Apache
/
lucene
mirror of https://github.com/apache/lucene.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

LUCENE-3892: backporting r1371010 and r1371011 to BlockPacked. 

I changed the comment on the value of BLOCK_SIZE. Indeed, since PackedInts
encoding/decoding is long-aligned, all numbers of bits per value that are
co-prime with 64 cannot encode/decode less than 64 values per iteration, so the
block size should be at least 64. Switching it to a lower value (eg. 32) should
work but will perform many unnecessary operations in the encoding/decoding step.


git-svn-id: https://svn.apache.org/repos/asf/lucene/dev/branches/pforcodec_3892@1371114 13f79535-47bb-0310-9956-ffa450edef68
This commit is contained in:
Adrien Grand 2012-08-09 09:16:09 +00:00
parent 98ae806ba6
commit 0be8bfb4fc
3 changed files with 72 additions and 49 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
lucene/core/src/java/org/apache/lucene/codecs/blockpacked/BlockPackedPostingsFormat.java
View File

@ -41,6 +41,10 @@ public final class BlockPackedPostingsFormat extends PostingsFormat {
private final int minTermBlockSize;
private final int maxTermBlockSize;
// nocommit is this right?:
// NOTE: should be at least 64 because of PackedInts long-aligned encoding/decoding
// NOTE: must be factor of ... 64?
public final static int BLOCK_SIZE = 128;
public BlockPackedPostingsFormat() {

109
lucene/core/src/java/org/apache/lucene/codecs/blockpacked/BlockPackedPostingsReader.java
View File

@ -384,7 +384,6 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
}
private void refillDocs() throws IOException {
//System.out.println("["+docFreq+"]"+" refillDoc");
final int left = docFreq - docUpto;
assert left > 0;
@ -426,7 +425,6 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
}
return doc = NO_MORE_DOCS;
}
//System.out.println("["+docFreq+"]"+" nextDoc");
if (docBufferUpto == BLOCK_SIZE) {
refillDocs();
}
@ -486,15 +484,15 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
skipped = true;
}
final int newDocUpto = skipper.skipTo(target);
final int newDocUpto = skipper.skipTo(target) + 1;
if (newDocUpto > docUpto) {
// Skipper moved
if (DEBUG) {
System.out.println("skipper moved to docUpto=" + newDocUpto + " vs current=" + docUpto + "; docID=" + skipper.getDoc() + " fp=" + skipper.getDocPointer());
}
assert newDocUpto % BLOCK_SIZE == (BLOCK_SIZE - 1): "got " + newDocUpto;
docUpto = newDocUpto+1;
assert newDocUpto % BLOCK_SIZE == 0 : "got " + newDocUpto;
docUpto = newDocUpto;
// Force to read next block
docBufferUpto = BLOCK_SIZE;
@ -503,6 +501,12 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
}
nextSkipDoc = skipper.getNextSkipDoc();
}
if (docUpto == docFreq) {
return doc = NO_MORE_DOCS;
}
if (docBufferUpto == BLOCK_SIZE) {
refillDocs();
}
// Now scan... this is an inlined/pared down version
// of nextDoc():
@ -510,18 +514,6 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
if (DEBUG) {
System.out.println(" scan doc=" + accum + " docBufferUpto=" + docBufferUpto);
}
if (docUpto == docFreq) {
return doc = NO_MORE_DOCS;
}
// nocommit: in theory we should not hit this? ie
// skipper should already have moved us to the block
// containing the doc? yet assert false trips ... i
// think because if you advance w/o having done a
// nextDoc yet()... can we assert/remove this?
if (docBufferUpto == BLOCK_SIZE) {
refillDocs();
}
accum += docDeltaBuffer[docBufferUpto];
docUpto++;
@ -529,6 +521,9 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
break;
}
docBufferUpto++;
if (docUpto == docFreq) {
return doc = NO_MORE_DOCS;
}
}
if (liveDocs == null || liveDocs.get(accum)) {
@ -666,9 +661,9 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
}
private void refillDocs() throws IOException {
//System.out.println("["+docFreq+"]"+" refillDoc");
final int left = docFreq - docUpto;
assert left > 0;
if (left >= BLOCK_SIZE) {
if (DEBUG) {
System.out.println(" fill doc block from fp=" + docIn.getFilePointer());
@ -797,7 +792,7 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
skipped = true;
}
final int newDocUpto = skipper.skipTo(target);
final int newDocUpto = skipper.skipTo(target) + 1;
if (newDocUpto > docUpto) {
// Skipper moved
@ -805,8 +800,8 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
System.out.println(" skipper moved to docUpto=" + newDocUpto + " vs current=" + docUpto + "; docID=" + skipper.getDoc() + " fp=" + skipper.getDocPointer() + " pos.fp=" + skipper.getPosPointer() + " pos.bufferUpto=" + skipper.getPosBufferUpto());
}
assert newDocUpto % BLOCK_SIZE == (BLOCK_SIZE - 1): "got " + newDocUpto;
docUpto = newDocUpto+1;
assert newDocUpto % BLOCK_SIZE == 0 : "got " + newDocUpto;
docUpto = newDocUpto;
// Force to read next block
docBufferUpto = BLOCK_SIZE;
@ -817,6 +812,12 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
}
nextSkipDoc = skipper.getNextSkipDoc();
}
if (docUpto == docFreq) {
return doc = NO_MORE_DOCS;
}
if (docBufferUpto == BLOCK_SIZE) {
refillDocs();
}
// Now scan... this is an inlined/pared down version
// of nextDoc():
@ -827,16 +828,6 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
if (docUpto == docFreq) {
return doc = NO_MORE_DOCS;
}
// nocommit: in theory we should not hit this? ie
// skipper should already have moved us to the block
// containing the doc? yet assert false trips ... i
// think because if you advance w/o having done a
// nextDoc yet()... can we assert/remove this?
if (docBufferUpto == BLOCK_SIZE) {
// nocommit hmm skip freq? but: we don't ever
// scan over more than one block?
refillDocs();
}
accum += docDeltaBuffer[docBufferUpto];
freq = freqBuffer[docBufferUpto];
posPendingCount += freq;
@ -846,6 +837,9 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
if (accum >= target) {
break;
}
if (docUpto == docFreq) {
return doc = NO_MORE_DOCS;
}
}
if (liveDocs == null || liveDocs.get(accum)) {
@ -1110,7 +1104,6 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
}
private void refillDocs() throws IOException {
//System.out.println("["+docFreq+"]"+" refillDoc");
final int left = docFreq - docUpto;
assert left > 0;
@ -1226,7 +1219,6 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
if (docUpto == docFreq) {
return doc = NO_MORE_DOCS;
}
//System.out.println("["+docFreq+"]"+" nextDoc");
if (docBufferUpto == BLOCK_SIZE) {
refillDocs();
}
@ -1293,15 +1285,15 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
skipped = true;
}
final int newDocUpto = skipper.skipTo(target);
final int newDocUpto = skipper.skipTo(target) + 1;
if (newDocUpto > docUpto) {
// Skipper moved
if (DEBUG) {
System.out.println(" skipper moved to docUpto=" + newDocUpto + " vs current=" + docUpto + "; docID=" + skipper.getDoc() + " fp=" + skipper.getDocPointer() + " pos.fp=" + skipper.getPosPointer() + " pos.bufferUpto=" + skipper.getPosBufferUpto() + " pay.fp=" + skipper.getPayPointer() + " lastStartOffset=" + lastStartOffset);
}
assert newDocUpto % BLOCK_SIZE == (BLOCK_SIZE - 1): "got " + newDocUpto;
docUpto = newDocUpto+1;
assert newDocUpto % BLOCK_SIZE == 0 : "got " + newDocUpto;
docUpto = newDocUpto;
// Force to read next block
docBufferUpto = BLOCK_SIZE;
@ -1315,24 +1307,51 @@ public final class BlockPackedPostingsReader extends PostingsReaderBase {
}
nextSkipDoc = skipper.getNextSkipDoc();
}
// nocommit inline nextDoc here
if (docUpto == docFreq) {
return doc = NO_MORE_DOCS;
}
if (docBufferUpto == BLOCK_SIZE) {
refillDocs();
}
// Now scan:
while (nextDoc() != NO_MORE_DOCS) {
if (doc >= target) {
// Now scan:
while (true) {
if (DEBUG) {
System.out.println(" advance return doc=" + doc);
System.out.println(" scan doc=" + accum + " docBufferUpto=" + docBufferUpto);
}
return doc;
accum += docDeltaBuffer[docBufferUpto];
freq = freqBuffer[docBufferUpto];
posPendingCount += freq;
docBufferUpto++;
docUpto++;
if (accum >= target) {
break;
}
if (docUpto == docFreq) {
return doc = NO_MORE_DOCS;
}
}
if (liveDocs == null || liveDocs.get(accum)) {
if (DEBUG) {
System.out.println(" advance return doc=END");
System.out.println(" return doc=" + accum);
}
if (indexHasPayloads) {
payloadByteUpto += payloadLength;
payloadLength = 0;
}
position = 0;
payloadLength = 0;
lastStartOffset = 0;
return doc = accum;
} else {
if (DEBUG) {
System.out.println(" now do nextDoc()");
}
return nextDoc();
}
return NO_MORE_DOCS;
}
// nocommit in theory we could avoid loading frq block

2
lucene/core/src/java/org/apache/lucene/codecs/blockpacked/BlockPackedSkipWriter.java
View File

@ -37,7 +37,7 @@ import org.apache.lucene.codecs.MultiLevelSkipListWriter;
* block, only record skip data at the start its start point(if it exist).
*
* For each skip point, we will record:
* 1. lastDocID,
* 1. docID in former position, i.e. for position 12, record docID[11], etc.
* 2. its related file points(position, payload),
* 3. related numbers or uptos(position, payload).
* 4. start offset.