src.it.unimi.dsi.big.mg4j.index.BitStreamIndexWriter Maven / Gradle / Ivy
Show all versions of mg4j-big Show documentation
package it.unimi.dsi.big.mg4j.index;
/*
* MG4J: Managing Gigabytes for Java (big)
*
* Copyright (C) 2003-2011 Paolo Boldi and Sebastiano Vigna
*
* This library is free software; you can redistribute it and/or modify it
* under the terms of the GNU Lesser General Public License as published by the Free
* Software Foundation; either version 3 of the License, or (at your option)
* any later version.
*
* This library is distributed in the hope that it will be useful, but
* WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
* or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License
* for more details.
*
* You should have received a copy of the GNU Lesser General Public License
* along with this program; if not, see Offsets bit stream
*
* An inverted index may have an associated {@link OutputBitStream} of
* offsets: this file contains T+1 integers, where T
* is the number of inverted lists (i.e., the number of terms), and the
* i-th entry is the position in bits where
* the i-th inverted list starts (the last entry is actually the
* length, in bytes, of the inverted index file itself).
*
*
The file actually contains γ-coded gaps: thus, in practice, it is formed by
* the number zero (the offset of the first list) followed by the length of each inverted list.
*
* @author Paolo Boldi
* @author Sebastiano Vigna
* @since 0.6
*/
public class BitStreamIndexWriter extends AbstractBitStreamIndexWriter {
private static final boolean ASSERTS = false;
/** This value of {@link #state} means that we should call {@link #newInvertedList()}.*/
protected static final int BEFORE_INVERTED_LIST = 0;
/** This value of {@link #state} means that we are positioned at the start of an inverted list,
* and we should call {@link #writeFrequency(long)}.*/
protected static final int BEFORE_FREQUENCY = 1;
/** This value of {@link #state} means that we are ready to call {@link #newDocumentRecord()}. */
protected static final int BEFORE_DOCUMENT_RECORD = 2;
/** This value of {@link #state} means that we just started a new document record, and we
* should call {@link #writeDocumentPointer(OutputBitStream, long)}. */
protected static final int BEFORE_POINTER = 3;
/** This value of {@link #state} can be assumed only in indices that contain payloads; it
* means that we are positioned just before the payload for the current document record. */
protected static final int BEFORE_PAYLOAD = 4;
/** This value of {@link #state} can be assumed only in indices that contain counts; it
* means that we are positioned just before the count for the current document record. */
protected static final int BEFORE_COUNT = 5;
/** This value of {@link #state} can be assumed only in indices that contain document positions;
* it means that we are positioned just before the position list of the current document record. */
protected static final int BEFORE_POSITIONS = 6;
/** This is the first unused state. Subclasses may start from this value to define new states. */
protected static final int FIRST_UNUSED_STATE = 7;
/** The underlying {@link OutputBitStream}. */
protected OutputBitStream obs;
/** The offsets {@link OutputBitStream}. */
private OutputBitStream offsets;
/** The {@link OutputBitStream} for the number of bits for positions. */
private OutputBitStream posNumBits;
/** The current state of the writer. */
protected int state;
/** The number of document records that the current inverted list will contain. */
protected long frequency;
/** The number of document records already written for the current inverted list. */
protected long writtenDocuments;
/** The current document pointer. */
protected long currentDocument;
/** The last document pointer in the current list. */
protected long lastDocument;
/** The position (in bytes) where the last inverted list started. */
protected long lastInvertedListPos;
/** The number of bits spent for positions in this the current inverted list. */
private long currPosNumBits;
/** The parameter b for Golomb coding of pointers. */
protected int b;
/** The parameter log2b for Golomb coding of pointers; it is the most significant bit of {@link #b}. */
protected int log2b;
/** The maximum number of positions in a document record so far. */
public int maxCount;
/** Creates a new index writer, with the specified basename. The index will be written on a file (stemmed with .index).
* If writeOffsets, also an offset file will be produced (stemmed with .offsets).
* When {@link #close()} will be called, the property file will also be produced (stemmed with .properties),
* or enriched if it already exists.
*
* @param basename the basename.
* @param numberOfDocuments the number of documents in the collection to be indexed.
* @param writeOffsets if true, the offset file will also be produced.
* @param flags a flag map setting the coding techniques to be used (see {@link CompressionFlags}).
*/
public BitStreamIndexWriter( final CharSequence basename, final long numberOfDocuments, final boolean writeOffsets, final Map flags ) throws IOException {
this(
new OutputBitStream( basename + DiskBasedIndex.INDEX_EXTENSION ),
writeOffsets ? new OutputBitStream( basename + DiskBasedIndex.OFFSETS_EXTENSION ) : null,
writeOffsets && flags.get( Component.POSITIONS ) != null ? new OutputBitStream( basename + DiskBasedIndex.POSITIONS_NUMBER_OF_BITS_EXTENSION ) : null,
numberOfDocuments,
flags
);
}
/** Creates a new index writer with payloads using the specified underlying {@link OutputBitStream}.
*
* @param obs the underlying output bit stream.
* @param offset the offset bit stream, or null if offsets should not be written.
* @param posNumBits the bit stream for positions bit lengths, or null if such lengths should not be written.
* @param numberOfDocuments the number of documents in the collection to be indexed.
* @param flags a flag map setting the coding techniques to be used (see {@link CompressionFlags}).
*/
public BitStreamIndexWriter( final OutputBitStream obs, final OutputBitStream offset, final OutputBitStream posNumBits, final long numberOfDocuments, final Map flags ) {
super( numberOfDocuments, flags );
this.obs = obs;
this.posNumBits = posNumBits;
this.offsets = offset;
this.frequency = -1;
this.currentTerm = -1;
this.maxCount = 0;
this.currPosNumBits = -1;
if ( ! hasCounts && hasPositions ) throw new IllegalArgumentException( "Index would have positions but no counts (this can't happen)" );
}
/** Creates a new index writer, with the specified underlying {@link OutputBitStream},
* without additional bit streams.
*
* @param obs the underlying output bit stream.
* @param numberOfDocuments the number of documents in the collection to be indexed.
* @param flags a flag map setting the coding techniques to be used (see {@link CompressionFlags}).
*/
public BitStreamIndexWriter( final OutputBitStream obs, final long numberOfDocuments, final Map flags ) {
this( obs, null,null, numberOfDocuments, flags );
}
public long newInvertedList() throws IOException {
if ( frequency >= 0 && frequency != writtenDocuments ) throw new IllegalStateException( "The number of document records (" + this.writtenDocuments + ") does not match the frequency (" + this.frequency + ")" );
if ( state != BEFORE_INVERTED_LIST && state != BEFORE_DOCUMENT_RECORD ) throw new IllegalStateException( "Trying to start new inverted list in state " + state );
// The position (in bits) where the new inverted list starts
long pos = obs.writtenBits();
// Reset variables
writtenDocuments = 0;
currentTerm++;
currentDocument = -1;
// If needed, write the offset
if ( offsets != null ) offsets.writeLongGamma( pos - lastInvertedListPos );
lastInvertedListPos = pos;
if ( posNumBits != null && currPosNumBits != -1 ) {
posNumBits.writeLongGamma( currPosNumBits );
currPosNumBits = 0;
}
state = BEFORE_FREQUENCY;
return pos;
}
public int writeFrequency( final long frequency ) throws IOException {
if ( state != BEFORE_FREQUENCY ) throw new IllegalStateException( "Trying to write frequency in state " + state );
int bitCount;
// Write the frequency
switch( frequencyCoding ) {
case SHIFTED_GAMMA:
bitCount = obs.writeLongShiftedGamma( frequency - 1 ); // frequency cannot be 0
break;
case GAMMA:
bitCount = obs.writeLongGamma( frequency - 1 ); // frequency cannot be 0
break;
case DELTA:
bitCount = obs.writeLongDelta( frequency - 1 ); // frequency cannot be 0
break;
default:
throw new IllegalStateException( "The required frequency coding (" + frequencyCoding + ") is not supported." );
}
this.frequency = frequency;
// We compute the modulus used for pointer Golomb coding
if ( pointerCoding == Coding.GOLOMB ) {
b = BitStreamIndex.golombModulus( frequency, numberOfDocuments );
log2b = Fast.mostSignificantBit( b );
}
state = BEFORE_DOCUMENT_RECORD;
bitsForFrequencies += bitCount;
return bitCount;
}
public OutputBitStream newDocumentRecord() throws IOException {
if ( frequency == writtenDocuments ) throw new IllegalStateException( "Document record overflow (written " + this.frequency + " already)" );
if ( state != BEFORE_DOCUMENT_RECORD ) throw new IllegalStateException( "Trying to start new document record in state " + state );
writtenDocuments++;
numberOfPostings++;
lastDocument = currentDocument;
state = BEFORE_POINTER;
return obs;
}
public long writeDocumentPointer( final OutputBitStream out, final long pointer ) throws IOException {
if ( state != BEFORE_POINTER ) throw new IllegalStateException( "Trying to write pointer in state " + state );
currentDocument = pointer;
long bitCount = 0;
if ( frequency != numberOfDocuments ) { // We do not write pointers for everywhere occurring documents.
switch( pointerCoding ) {
case SHIFTED_GAMMA:
bitCount = out.writeLongShiftedGamma( pointer - lastDocument - 1 );
break;
case UNARY:
bitCount = out.writeLongUnary( pointer - lastDocument - 1 );
break;
case GAMMA:
bitCount = out.writeLongGamma( pointer - lastDocument - 1 );
break;
case DELTA:
bitCount = out.writeLongDelta( pointer - lastDocument - 1 );
break;
case GOLOMB:
bitCount = out.writeLongGolomb( pointer - lastDocument - 1, b, log2b );
break;
default:
throw new IllegalStateException( "The required pointer coding (" + pointerCoding + ") is not supported." );
}
}
else if ( pointer - lastDocument != 1 ) throw new IllegalStateException( "Term " + currentTerm + " has frequency equal to the number of documents, but pointers are not consecutive integers" );
state = hasPayloads ? BEFORE_PAYLOAD : hasCounts ? BEFORE_COUNT : BEFORE_DOCUMENT_RECORD;
bitsForPointers += bitCount;
return bitCount;
}
public int writePayload( final OutputBitStream out, final Payload payload ) throws IOException {
if ( frequency < 0 ) throw new IllegalStateException( "Trying to write payload without calling newInvertedList" );
if ( state != BEFORE_PAYLOAD ) throw new IllegalStateException( "Trying to write payload in state " + state );
final int count = payload.write( out );
bitsForPayloads += count;
state = hasCounts ? BEFORE_COUNT : BEFORE_DOCUMENT_RECORD;
return count;
}
public void close() throws IOException {
if ( state != BEFORE_DOCUMENT_RECORD && state != BEFORE_INVERTED_LIST ) throw new IllegalStateException( "Trying to close index in state " + state );
if ( frequency >= 0 && frequency != writtenDocuments ) throw new IllegalStateException( "The number of document records (" + this.writtenDocuments + ") does not match the frequency (" + this.frequency + ")" );
if ( writtenBits() != obs.writtenBits() )
throw new IllegalStateException( "Written bits count mismatch: we say " + writtenBits() + ", the stream says " + obs.writtenBits() );
if ( offsets != null ) {
offsets.writeLongGamma( obs.writtenBits() - lastInvertedListPos );
offsets.close();
}
if ( posNumBits != null ) {
if ( currPosNumBits != -1 ) posNumBits.writeLongGamma( currPosNumBits );
posNumBits.close();
}
obs.close();
}
public int writePositionCount( final OutputBitStream out, final int count ) throws IOException {
if ( frequency < 0 ) throw new IllegalStateException( "Trying to write count without calling newInvertedList" );
if ( state != BEFORE_COUNT ) throw new IllegalStateException( "Trying to write count in state " + state );
final int bitCount;
numberOfOccurrences += count;
switch( countCoding ) {
case SHIFTED_GAMMA:
bitCount = out.writeShiftedGamma( count - 1 );
break;
case GAMMA:
bitCount = out.writeGamma( count - 1 );
break;
case UNARY:
bitCount = out.writeUnary( count - 1 );
break;
case DELTA:
bitCount = out.writeDelta( count - 1 );
break;
default:
throw new IllegalStateException( "The required count coding (" + countCoding + ") is not supported." );
}
state = hasPositions ? BEFORE_POSITIONS : BEFORE_DOCUMENT_RECORD;
bitsForCounts += bitCount;
return bitCount;
}
public int writeDocumentPositions( final OutputBitStream out, final int[] occ, final int offset, final int len, final int docSize ) throws IOException {
if ( frequency < 0 ) throw new IllegalStateException( "Trying to write occurrences without calling newInvertedList" );
if ( state != BEFORE_POSITIONS ) throw new IllegalStateException( "Trying to write positions in state " + state );
if ( ASSERTS ) {
if ( docSize > 0 ) for( int i = 0; i< len; i++ ) assert occ[ offset + i ] < docSize : "Position " + occ[ offset + i ] + " for document " + currentDocument + " is too large; size is " + docSize;
}
int i;
int prev = -1;
int bitCount = 0;
final int end = offset + len;
switch( positionCoding ) {
case GAMMA:
for( i = offset; i < end; i++ ) {
bitCount += out.writeGamma( occ[ i ] - prev - 1 );
prev = occ[ i ];
}
break;
case DELTA:
for( i = offset; i < end; i++ ) {
bitCount += out.writeDelta( occ[ i ] - prev - 1 );
prev = occ[ i ];
}
break;
case SHIFTED_GAMMA:
for( i = offset; i < end; i++ ) {
bitCount += out.writeShiftedGamma( occ[ i ] - prev - 1 );
prev = occ[ i ];
}
break;
case GOLOMB:
if ( len < 3 ) {
for( i = 0; i < len; i++ ) bitCount += out.writeMinimalBinary( occ[ i ], docSize );
break;
}
// We compute b and log2b for positions
final int positionB = BitStreamIndex.golombModulus( len, docSize );
final int positionLog2b = Fast.mostSignificantBit( positionB );
for( i = offset; i < end; i++ ) {
bitCount += out.writeGolomb( occ[ i ] - prev - 1, positionB, positionLog2b );
prev = occ[ i ];
}
break;
case INTERPOLATIVE:
bitCount = InterpolativeCoding.write( out, occ, 0, len, 0, docSize - 1 );
break;
default:
throw new IllegalStateException( "The required position coding (" + positionCoding + ") is not supported." );
}
state = BEFORE_DOCUMENT_RECORD;
bitsForPositions += bitCount;
currPosNumBits += bitCount;
if ( len > maxCount ) maxCount = len;
return bitCount;
}
public long writtenBits() {
return bitsForFrequencies + bitsForPointers + bitsForPayloads + bitsForCounts + bitsForPositions;
}
public Properties properties() {
Properties result = new Properties();
result.setProperty( Index.PropertyKeys.DOCUMENTS, numberOfDocuments );
result.setProperty( Index.PropertyKeys.TERMS, currentTerm + 1 );
result.setProperty( Index.PropertyKeys.POSTINGS, numberOfPostings );
result.setProperty( Index.PropertyKeys.MAXCOUNT, maxCount );
result.setProperty( Index.PropertyKeys.INDEXCLASS, FileIndex.class.getName() );
// We save all flags, except for the PAYLOAD component, which is just used internally.
for( Map.Entry e: flags.entrySet() )
if ( e.getKey() != Component.PAYLOADS ) result.addProperty( Index.PropertyKeys.CODING, new MutableString().append( e.getKey() ).append( ':' ).append( e.getValue() ) );
return result;
}
}