTableInd[indirect];
michael@0:     uint32 langprob2 = obj->kCLDTableInd[indirect + 1];
michael@0:     return (GetLangScore(langprob, lang1) + GetLangScore(langprob2, lang1)) -
michael@0:       (GetLangScore(langprob, lang2) + GetLangScore(langprob2, lang2));
michael@0:   }
michael@0: 
michael@0: }
michael@0: 
michael@0: // Score all the bases, deltas, distincts, boosts for one chunk into chunk_tote
michael@0: // After last chunk there is always a hitbuffer entry with an offset just off
michael@0: // the end of the text.
michael@0: // Sets delta_len, and distinct_len
michael@0: void ScoreOneChunk(const char* text, ULScript ulscript,
michael@0:                    const ScoringHitBuffer* hitbuffer,
michael@0:                    int chunk_i,
michael@0:                    ScoringContext* scoringcontext,
michael@0:                    ChunkSpan* cspan, Tote* chunk_tote,
michael@0:                    ChunkSummary* chunksummary) {
michael@0:   int first_linear_in_chunk = hitbuffer->chunk_start[chunk_i];
michael@0:   int first_linear_in_next_chunk = hitbuffer->chunk_start[chunk_i + 1];
michael@0: 
michael@0:   chunk_tote->Reinit();
michael@0:   cspan->delta_len = 0;
michael@0:   cspan->distinct_len = 0;
michael@0:   if (scoringcontext->flags_cld2_verbose) {
michael@0:     fprintf(scoringcontext->debug_file, "<br>ScoreOneChunk[%d..%d) ",
michael@0:             first_linear_in_chunk, first_linear_in_next_chunk);
michael@0:   }
michael@0: 
michael@0:   // 2013.02.05 linear design: just use base and base_len for the span
michael@0:   cspan->chunk_base = first_linear_in_chunk;
michael@0:   cspan->base_len = first_linear_in_next_chunk - first_linear_in_chunk;
michael@0:   for (int i = first_linear_in_chunk; i < first_linear_in_next_chunk; ++i) {
michael@0:     uint32 langprob = hitbuffer->linear[i].langprob;
michael@0:     AddLangProb(langprob, chunk_tote);
michael@0:     if (hitbuffer->linear[i].type <= QUADHIT) {
michael@0:       chunk_tote->AddScoreCount();      // Just count quads, not octas
michael@0:     }
michael@0:     if (hitbuffer->linear[i].type == DISTINCTHIT) {
michael@0:       AddDistinctBoost2(langprob, scoringcontext);
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   // Score language prior boosts
michael@0:   // Score distinct word boost
michael@0:   ScoreBoosts(scoringcontext, chunk_tote);
michael@0: 
michael@0:   int lo = hitbuffer->linear[first_linear_in_chunk].offset;
michael@0:   int hi = hitbuffer->linear[first_linear_in_next_chunk].offset;
michael@0: 
michael@0:   // Chunk_tote: get top langs, scores, etc. and fill in chunk summary
michael@0:   SetChunkSummary(ulscript, first_linear_in_chunk, lo, hi - lo,
michael@0:                   scoringcontext, chunk_tote, chunksummary);
michael@0: 
michael@0:   bool more_to_come = false;
michael@0:   bool score_cjk = false;
michael@0:   if (scoringcontext->flags_cld2_html) {
michael@0:     // Show one chunk in readable output
michael@0:     CLD2_Debug(text, lo, hi, more_to_come, score_cjk, hitbuffer,
michael@0:                scoringcontext, cspan, chunksummary);
michael@0:   }
michael@0: 
michael@0:   scoringcontext->prior_chunk_lang = static_cast<Language>(chunksummary->lang1);
michael@0: }
michael@0: 
michael@0: 
michael@0: // Score chunks of text described by hitbuffer, allowing each to be in a
michael@0: // different language, and optionally adjusting the boundaries inbetween.
michael@0: // Set last_cspan to the last chunkspan used
michael@0: void ScoreAllHits(const char* text,  ULScript ulscript,
michael@0:                   bool more_to_come, bool score_cjk,
michael@0:                   const ScoringHitBuffer* hitbuffer,
michael@0:                   ScoringContext* scoringcontext,
michael@0:                   SummaryBuffer* summarybuffer, ChunkSpan* last_cspan) {
michael@0:   ChunkSpan prior_cspan = {0, 0, 0, 0, 0, 0};
michael@0:   ChunkSpan cspan = {0, 0, 0, 0, 0, 0};
michael@0: 
michael@0:   for (int i = 0; i < hitbuffer->next_chunk_start; ++i) {
michael@0:     // Score one chunk
michael@0:     // Sets delta_len, and distinct_len
michael@0:     Tote chunk_tote;
michael@0:     ChunkSummary chunksummary;
michael@0:     ScoreOneChunk(text, ulscript,
michael@0:                   hitbuffer, i,
michael@0:                   scoringcontext, &cspan, &chunk_tote, &chunksummary);
michael@0: 
michael@0:     // Put result in summarybuffer
michael@0:     if (summarybuffer->n < kMaxSummaries) {
michael@0:       summarybuffer->chunksummary[summarybuffer->n] = chunksummary;
michael@0:       summarybuffer->n += 1;
michael@0:     }
michael@0: 
michael@0:     prior_cspan = cspan;
michael@0:     cspan.chunk_base += cspan.base_len;
michael@0:     cspan.chunk_delta += cspan.delta_len;
michael@0:     cspan.chunk_distinct += cspan.distinct_len;
michael@0:   }
michael@0: 
michael@0:   // Add one dummy off the end to hold first unused linear_in_chunk
michael@0:   int linear_off_end = hitbuffer->next_linear;
michael@0:   int offset_off_end = hitbuffer->linear[linear_off_end].offset;
michael@0:   ChunkSummary* cs = &summarybuffer->chunksummary[summarybuffer->n];
michael@0:   memset(cs, 0, sizeof(ChunkSummary));
michael@0:   cs->offset = offset_off_end;
michael@0:   cs->chunk_start = linear_off_end;
michael@0:   *last_cspan = prior_cspan;
michael@0: }
michael@0: 
michael@0: 
michael@0: void SummaryBufferToDocTote(const SummaryBuffer* summarybuffer,
michael@0:                             bool more_to_come, DocTote* doc_tote) {
michael@0:   int cs_bytes_sum = 0;
michael@0:   for (int i = 0; i < summarybuffer->n; ++i) {
michael@0:     const ChunkSummary* cs = &summarybuffer->chunksummary[i];
michael@0:     int reliability = minint(cs->reliability_delta, cs->reliability_score);
michael@0:     // doc_tote uses full languages
michael@0:     doc_tote->Add(cs->lang1, cs->bytes, cs->score1, reliability);
michael@0:     cs_bytes_sum += cs->bytes;
michael@0:   }
michael@0: }
michael@0: 
michael@0: // Turn on for debugging vectors
michael@0: static const bool kShowLettersOriginal = false;
michael@0: 
michael@0: 
michael@0: // If next chunk language matches last vector language, extend last element
michael@0: // Otherwise add new element to vector
michael@0: void ItemToVector(ScriptScanner* scanner,
michael@0:                   ResultChunkVector* vec, Language new_lang,
michael@0:                   int mapped_offset, int mapped_len) {
michael@0:   uint16 last_vec_lang = static_cast<uint16>(UNKNOWN_LANGUAGE);
michael@0:   int last_vec_subscr = vec->size() - 1;
michael@0:   if (last_vec_subscr >= 0) {
michael@0:     ResultChunk* priorrc = &(*vec)[last_vec_subscr];
michael@0:     last_vec_lang = priorrc->lang1;
michael@0:     if (new_lang == last_vec_lang) {
michael@0:       // Extend prior. Current mapped_offset may be beyond prior end, so do
michael@0:       // the arithmetic to include any such gap
michael@0:       priorrc->bytes = minint((mapped_offset + mapped_len) - priorrc->offset,
michael@0:                               kMaxResultChunkBytes);
michael@0:       if (kShowLettersOriginal) {
michael@0:         // Optionally print the new chunk original text
michael@0:         string temp2(&scanner->GetBufferStart()[priorrc->offset],
michael@0:                      priorrc->bytes);
michael@0:         fprintf(stderr, "Item[%d..%d) '%s'<br>\n",
michael@0:                 priorrc->offset, priorrc->offset + priorrc->bytes,
michael@0:                 GetHtmlEscapedText(temp2).c_str());
michael@0:       }
michael@0:       return;
michael@0:     }
michael@0:   }
michael@0:   // Add new vector element
michael@0:   ResultChunk rc;
michael@0:   rc.offset = mapped_offset;
michael@0:   rc.bytes = minint(mapped_len, kMaxResultChunkBytes);
michael@0:   rc.lang1 = static_cast<uint16>(new_lang);
michael@0:   vec->push_back(rc);
michael@0:   if (kShowLettersOriginal) {
michael@0:     // Optionally print the new chunk original text
michael@0:     string temp2(&scanner->GetBufferStart()[rc.offset], rc.bytes);
michael@0:     fprintf(stderr, "Item[%d..%d) '%s'<br>\n",
michael@0:             rc.offset, rc.offset + rc.bytes,
michael@0:             GetHtmlEscapedText(temp2).c_str());
michael@0:   }
michael@0: }
michael@0: 
michael@0: uint16 PriorVecLang(const ResultChunkVector* vec) {
michael@0:   if (vec->empty()) {return static_cast<uint16>(UNKNOWN_LANGUAGE);}
michael@0:   return (*vec)[vec->size() - 1].lang1;
michael@0: }
michael@0: 
michael@0: uint16 NextChunkLang(const SummaryBuffer* summarybuffer, int i) {
michael@0:   if ((i + 1) >= summarybuffer->n) {
michael@0:     return static_cast<uint16>(UNKNOWN_LANGUAGE);
michael@0:   }
michael@0:   return summarybuffer->chunksummary[i + 1].lang1;
michael@0: }
michael@0: 
michael@0: 
michael@0: 
michael@0: // Add n elements of summarybuffer to resultchunk vector:
michael@0: // Each element is letters-only text [offset..offset+bytes)
michael@0: // This maps back to original[Back(offset)..Back(offset+bytes))
michael@0: //
michael@0: // We go out of our way to minimize the variation in the ResultChunkVector,
michael@0: // so that the caller has fewer but more meaningful spans in different
michael@0: // lanaguges, for the likely purpose of translation or spell-check.
michael@0: //
michael@0: // The language of each chunk is lang1, but it might be unreliable for
michael@0: // either of two reasons: its score is relatively too close to the score of
michael@0: // lang2, or its score is too far away from the expected score of real text in
michael@0: // the given language. Unreliable languages are mapped to Unknown.
michael@0: //
michael@0: void SummaryBufferToVector(ScriptScanner* scanner, const char* text,
michael@0:                            const SummaryBuffer* summarybuffer,
michael@0:                            bool more_to_come, ResultChunkVector* vec) {
michael@0:   if (vec == NULL) {return;}
michael@0: 
michael@0:   if (kShowLettersOriginal) {
michael@0:     fprintf(stderr, "map2original_ ");
michael@0:     scanner->map2original_.DumpWindow();
michael@0:     fprintf(stderr, "<br>\n");
michael@0:     fprintf(stderr, "map2uplow_ ");
michael@0:     scanner->map2uplow_.DumpWindow();
michael@0:     fprintf(stderr, "<br>\n");
michael@0:   }
michael@0: 
michael@0:   for (int i = 0; i < summarybuffer->n; ++i) {
michael@0:     const ChunkSummary* cs = &summarybuffer->chunksummary[i];
michael@0:     int unmapped_offset = cs->offset;
michael@0:     int unmapped_len = cs->bytes;
michael@0: 
michael@0:     if (kShowLettersOriginal) {
michael@0:       // Optionally print the chunk lowercase letters/marks text
michael@0:       string temp(&text[unmapped_offset], unmapped_len);
michael@0:       fprintf(stderr, "Letters [%d..%d) '%s'<br>\n",
michael@0:               unmapped_offset, unmapped_offset + unmapped_len,
michael@0:               GetHtmlEscapedText(temp).c_str());
michael@0:     }
michael@0: 
michael@0:     int mapped_offset = scanner->MapBack(unmapped_offset);
michael@0: 
michael@0:     // Trim back a little to prefer splicing original at word boundaries
michael@0:     if (mapped_offset > 0) {
michael@0:       // Size of prior vector entry, if any
michael@0:       int prior_size = 0;
michael@0:       if (!vec->empty()) {
michael@0:         ResultChunk* rc = &(*vec)[vec->size() - 1];
michael@0:         prior_size = rc->bytes;
michael@0:       }
michael@0:       // Maximum back up size to leave at least 3 bytes in prior,
michael@0:       // and not entire buffer, and no more than 12 bytes total backup
michael@0:       int n_limit = minint(prior_size - 3, mapped_offset);
michael@0:       n_limit = minint(n_limit, 12);
michael@0: 
michael@0:       // Backscan over letters, stopping if prior byte is < 0x41
michael@0:       // There is some possibility that we will backscan over a different script
michael@0:       const char* s = &scanner->GetBufferStart()[mapped_offset];
michael@0:       const unsigned char* us = reinterpret_cast<const unsigned char*>(s);
michael@0:       int n = 0;
michael@0:       while ((n < n_limit) && (us[-n - 1] >= 0x41)) {++n;}
michael@0:       if (n >= n_limit) {n = 0;} // New boundary not found within range
michael@0: 
michael@0:       // Also back up exactly one leading punctuation character if '"#@
michael@0:       if (n < n_limit) {
michael@0:         unsigned char c = us[-n - 1];
michael@0:         if ((c == '\'') || (c == '"') || (c == '#') || (c == '@')) {++n;}
michael@0:       }
michael@0:       // Shrink the previous chunk slightly
michael@0:       if (n > 0) {
michael@0:         ResultChunk* rc = &(*vec)[vec->size() - 1];
michael@0:         rc->bytes -= n;
michael@0:         mapped_offset -= n;
michael@0:         if (kShowLettersOriginal) {
michael@0:           fprintf(stderr, "Back up %d bytes<br>\n", n);
michael@0:           // Optionally print the prior chunk original text
michael@0:           string temp2(&scanner->GetBufferStart()[rc->offset], rc->bytes);
michael@0:           fprintf(stderr, "Prior   [%d..%d) '%s'<br>\n",
michael@0:                   rc->offset, rc->offset + rc->bytes,
michael@0:                   GetHtmlEscapedText(temp2).c_str());
michael@0:         }
michael@0:       }
michael@0:     }
michael@0: 
michael@0:     int mapped_len =
michael@0:       scanner->MapBack(unmapped_offset + unmapped_len) - mapped_offset;
michael@0: 
michael@0:     if (kShowLettersOriginal) {
michael@0:       // Optionally print the chunk original text
michael@0:       string temp2(&scanner->GetBufferStart()[mapped_offset], mapped_len);
michael@0:       fprintf(stderr, "Original[%d..%d) '%s'<br>\n",
michael@0:               mapped_offset, mapped_offset + mapped_len,
michael@0:               GetHtmlEscapedText(temp2).c_str());
michael@0:     }
michael@0: 
michael@0:     Language new_lang = static_cast<Language>(cs->lang1);
michael@0:     bool reliability_delta_bad =
michael@0:       (cs->reliability_delta < kUnreliablePercentThreshold);
michael@0:     bool reliability_score_bad =
michael@0:       (cs->reliability_score < kUnreliablePercentThreshold);
michael@0: 
michael@0:     // If the top language matches last vector, ignore reliability_delta
michael@0:     uint16 prior_lang = PriorVecLang(vec);
michael@0:     if (prior_lang == cs->lang1) {
michael@0:       reliability_delta_bad = false;
michael@0:     }
michael@0:     // If the top language is in same close set as last vector, set up to merge
michael@0:     if (SameCloseSet(cs->lang1, prior_lang)) {
michael@0:       new_lang = static_cast<Language>(prior_lang);
michael@0:       reliability_delta_bad = false;
michael@0:     }
michael@0:     // If the top two languages are in the same close set and the last vector
michael@0:     // language is the second language, set up to merge
michael@0:     if (SameCloseSet(cs->lang1, cs->lang2) &&
michael@0:         (prior_lang == cs->lang2)) {
michael@0:       new_lang = static_cast<Language>(prior_lang);
michael@0:       reliability_delta_bad = false;
michael@0:     }
michael@0:     // If unreliable and the last and next vector languages are both
michael@0:     // the second language, set up to merge
michael@0:     uint16 next_lang = NextChunkLang(summarybuffer, i);
michael@0:     if (reliability_delta_bad &&
michael@0:         (prior_lang == cs->lang2) && (next_lang == cs->lang2)) {
michael@0:       new_lang = static_cast<Language>(prior_lang);
michael@0:       reliability_delta_bad = false;
michael@0:     }
michael@0: 
michael@0:     if (reliability_delta_bad || reliability_score_bad) {
michael@0:       new_lang = UNKNOWN_LANGUAGE;
michael@0:     }
michael@0:     ItemToVector(scanner, vec, new_lang, mapped_offset, mapped_len);
michael@0:   }
michael@0: }
michael@0: 
michael@0: // Add just one element to resultchunk vector:
michael@0: // For RTypeNone or RTypeOne
michael@0: void JustOneItemToVector(ScriptScanner* scanner, const char* text,
michael@0:                          Language lang1, int unmapped_offset, int unmapped_len,
michael@0:                          ResultChunkVector* vec) {
michael@0:   if (vec == NULL) {return;}
michael@0: 
michael@0:   if (kShowLettersOriginal) {
michael@0:     fprintf(stderr, "map2original_ ");
michael@0:     scanner->map2original_.DumpWindow();
michael@0:     fprintf(stderr, "<br>\n");
michael@0:     fprintf(stderr, "map2uplow_ ");
michael@0:     scanner->map2uplow_.DumpWindow();
michael@0:     fprintf(stderr, "<br>\n");
michael@0:   }
michael@0: 
michael@0:   if (kShowLettersOriginal) {
michael@0:    // Optionally print the chunk lowercase letters/marks text
michael@0:    string temp(&text[unmapped_offset], unmapped_len);
michael@0:    fprintf(stderr, "Letters1 [%d..%d) '%s'<br>\n",
michael@0:            unmapped_offset, unmapped_offset + unmapped_len,
michael@0:            GetHtmlEscapedText(temp).c_str());
michael@0:   }
michael@0: 
michael@0:   int mapped_offset = scanner->MapBack(unmapped_offset);
michael@0:   int mapped_len =
michael@0:     scanner->MapBack(unmapped_offset + unmapped_len) - mapped_offset;
michael@0: 
michael@0:   if (kShowLettersOriginal) {
michael@0:     // Optionally print the chunk original text
michael@0:     string temp2(&scanner->GetBufferStart()[mapped_offset], mapped_len);
michael@0:     fprintf(stderr, "Original1[%d..%d) '%s'<br>\n",
michael@0:             mapped_offset, mapped_offset + mapped_len,
michael@0:             GetHtmlEscapedText(temp2).c_str());
michael@0:   }
michael@0: 
michael@0:   ItemToVector(scanner, vec, lang1, mapped_offset, mapped_len);
michael@0: }
michael@0: 
michael@0: 
michael@0: // Debugging. Not thread safe. Defined in getonescriptspan
michael@0: char* DisplayPiece(const char* next_byte_, int byte_length_);
michael@0: 
michael@0: // If high bit is on, take out high bit and add 2B to make table2 entries easy
michael@0: inline int PrintableIndirect(int x) {
michael@0:   if ((x & 0x80000000u) != 0) {
michael@0:     return (x & ~0x80000000u) + 2000000000;
michael@0:   }
michael@0:   return x;
michael@0: }
michael@0: void DumpHitBuffer(FILE* df, const char* text,
michael@0:                    const ScoringHitBuffer* hitbuffer) {
michael@0:   fprintf(df,
michael@0:           "<br>DumpHitBuffer[%s, next_base/delta/distinct %d, %d, %d)<br>\n",
michael@0:           ULScriptCode(hitbuffer->ulscript),
michael@0:           hitbuffer->next_base, hitbuffer->next_delta,
michael@0:           hitbuffer->next_distinct);
michael@0:   for (int i = 0; i < hitbuffer->maxscoringhits; ++i) {
michael@0:     if (i < hitbuffer->next_base) {
michael@0:       fprintf(df, "Q[%d]%d,%d,%s ",
michael@0:               i, hitbuffer->base[i].offset,
michael@0:               PrintableIndirect(hitbuffer->base[i].indirect),
michael@0:               DisplayPiece(&text[hitbuffer->base[i].offset], 6));
michael@0:     }
michael@0:     if (i < hitbuffer->next_delta) {
michael@0:       fprintf(df, "DL[%d]%d,%d,%s ",
michael@0:               i, hitbuffer->delta[i].offset, hitbuffer->delta[i].indirect,
michael@0:               DisplayPiece(&text[hitbuffer->delta[i].offset], 12));
michael@0:     }
michael@0:     if (i < hitbuffer->next_distinct) {
michael@0:       fprintf(df, "D[%d]%d,%d,%s ",
michael@0:               i, hitbuffer->distinct[i].offset, hitbuffer->distinct[i].indirect,
michael@0:               DisplayPiece(&text[hitbuffer->distinct[i].offset], 12));
michael@0:     }
michael@0:     if (i < hitbuffer->next_base) {
michael@0:       fprintf(df, "<br>\n");
michael@0:     }
michael@0:     if (i > 50) {break;}
michael@0:   }
michael@0:   if (hitbuffer->next_base > 50) {
michael@0:     int i = hitbuffer->next_base;
michael@0:     fprintf(df, "Q[%d]%d,%d,%s ",
michael@0:             i, hitbuffer->base[i].offset,
michael@0:             PrintableIndirect(hitbuffer->base[i].indirect),
michael@0:             DisplayPiece(&text[hitbuffer->base[i].offset], 6));
michael@0:   }
michael@0:   if (hitbuffer->next_delta > 50) {
michael@0:     int i = hitbuffer->next_delta;
michael@0:     fprintf(df, "DL[%d]%d,%d,%s ",
michael@0:             i, hitbuffer->delta[i].offset, hitbuffer->delta[i].indirect,
michael@0:             DisplayPiece(&text[hitbuffer->delta[i].offset], 12));
michael@0:   }
michael@0:   if (hitbuffer->next_distinct > 50) {
michael@0:     int i = hitbuffer->next_distinct;
michael@0:     fprintf(df, "D[%d]%d,%d,%s ",
michael@0:             i, hitbuffer->distinct[i].offset, hitbuffer->distinct[i].indirect,
michael@0:             DisplayPiece(&text[hitbuffer->distinct[i].offset], 12));
michael@0:   }
michael@0:   fprintf(df, "<br>\n");
michael@0: }
michael@0: 
michael@0: 
michael@0: void DumpLinearBuffer(FILE* df, const char* text,
michael@0:                       const ScoringHitBuffer* hitbuffer) {
michael@0:   fprintf(df, "<br>DumpLinearBuffer[%d)<br>\n",
michael@0:           hitbuffer->next_linear);
michael@0:   // Include the dummy entry off the end
michael@0:   for (int i = 0; i < hitbuffer->next_linear + 1; ++i) {
michael@0:     if ((50 < i) && (i < (hitbuffer->next_linear - 1))) {continue;}
michael@0:     fprintf(df, "[%d]%d,%c=%08x,%s<br>\n",
michael@0:             i, hitbuffer->linear[i].offset,
michael@0:             "UQLD"[hitbuffer->linear[i].type],
michael@0:             hitbuffer->linear[i].langprob,
michael@0:             DisplayPiece(&text[hitbuffer->linear[i].offset], 6));
michael@0:   }
michael@0:   fprintf(df, "<br>\n");
michael@0: 
michael@0:   fprintf(df, "DumpChunkStart[%d]<br>\n", hitbuffer->next_chunk_start);
michael@0:   for (int i = 0; i < hitbuffer->next_chunk_start + 1; ++i) {
michael@0:     fprintf(df, "[%d]%d\n", i, hitbuffer->chunk_start[i]);
michael@0:   }
michael@0:   fprintf(df, "<br>\n");
michael@0: }
michael@0: 
michael@0: // Move this verbose debugging output to debug.cc eventually
michael@0: void DumpChunkSummary(FILE* df, const ChunkSummary* cs) {
michael@0:   // Print chunksummary
michael@0:   fprintf(df, "%d lin[%d] %s.%d %s.%d %dB %d# %s %dRd %dRs<br>\n",
michael@0:           cs->offset,
michael@0:           cs->chunk_start,
michael@0:           LanguageCode(static_cast<Language>(cs->lang1)),
michael@0:           cs->score1,
michael@0:           LanguageCode(static_cast<Language>(cs->lang2)),
michael@0:           cs->score2,
michael@0:           cs->bytes,
michael@0:           cs->grams,
michael@0:           ULScriptCode(static_cast<ULScript>(cs->ulscript)),
michael@0:           cs->reliability_delta,
michael@0:           cs->reliability_score);
michael@0: }
michael@0: 
michael@0: void DumpSummaryBuffer(FILE* df, const SummaryBuffer* summarybuffer) {
michael@0:   fprintf(df, "<br>DumpSummaryBuffer[%d]<br>\n", summarybuffer->n);
michael@0:   fprintf(df, "[i] offset linear[chunk_start] lang.score1 lang.score2 "
michael@0:               "bytesB ngrams# script rel_delta rel_score<br>\n");
michael@0:   for (int i = 0; i <= summarybuffer->n; ++i) {
michael@0:     fprintf(df, "[%d] ", i);
michael@0:     DumpChunkSummary(df, &summarybuffer->chunksummary[i]);
michael@0:   }
michael@0:   fprintf(df, "<br>\n");
michael@0: }
michael@0: 
michael@0: 
michael@0: 
michael@0: // Within hitbufer->linear[]
michael@0: // <-- prior chunk --><-- this chunk -->
michael@0: // |                  |                 |
michael@0: // linear0            linear1           linear2
michael@0: //     lang0              lang1
michael@0: // The goal of sharpening is to move this_linear to better separate langs
michael@0: int BetterBoundary(const char* text,
michael@0:                    ScoringHitBuffer* hitbuffer,
michael@0:                    ScoringContext* scoringcontext,
michael@0:                    uint16 pslang0, uint16 pslang1,
michael@0:                    int linear0, int linear1, int linear2) {
michael@0:   // Degenerate case, no change
michael@0:   if ((linear2 - linear0) <= 8) {return linear1;}
michael@0: 
michael@0:   // Each diff gives pslang0 score - pslang1 score
michael@0:   // Running diff has four entries + + + + followed by four entries - - - -
michael@0:   // so that this value is maximal at the sharpest boundary between pslang0
michael@0:   // (positive diffs) and pslang1 (negative diffs)
michael@0:   int running_diff = 0;
michael@0:   int diff[8];    // Ring buffer of pslang0-pslang1 differences
michael@0:   // Initialize with first 8 diffs
michael@0:   for (int i = linear0; i < linear0 + 8; ++i) {
michael@0:     int j = i & 7;
michael@0:     uint32 langprob = hitbuffer->linear[i].langprob;
michael@0:     diff[j] = GetLangScore(langprob, pslang0) -
michael@0:        GetLangScore(langprob, pslang1);
michael@0:     if (i < linear0 + 4) {
michael@0:       // First four diffs pslang0 - pslang1
michael@0:       running_diff += diff[j];
michael@0:     } else {
michael@0:       // Second four diffs -(pslang0 - pslang1)
michael@0:       running_diff -= diff[j];
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   // Now scan for sharpest boundary. j is at left end of 8 entries
michael@0:   // To be a boundary, there must be both >0 and <0 entries in the window
michael@0:   int better_boundary_value = 0;
michael@0:   int better_boundary = linear1;
michael@0:   for (int i = linear0; i < linear2 - 8; ++i) {
michael@0:     int j = i & 7;
michael@0:     if (better_boundary_value < running_diff) {
michael@0:       bool has_plus = false;
michael@0:       bool has_minus = false;
michael@0:       for (int kk = 0; kk < 8; ++kk) {
michael@0:         if (diff[kk] > 0) {has_plus = true;}
michael@0:         if (diff[kk] < 0) {has_minus = true;}
michael@0:       }
michael@0:       if (has_plus && has_minus) {
michael@0:         better_boundary_value = running_diff;
michael@0:         better_boundary = i + 4;
michael@0:       }
michael@0:     }
michael@0:     // Shift right one entry
michael@0:     uint32 langprob = hitbuffer->linear[i + 8].langprob;
michael@0:     int newdiff = GetLangScore(langprob, pslang0) -
michael@0:        GetLangScore(langprob, pslang1);
michael@0:     int middiff = diff[(i + 4) & 7];
michael@0:     int olddiff = diff[j];
michael@0:     diff[j] = newdiff;
michael@0:     running_diff -= olddiff;                  // Remove left
michael@0:     running_diff += 2 * middiff;              // Convert middle from - to +
michael@0:     running_diff -= newdiff;                  // Insert right
michael@0:   }
michael@0: 
michael@0:   if (scoringcontext->flags_cld2_verbose && (linear1 != better_boundary)) {
michael@0:     Language lang0 = FromPerScriptNumber(scoringcontext->ulscript, pslang0);
michael@0:     Language lang1 = FromPerScriptNumber(scoringcontext->ulscript, pslang1);
michael@0:     fprintf(scoringcontext->debug_file, " Better lin[%d=>%d] %s^^%s <br>\n",
michael@0:             linear1, better_boundary,
michael@0:             LanguageCode(lang0), LanguageCode(lang1));
michael@0:     int lin0_off = hitbuffer->linear[linear0].offset;
michael@0:     int lin1_off = hitbuffer->linear[linear1].offset;
michael@0:     int lin2_off = hitbuffer->linear[linear2].offset;
michael@0:     int better_offm1 = hitbuffer->linear[better_boundary - 1].offset;
michael@0:     int better_off = hitbuffer->linear[better_boundary].offset;
michael@0:     int better_offp1 = hitbuffer->linear[better_boundary + 1].offset;
michael@0:     string old0(&text[lin0_off], lin1_off - lin0_off);
michael@0:     string old1(&text[lin1_off], lin2_off - lin1_off);
michael@0:     string new0(&text[lin0_off], better_offm1 - lin0_off);
michael@0:     string new0m1(&text[better_offm1], better_off - better_offm1);
michael@0:     string new1(&text[better_off], better_offp1 - better_off);
michael@0:     string new1p1(&text[better_offp1], lin2_off - better_offp1);
michael@0:     fprintf(scoringcontext->debug_file, "%s^^%s => <br>\n%s^%s^^%s^%s<br>\n",
michael@0:             GetHtmlEscapedText(old0).c_str(),
michael@0:             GetHtmlEscapedText(old1).c_str(),
michael@0:             GetHtmlEscapedText(new0).c_str(),
michael@0:             GetHtmlEscapedText(new0m1).c_str(),
michael@0:             GetHtmlEscapedText(new1).c_str(),
michael@0:             GetHtmlEscapedText(new1p1).c_str());
michael@0:     // Slow picture of differences per linear entry
michael@0:     int d;
michael@0:     for (int i = linear0; i < linear2; ++i) {
michael@0:       if (i == better_boundary) {
michael@0:         fprintf(scoringcontext->debug_file, "^^ ");
michael@0:       }
michael@0:       uint32 langprob = hitbuffer->linear[i].langprob;
michael@0:       d = GetLangScore(langprob, pslang0) - GetLangScore(langprob, pslang1);
michael@0:       const char* s = "=";
michael@0:       //if (d > 2) {s = "\xc2\xaf";}    // Macron
michael@0:       if (d > 2) {s = "#";}
michael@0:       else if (d > 0) {s = "+";}
michael@0:       else if (d < -2) {s = "_";}
michael@0:       else if (d < 0) {s = "-";}
michael@0:       fprintf(scoringcontext->debug_file, "%s ", s);
michael@0:     }
michael@0:     fprintf(scoringcontext->debug_file, " &nbsp;&nbsp;(scale: #+=-_)<br>\n");
michael@0:   }
michael@0:   return better_boundary;
michael@0: }
michael@0: 
michael@0: 
michael@0: // For all but the first summary, if its top language differs from
michael@0: // the previous chunk, refine the boundary
michael@0: // Linearized version
michael@0: void SharpenBoundaries(const char* text,
michael@0:                        bool more_to_come,
michael@0:                        ScoringHitBuffer* hitbuffer,
michael@0:                        ScoringContext* scoringcontext,
michael@0:                        SummaryBuffer* summarybuffer) {
michael@0: 
michael@0:   int prior_linear = summarybuffer->chunksummary[0].chunk_start;
michael@0:   uint16 prior_lang = summarybuffer->chunksummary[0].lang1;
michael@0: 
michael@0:   if (scoringcontext->flags_cld2_verbose) {
michael@0:     fprintf(scoringcontext->debug_file, "<br>SharpenBoundaries<br>\n");
michael@0:   }
michael@0:   for (int i = 1; i < summarybuffer->n; ++i) {
michael@0:     ChunkSummary* cs = &summarybuffer->chunksummary[i];
michael@0:     uint16 this_lang = cs->lang1;
michael@0:     if (this_lang == prior_lang) {
michael@0:       prior_linear = cs->chunk_start;
michael@0:       continue;
michael@0:     }
michael@0: 
michael@0:     int this_linear = cs->chunk_start;
michael@0:     int next_linear = summarybuffer->chunksummary[i + 1].chunk_start;
michael@0: 
michael@0:     // If this/prior in same close set, don't move boundary
michael@0:     if (SameCloseSet(prior_lang, this_lang)) {
michael@0:       prior_linear = this_linear;
michael@0:       prior_lang = this_lang;
michael@0:       continue;
michael@0:     }
michael@0: 
michael@0: 
michael@0:     // Within hitbuffer->linear[]
michael@0:     // <-- prior chunk --><-- this chunk -->
michael@0:     // |                  |                 |
michael@0:     // prior_linear       this_linear       next_linear
michael@0:     //     prior_lang         this_lang
michael@0:     // The goal of sharpening is to move this_linear to better separate langs
michael@0: 
michael@0:     uint8 pslang0 = PerScriptNumber(scoringcontext->ulscript,
michael@0:                                     static_cast<Language>(prior_lang));
michael@0:     uint8 pslang1 = PerScriptNumber(scoringcontext->ulscript,
michael@0:                                     static_cast<Language>(this_lang));
michael@0:     int better_linear = BetterBoundary(text,
michael@0:                                        hitbuffer,
michael@0:                                        scoringcontext,
michael@0:                                        pslang0, pslang1,
michael@0:                                        prior_linear, this_linear, next_linear);
michael@0: 
michael@0:     int old_offset = hitbuffer->linear[this_linear].offset;
michael@0:     int new_offset = hitbuffer->linear[better_linear].offset;
michael@0:     cs->chunk_start = better_linear;
michael@0:     cs->offset = new_offset;
michael@0:     // If this_linear moved right, make bytes smaller for this, larger for prior
michael@0:     // If this_linear moved left, make bytes larger for this, smaller for prior
michael@0:     cs->bytes -= (new_offset - old_offset);
michael@0:     summarybuffer->chunksummary[i - 1].bytes += (new_offset - old_offset);
michael@0: 
michael@0:     this_linear = better_linear;    // Update so that next chunk doesn't intrude
michael@0: 
michael@0:     // Consider rescoring the two chunks
michael@0: 
michael@0:     // Update for next round (note: using pre-updated boundary)
michael@0:     prior_linear = this_linear;
michael@0:     prior_lang = this_lang;
michael@0:   }
michael@0: }
michael@0: 
michael@0: // Make a langprob that gives small weight to the default language for ulscript
michael@0: uint32 DefaultLangProb(ULScript ulscript) {
michael@0:   Language default_lang = DefaultLanguage(ulscript);
michael@0:   return MakeLangProb(default_lang, 1);
michael@0: }
michael@0: 
michael@0: // Effectively, do a merge-sort based on text offsets
michael@0: // Look up each indirect value in appropriate scoring table and keep
michael@0: // just the resulting langprobs
michael@0: void LinearizeAll(ScoringContext* scoringcontext, bool score_cjk,
michael@0:                   ScoringHitBuffer* hitbuffer) {
michael@0:   const CLD2TableSummary* base_obj;       // unigram or quadgram
michael@0:   const CLD2TableSummary* base_obj2;      // quadgram dual table
michael@0:   const CLD2TableSummary* delta_obj;      // bigram or octagram
michael@0:   const CLD2TableSummary* distinct_obj;   // bigram or octagram
michael@0:   uint16 base_hit;
michael@0:   if (score_cjk) {
michael@0:     base_obj = scoringcontext->scoringtables->unigram_compat_obj;
michael@0:     base_obj2 = scoringcontext->scoringtables->unigram_compat_obj;
michael@0:     delta_obj = scoringcontext->scoringtables->deltabi_obj;
michael@0:     distinct_obj = scoringcontext->scoringtables->distinctbi_obj;
michael@0:     base_hit = UNIHIT;
michael@0:   } else {
michael@0:     base_obj = scoringcontext->scoringtables->quadgram_obj;
michael@0:     base_obj2 = scoringcontext->scoringtables->quadgram_obj2;
michael@0:     delta_obj = scoringcontext->scoringtables->deltaocta_obj;
michael@0:     distinct_obj = scoringcontext->scoringtables->distinctocta_obj;
michael@0:     base_hit = QUADHIT;
michael@0:   }
michael@0: 
michael@0:   int base_limit = hitbuffer->next_base;
michael@0:   int delta_limit = hitbuffer->next_delta;
michael@0:   int distinct_limit = hitbuffer->next_distinct;
michael@0:   int base_i = 0;
michael@0:   int delta_i = 0;
michael@0:   int distinct_i = 0;
michael@0:   int linear_i = 0;
michael@0: 
michael@0:   // Start with an initial base hit for the default language for this script
michael@0:   // Inserting this avoids edge effects with no hits at all
michael@0:   hitbuffer->linear[linear_i].offset = hitbuffer->lowest_offset;
michael@0:   hitbuffer->linear[linear_i].type = base_hit;
michael@0:   hitbuffer->linear[linear_i].langprob =
michael@0:     DefaultLangProb(scoringcontext->ulscript);
michael@0:   ++linear_i;
michael@0: 
michael@0:   while ((base_i < base_limit) || (delta_i < delta_limit) ||
michael@0:          (distinct_i < distinct_limit)) {
michael@0:     int base_off = hitbuffer->base[base_i].offset;
michael@0:     int delta_off = hitbuffer->delta[delta_i].offset;
michael@0:     int distinct_off = hitbuffer->distinct[distinct_i].offset;
michael@0: 
michael@0:     // Do delta and distinct first, so that they are not lost at base_limit
michael@0:     if ((delta_i < delta_limit) &&
michael@0:         (delta_off <= base_off) && (delta_off <= distinct_off)) {
michael@0:       // Add delta entry
michael@0:       int indirect = hitbuffer->delta[delta_i].indirect;
michael@0:       ++delta_i;
michael@0:       uint32 langprob = delta_obj->kCLDTableInd[indirect];
michael@0:       if (langprob > 0) {
michael@0:         hitbuffer->linear[linear_i].offset = delta_off;
michael@0:         hitbuffer->linear[linear_i].type = DELTAHIT;
michael@0:         hitbuffer->linear[linear_i].langprob = langprob;
michael@0:         ++linear_i;
michael@0:       }
michael@0:     }
michael@0:     else if ((distinct_i < distinct_limit) &&
michael@0:              (distinct_off <= base_off) && (distinct_off <= delta_off)) {
michael@0:       // Add distinct entry
michael@0:       int indirect = hitbuffer->distinct[distinct_i].indirect;
michael@0:       ++distinct_i;
michael@0:       uint32 langprob = distinct_obj->kCLDTableInd[indirect];
michael@0:       if (langprob > 0) {
michael@0:         hitbuffer->linear[linear_i].offset = distinct_off;
michael@0:         hitbuffer->linear[linear_i].type = DISTINCTHIT;
michael@0:         hitbuffer->linear[linear_i].langprob = langprob;
michael@0:         ++linear_i;
michael@0:       }
michael@0:     }
michael@0:     else {
michael@0:       // Add one or two base entries
michael@0:       int indirect = hitbuffer->base[base_i].indirect;
michael@0:       // First, get right scoring table
michael@0:       const CLD2TableSummary* local_base_obj = base_obj;
michael@0:       if ((indirect & 0x80000000u) != 0) {
michael@0:         local_base_obj = base_obj2;
michael@0:         indirect &= ~0x80000000u;
michael@0:       }
michael@0:       ++base_i;
michael@0:       // One langprob in kQuadInd[0..SingleSize),
michael@0:       // two in kQuadInd[SingleSize..Size)
michael@0:       if (indirect < static_cast<int>(local_base_obj->kCLDTableSizeOne)) {
michael@0:         // Up to three languages at indirect
michael@0:         uint32 langprob = local_base_obj->kCLDTableInd[indirect];
michael@0:         if (langprob > 0) {
michael@0:           hitbuffer->linear[linear_i].offset = base_off;
michael@0:           hitbuffer->linear[linear_i].type = base_hit;
michael@0:           hitbuffer->linear[linear_i].langprob = langprob;
michael@0:           ++linear_i;
michael@0:         }
michael@0:       } else {
michael@0:         // Up to six languages at start + 2 * (indirect - start)
michael@0:         indirect += (indirect - local_base_obj->kCLDTableSizeOne);
michael@0:         uint32 langprob = local_base_obj->kCLDTableInd[indirect];
michael@0:         uint32 langprob2 = local_base_obj->kCLDTableInd[indirect + 1];
michael@0:         if (langprob > 0) {
michael@0:           hitbuffer->linear[linear_i].offset = base_off;
michael@0:           hitbuffer->linear[linear_i].type = base_hit;
michael@0:           hitbuffer->linear[linear_i].langprob = langprob;
michael@0:           ++linear_i;
michael@0:         }
michael@0:         if (langprob2 > 0) {
michael@0:           hitbuffer->linear[linear_i].offset = base_off;
michael@0:           hitbuffer->linear[linear_i].type = base_hit;
michael@0:           hitbuffer->linear[linear_i].langprob = langprob2;
michael@0:           ++linear_i;
michael@0:         }
michael@0:       }
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   // Update
michael@0:   hitbuffer->next_linear = linear_i;
michael@0: 
michael@0:   // Add a dummy entry off the end, just to capture final offset
michael@0:   hitbuffer->linear[linear_i].offset =
michael@0:   hitbuffer->base[hitbuffer->next_base].offset;
michael@0:   hitbuffer->linear[linear_i].langprob = 0;
michael@0: }
michael@0: 
michael@0: // Break linear array into chunks of ~20 quadgram hits or ~50 CJK unigram hits
michael@0: void ChunkAll(int letter_offset, bool score_cjk, ScoringHitBuffer* hitbuffer) {
michael@0:   int chunksize;
michael@0:   uint16 base_hit;
michael@0:   if (score_cjk) {
michael@0:     chunksize = kChunksizeUnis;
michael@0:     base_hit = UNIHIT;
michael@0:   } else {
michael@0:     chunksize = kChunksizeQuads;
michael@0:     base_hit = QUADHIT;
michael@0:   }
michael@0: 
michael@0:   int linear_i = 0;
michael@0:   int linear_off_end = hitbuffer->next_linear;
michael@0:   int text_i = letter_offset;               // Next unseen text offset
michael@0:   int next_chunk_start = 0;
michael@0:   int bases_left = hitbuffer->next_base;
michael@0:   while (bases_left > 0) {
michael@0:     // Linearize one chunk
michael@0:     int base_len = chunksize;     // Default; may be changed below
michael@0:     if (bases_left < (chunksize + (chunksize >> 1))) {
michael@0:       // If within 1.5 chunks of the end, avoid runts by using it all
michael@0:       base_len = bases_left;
michael@0:     } else if (bases_left < (2 * chunksize)) {
michael@0:       // Avoid runts by splitting 1.5 to 2 chunks in half (about 3/4 each)
michael@0:       base_len = (bases_left + 1) >> 1;
michael@0:     }
michael@0: 
michael@0:     hitbuffer->chunk_start[next_chunk_start] = linear_i;
michael@0:     hitbuffer->chunk_offset[next_chunk_start] = text_i;
michael@0:     ++next_chunk_start;
michael@0: 
michael@0:     int base_count = 0;
michael@0:     while ((base_count < base_len) && (linear_i < linear_off_end)) {
michael@0:       if (hitbuffer->linear[linear_i].type == base_hit) {++base_count;}
michael@0:       ++linear_i;
michael@0:     }
michael@0:     text_i = hitbuffer->linear[linear_i].offset;    // Next unseen text offset
michael@0:     bases_left -= base_len;
michael@0:   }
michael@0: 
michael@0:   // If no base hits at all, make a single dummy chunk
michael@0:   if (next_chunk_start == 0) {
michael@0:      hitbuffer->chunk_start[next_chunk_start] = 0;
michael@0:      hitbuffer->chunk_offset[next_chunk_start] = hitbuffer->linear[0].offset;
michael@0:      ++next_chunk_start;
michael@0:   }
michael@0: 
michael@0:   // Remember the linear array start of dummy entry
michael@0:   hitbuffer->next_chunk_start = next_chunk_start;
michael@0: 
michael@0:   // Add a dummy entry off the end, just to capture final linear subscr
michael@0:   hitbuffer->chunk_start[next_chunk_start] = hitbuffer->next_linear;
michael@0:   hitbuffer->chunk_offset[next_chunk_start] = text_i;
michael@0: }
michael@0: 
michael@0: 
michael@0: // Merge-sort the individual hit arrays, go indirect on the scoring subscripts,
michael@0: // break linear array into chunks.
michael@0: //
michael@0: // Input:
michael@0: //  hitbuffer base, delta, distinct arrays
michael@0: // Output:
michael@0: //  linear array
michael@0: //  chunk_start array
michael@0: //
michael@0: void LinearizeHitBuffer(int letter_offset,
michael@0:                         ScoringContext* scoringcontext,
michael@0:                         bool more_to_come, bool score_cjk,
michael@0:                         ScoringHitBuffer* hitbuffer) {
michael@0:   LinearizeAll(scoringcontext, score_cjk, hitbuffer);
michael@0:   ChunkAll(letter_offset, score_cjk, hitbuffer);
michael@0: }
michael@0: 
michael@0: 
michael@0: 
michael@0: // The hitbuffer is in an awkward form -- three sets of base/delta/distinct
michael@0: // scores, each with an indirect subscript to one of six scoring tables, some
michael@0: // of which can yield two langprobs for six languages, others one langprob for
michael@0: // three languages. The only correlation between base/delta/distinct is their
michael@0: // offsets into the letters-only text buffer.
michael@0: //
michael@0: // SummaryBuffer needs to be built to linear, giving linear offset of start of
michael@0: // each chunk
michael@0: //
michael@0: // So we first do all the langprob lookups and merge-sort by offset to make
michael@0: // a single linear vector, building a side vector of chunk beginnings as we go.
michael@0: // The sharpening is simply moving the beginnings, scoring is a simple linear
michael@0: // sweep, etc.
michael@0: 
michael@0: void ProcessHitBuffer(const LangSpan& scriptspan,
michael@0:                       int letter_offset,
michael@0:                       ScoringContext* scoringcontext,
michael@0:                       DocTote* doc_tote,
michael@0:                       ResultChunkVector* vec,
michael@0:                       bool more_to_come, bool score_cjk,
michael@0:                       ScoringHitBuffer* hitbuffer) {
michael@0:   if (scoringcontext->flags_cld2_verbose) {
michael@0:     fprintf(scoringcontext->debug_file, "Hitbuffer[) ");
michael@0:     DumpHitBuffer(scoringcontext->debug_file, scriptspan.text, hitbuffer);
michael@0:   }
michael@0: 
michael@0:   LinearizeHitBuffer(letter_offset, scoringcontext, more_to_come, score_cjk,
michael@0:                      hitbuffer);
michael@0: 
michael@0:   if (scoringcontext->flags_cld2_verbose) {
michael@0:     fprintf(scoringcontext->debug_file, "Linear[) ");
michael@0:     DumpLinearBuffer(scoringcontext->debug_file, scriptspan.text, hitbuffer);
michael@0:   }
michael@0: 
michael@0:   SummaryBuffer summarybuffer;
michael@0:   summarybuffer.n = 0;
michael@0:   ChunkSpan last_cspan;
michael@0:   ScoreAllHits(scriptspan.text, scriptspan.ulscript,
michael@0:                     more_to_come, score_cjk, hitbuffer,
michael@0:                     scoringcontext,
michael@0:                     &summarybuffer, &last_cspan);
michael@0: 
michael@0:   if (scoringcontext->flags_cld2_verbose) {
michael@0:     DumpSummaryBuffer(scoringcontext->debug_file, &summarybuffer);
michael@0:   }
michael@0: 
michael@0:   if (vec != NULL) {
michael@0:     // Sharpen boundaries of summarybuffer
michael@0:     // This is not a high-performance path
michael@0:     SharpenBoundaries(scriptspan.text, more_to_come, hitbuffer, scoringcontext,
michael@0:                       &summarybuffer);
michael@0:     // Show after the sharpening
michael@0:     // CLD2_Debug2(scriptspan.text, more_to_come, score_cjk,
michael@0:     //             hitbuffer, scoringcontext, &summarybuffer);
michael@0: 
michael@0:     if (scoringcontext->flags_cld2_verbose) {
michael@0:       DumpSummaryBuffer(scoringcontext->debug_file, &summarybuffer);
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   SummaryBufferToDocTote(&summarybuffer, more_to_come, doc_tote);
michael@0:   SummaryBufferToVector(scoringcontext->scanner, scriptspan.text,
michael@0:                         &summarybuffer, more_to_come, vec);
michael@0: }
michael@0: 
michael@0: void SpliceHitBuffer(ScoringHitBuffer* hitbuffer, int next_offset) {
michael@0:   // Splice hitbuffer and summarybuffer for next round. With big chunks and
michael@0:   // distinctive-word state carried across chunks, we might not need to do this.
michael@0:   hitbuffer->next_base = 0;
michael@0:   hitbuffer->next_delta = 0;
michael@0:   hitbuffer->next_distinct = 0;
michael@0:   hitbuffer->next_linear = 0;
michael@0:   hitbuffer->next_chunk_start = 0;
michael@0:   hitbuffer->lowest_offset = next_offset;
michael@0: }
michael@0: 
michael@0: 
michael@0: // Score RTypeNone or RTypeOne scriptspan into doc_tote and vec, updating
michael@0: // scoringcontext
michael@0: void ScoreEntireScriptSpan(const LangSpan& scriptspan,
michael@0:                            ScoringContext* scoringcontext,
michael@0:                            DocTote* doc_tote,
michael@0:                            ResultChunkVector* vec) {
michael@0:   int bytes = scriptspan.text_bytes;
michael@0:   // Artificially set score to 1024 per 1KB, or 1 per byte
michael@0:   int score = bytes;
michael@0:   int reliability = 100;
michael@0:   // doc_tote uses full languages
michael@0:   Language one_one_lang = DefaultLanguage(scriptspan.ulscript);
michael@0:   doc_tote->Add(one_one_lang, bytes, score, reliability);
michael@0: 
michael@0:   if (scoringcontext->flags_cld2_html) {
michael@0:     ChunkSummary chunksummary = {
michael@0:       1, 0,
michael@0:       one_one_lang, UNKNOWN_LANGUAGE, score, 1,
michael@0:       bytes, 0, scriptspan.ulscript, reliability, reliability
michael@0:     };
michael@0:     CLD2_Debug(scriptspan.text, 1, scriptspan.text_bytes,
michael@0:                false, false, NULL,
michael@0:                scoringcontext, NULL, &chunksummary);
michael@0:   }
michael@0: 
michael@0:   // First byte is always a space
michael@0:   JustOneItemToVector(scoringcontext->scanner, scriptspan.text,
michael@0:                       one_one_lang, 1, bytes - 1, vec);
michael@0: 
michael@0:   scoringcontext->prior_chunk_lang = UNKNOWN_LANGUAGE;
michael@0: }
michael@0: 
michael@0: // Score RTypeCJK scriptspan into doc_tote and vec, updating scoringcontext
michael@0: void ScoreCJKScriptSpan(const LangSpan& scriptspan,
michael@0:                         ScoringContext* scoringcontext,
michael@0:                         DocTote* doc_tote,
michael@0:                         ResultChunkVector* vec) {
michael@0:   // Allocate three parallel arrays of scoring hits
michael@0:   ScoringHitBuffer* hitbuffer = new ScoringHitBuffer;
michael@0:   hitbuffer->init();
michael@0:   hitbuffer->ulscript = scriptspan.ulscript;
michael@0: 
michael@0:   scoringcontext->prior_chunk_lang = UNKNOWN_LANGUAGE;
michael@0:   scoringcontext->oldest_distinct_boost = 0;
michael@0: 
michael@0:   // Incoming scriptspan has a single leading space at scriptspan.text[0]
michael@0:   // and three trailing spaces then NUL at scriptspan.text[text_bytes + 0/1/2/3]
michael@0: 
michael@0:   int letter_offset = 1;        // Skip initial space
michael@0:   hitbuffer->lowest_offset = letter_offset;
michael@0:   int letter_limit = scriptspan.text_bytes;
michael@0:   while (letter_offset < letter_limit) {
michael@0:     if (scoringcontext->flags_cld2_verbose) {
michael@0:       fprintf(scoringcontext->debug_file, " ScoreCJKScriptSpan[%d,%d)<br>\n",
michael@0:               letter_offset, letter_limit);
michael@0:     }
michael@0:     //
michael@0:     // Fill up one hitbuffer, possibly splicing onto previous fragment
michael@0:     //
michael@0:     // NOTE: GetUniHits deals with close repeats
michael@0:     // NOTE: After last chunk there is always a hitbuffer entry with an offset
michael@0:     // just off the end of the text = next_offset.
michael@0:     int next_offset = GetUniHits(scriptspan.text, letter_offset, letter_limit,
michael@0:                                   scoringcontext, hitbuffer);
michael@0:     // NOTE: GetBiHitVectors deals with close repeats,
michael@0:     // does one hash and two lookups (delta and distinct) per word
michael@0:     GetBiHits(scriptspan.text, letter_offset, next_offset,
michael@0:                 scoringcontext, hitbuffer);
michael@0: 
michael@0:     //
michael@0:     // Score one hitbuffer in chunks to summarybuffer
michael@0:     //
michael@0:     bool more_to_come = next_offset < letter_limit;
michael@0:     bool score_cjk = true;
michael@0:     ProcessHitBuffer(scriptspan, letter_offset, scoringcontext, doc_tote, vec,
michael@0:                      more_to_come, score_cjk, hitbuffer);
michael@0:     SpliceHitBuffer(hitbuffer, next_offset);
michael@0: 
michael@0:     letter_offset = next_offset;
michael@0:   }
michael@0: 
michael@0:   delete hitbuffer;
michael@0:   // Context across buffers is not connected yet
michael@0:   scoringcontext->prior_chunk_lang = UNKNOWN_LANGUAGE;
michael@0: }
michael@0: 
michael@0: 
michael@0: 
michael@0: // Score RTypeMany scriptspan into doc_tote and vec, updating scoringcontext
michael@0: // We have a scriptspan with all lowercase text in one script. Look up
michael@0: // quadgrams and octagrams, saving the hits in three parallel vectors.
michael@0: // Score from those vectors in chunks, toting each chunk to get a single
michael@0: // language, and combining into the overall document score. The hit vectors
michael@0: // in general are not big enough to handle and entire scriptspan, so
michael@0: // repeat until the entire scriptspan is scored.
michael@0: // Caller deals with minimizing numbr of runt scriptspans
michael@0: // This routine deals with minimizing number of runt chunks.
michael@0: //
michael@0: // Returns updated scoringcontext
michael@0: // Returns updated doc_tote
michael@0: // If vec != NULL, appends to that vector of ResultChunk's
michael@0: void ScoreQuadScriptSpan(const LangSpan& scriptspan,
michael@0:                          ScoringContext* scoringcontext,
michael@0:                          DocTote* doc_tote,
michael@0:                          ResultChunkVector* vec) {
michael@0:   // Allocate three parallel arrays of scoring hits
michael@0:   ScoringHitBuffer* hitbuffer = new ScoringHitBuffer;
michael@0:   hitbuffer->init();
michael@0:   hitbuffer->ulscript = scriptspan.ulscript;
michael@0: 
michael@0:   scoringcontext->prior_chunk_lang = UNKNOWN_LANGUAGE;
michael@0:   scoringcontext->oldest_distinct_boost = 0;
michael@0: 
michael@0:   // Incoming scriptspan has a single leading space at scriptspan.text[0]
michael@0:   // and three trailing spaces then NUL at scriptspan.text[text_bytes + 0/1/2/3]
michael@0: 
michael@0:   int letter_offset = 1;        // Skip initial space
michael@0:   hitbuffer->lowest_offset = letter_offset;
michael@0:   int letter_limit = scriptspan.text_bytes;
michael@0:   while (letter_offset < letter_limit) {
michael@0:     //
michael@0:     // Fill up one hitbuffer, possibly splicing onto previous fragment
michael@0:     //
michael@0:     // NOTE: GetQuadHits deals with close repeats
michael@0:     // NOTE: After last chunk there is always a hitbuffer entry with an offset
michael@0:     // just off the end of the text = next_offset.
michael@0:     int next_offset = GetQuadHits(scriptspan.text, letter_offset, letter_limit,
michael@0:                                   scoringcontext, hitbuffer);
michael@0:     // If true, there is more text to process in this scriptspan
michael@0:     // NOTE: GetOctaHitVectors deals with close repeats,
michael@0:     // does one hash and two lookups (delta and distinct) per word
michael@0:     GetOctaHits(scriptspan.text, letter_offset, next_offset,
michael@0:                 scoringcontext, hitbuffer);
michael@0: 
michael@0:     //
michael@0:     // Score one hitbuffer in chunks to summarybuffer
michael@0:     //
michael@0:     bool more_to_come = next_offset < letter_limit;
michael@0:     bool score_cjk = false;
michael@0:     ProcessHitBuffer(scriptspan, letter_offset, scoringcontext, doc_tote, vec,
michael@0:                      more_to_come, score_cjk, hitbuffer);
michael@0:     SpliceHitBuffer(hitbuffer, next_offset);
michael@0: 
michael@0:     letter_offset = next_offset;
michael@0:   }
michael@0: 
michael@0:   delete hitbuffer;
michael@0: }
michael@0: 
michael@0: 
michael@0: // Score one scriptspan into doc_tote and vec, updating scoringcontext
michael@0: // Inputs:
michael@0: //  One scriptspan of perhaps 40-60KB, all same script lower-case letters
michael@0: //    and single ASCII spaces. First character is a space to allow simple
michael@0: //    begining-of-word detect. End of buffer has three spaces and NUL to
michael@0: //    allow easy scan-to-end-of-word.
michael@0: //  Scoring context of
michael@0: //    scoring tables
michael@0: //    flags
michael@0: //    running boosts
michael@0: // Outputs:
michael@0: //  Updated doc_tote giving overall languages and byte counts
michael@0: //  Optional updated chunk vector giving offset, length, language
michael@0: //
michael@0: // Caller initializes flags, boosts, doc_tote and vec.
michael@0: // Caller aggregates across multiple scriptspans
michael@0: // Caller calculates final document result
michael@0: // Caller deals with detecting and triggering suppression of repeated text.
michael@0: //
michael@0: // This top-level routine just chooses the recognition type and calls one of
michael@0: // the next-level-down routines.
michael@0: //
michael@0: void ScoreOneScriptSpan(const LangSpan& scriptspan,
michael@0:                         ScoringContext* scoringcontext,
michael@0:                         DocTote* doc_tote,
michael@0:                         ResultChunkVector* vec) {
michael@0:   if (scoringcontext->flags_cld2_verbose) {
michael@0:     fprintf(scoringcontext->debug_file, "<br>ScoreOneScriptSpan(%s,%d) ",
michael@0:             ULScriptCode(scriptspan.ulscript), scriptspan.text_bytes);
michael@0:     // Optionally print the chunk lowercase letters/marks text
michael@0:     string temp(&scriptspan.text[0], scriptspan.text_bytes);
michael@0:     fprintf(scoringcontext->debug_file, "'%s'",
michael@0:             GetHtmlEscapedText(temp).c_str());
michael@0:     fprintf(scoringcontext->debug_file, "<br>\n");
michael@0:   }
michael@0:   scoringcontext->prior_chunk_lang = UNKNOWN_LANGUAGE;
michael@0:   scoringcontext->oldest_distinct_boost = 0;
michael@0:   ULScriptRType rtype = ULScriptRecognitionType(scriptspan.ulscript);
michael@0:   if (scoringcontext->flags_cld2_score_as_quads && (rtype != RTypeCJK)) {
michael@0:     rtype = RTypeMany;
michael@0:   }
michael@0:   switch (rtype) {
michael@0:   case RTypeNone:
michael@0:   case RTypeOne:
michael@0:     ScoreEntireScriptSpan(scriptspan, scoringcontext, doc_tote, vec);
michael@0:     break;
michael@0:   case RTypeCJK:
michael@0:     ScoreCJKScriptSpan(scriptspan, scoringcontext, doc_tote, vec);
michael@0:     break;
michael@0:   case RTypeMany:
michael@0:     ScoreQuadScriptSpan(scriptspan, scoringcontext, doc_tote, vec);
michael@0:     break;
michael@0:   }
michael@0: }
michael@0: 
michael@0: }       // End namespace CLD2
michael@0: