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

LUCENE-4171: Performance improvements to Packed64 (Toke Eskildsen). 

git-svn-id: https://svn.apache.org/repos/asf/lucene/dev/trunk@1355346 13f79535-47bb-0310-9956-ffa450edef68
This commit is contained in:
Adrien Grand 2012-06-29 12:52:54 +00:00
parent 32a7d50eb3
commit 1e21e6ceec
2 changed files with 82 additions and 117 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
lucene/CHANGES.txt
View File

@ -1036,6 +1036,9 @@ Optimizations
the cloned instances. WeakIdentityMap was extended to support
iterating over its keys. (Uwe Schindler)
* LUCENE-4171: Performance improvements to Packed64.
(Toke Eskildsen via Adrien Grand)
Bug fixes
* LUCENE-2803: The FieldCache can miss values if an entry for a reader

196
lucene/core/src/java/org/apache/lucene/util/packed/Packed64.java
View File

@ -25,99 +25,40 @@ import java.util.Arrays;
/**
* Space optimized random access capable array of values with a fixed number of
* bits. For 32 bits/value and less, performance on 32 bit machines is not
* optimal. Consider using {@link Packed32} for such a setup.
* bits/value. Values are packed contiguously.
* </p><p>
* The implementation strives to avoid conditionals and expensive operations,
* sacrificing code clarity to achieve better performance.
* The implementation strives to perform af fast as possible under the
* constraint of contiguous bits, by avoiding expensive operations. This comes
* at the cost of code clarity.
* </p><p>
* Technical details: This implementation is a refinement of a non-branching
* version. The non-branching get and set methods meant that 2 or 4 atomics in
* the underlying array were always accessed, even for the cases where only
* 1 or 2 were needed. Even with caching, this had a detrimental effect on
* performance.
* Related to this issue, the old implementation used lookup tables for shifts
* and masks, which also proved to be a bit slower than calculating the shifts
* and masks on the fly.
* See https://issues.apache.org/jira/browse/LUCENE-4062 for details.
*
*/
class Packed64 extends PackedInts.MutableImpl {
static final int BLOCK_SIZE = 64; // 32 = int, 64 = long
static final int BLOCK_BITS = 6; // The #bits representing BLOCK_SIZE
static final int MOD_MASK = BLOCK_SIZE - 1; // x % BLOCK_SIZE
private static final int ENTRY_SIZE = BLOCK_SIZE + 1;
static final int FAC_BITPOS = 3;
/*
* In order to make an efficient value-getter, conditionals should be
* avoided. A value can be positioned inside of a block, requiring shifting
* left or right or it can span two blocks, requiring a left-shift on the
* first block and a right-shift on the right block.
* </p><p>
* By always shifting the first block both left and right, we get exactly
* the right bits. By always shifting the second block right and applying
* a mask, we get the right bits there. After that, we | the two bitsets.
*/
static final int[][] SHIFTS =
new int[ENTRY_SIZE][ENTRY_SIZE * FAC_BITPOS];
static final long[][] MASKS = new long[ENTRY_SIZE][ENTRY_SIZE];
static { // Generate shifts
for (int elementBits = 1 ; elementBits <= BLOCK_SIZE ; elementBits++) {
for (int bitPos = 0 ; bitPos < BLOCK_SIZE ; bitPos++) {
int[] currentShifts = SHIFTS[elementBits];
int base = bitPos * FAC_BITPOS;
currentShifts[base ] = bitPos;
currentShifts[base + 1] = BLOCK_SIZE - elementBits;
if (bitPos <= BLOCK_SIZE - elementBits) { // Single block
currentShifts[base + 2] = 0;
MASKS[elementBits][bitPos] = 0;
} else { // Two blocks
int rBits = elementBits - (BLOCK_SIZE - bitPos);
currentShifts[base + 2] = BLOCK_SIZE - rBits;
MASKS[elementBits][bitPos] = ~(~0L << rBits);
}
}
}
}
/*
* The setter requires more masking than the getter.
*/
private static final long[][] WRITE_MASKS =
new long[ENTRY_SIZE][ENTRY_SIZE * FAC_BITPOS];
static {
for (int elementBits = 1 ; elementBits <= BLOCK_SIZE ; elementBits++) {
long elementPosMask = ~(~0L << elementBits);
int[] currentShifts = SHIFTS[elementBits];
long[] currentMasks = WRITE_MASKS[elementBits];
for (int bitPos = 0 ; bitPos < BLOCK_SIZE ; bitPos++) {
int base = bitPos * FAC_BITPOS;
currentMasks[base ] =~((elementPosMask
<< currentShifts[base + 1])
>>> currentShifts[base]);
if (bitPos <= BLOCK_SIZE - elementBits) { // Second block not used
currentMasks[base+1] = ~0; // Keep all bits
currentMasks[base+2] = 0; // Or with 0
} else {
currentMasks[base+1] = ~(elementPosMask
<< currentShifts[base + 2]);
currentMasks[base+2] = currentShifts[base + 2] == 0 ? 0 : ~0;
}
}
}
}
private static int pgcd(int a, int b) {
if (a < b) {
return pgcd(b, a);
} else if (b == 0) {
return a;
} else {
return pgcd(b, a % b);
}
}
/* The bits */
/**
* Values are stores contiguously in the blocks array.
*/
private final long[] blocks;
// Cached calculations
private int maxPos; // blocks.length * BLOCK_SIZE / elementBits - 1
private int[] shifts; // The shifts for the current elementBits
private long[] readMasks;
private long[] writeMasks;
/**
* A right-aligned mask of width BitsPerValue used by {@link #get(int)}.
*/
private final long maskRight;
/**
* Optimization: Saves one lookup in {@link #get(int)}.
*/
private final int bpvMinusBlockSize;
/**
* Creates an array with the internal structures adjusted for the given
@ -126,18 +67,18 @@ class Packed64 extends PackedInts.MutableImpl {
* @param bitsPerValue the number of bits available for any given value.
*/
public Packed64(int valueCount, int bitsPerValue) {
// TODO: Test for edge-cases (2^31 values, 63 bitsPerValue)
// +2 due to the avoid-conditionals-trick. The last entry is always 0
this(new long[(int)((long)valueCount * bitsPerValue / BLOCK_SIZE + 2)],
// NOTE: block-size was previously calculated as
// valueCount * bitsPerValue / BLOCK_SIZE + 1
// due to memory layout requirements dictated by non-branching code
this(new long[size(valueCount, bitsPerValue)],
valueCount, bitsPerValue);
}
/**
* Creates an array backed by the given blocks.
* </p><p>
* Note: The blocks are used directly, so changes to the given block will
* affect the Packed32-structure.
* affect the Packed64-structure.
* @param blocks used as the internal backing array. Not that the last
* element cannot be addressed directly.
* @param valueCount the number of values.
@ -146,7 +87,8 @@ class Packed64 extends PackedInts.MutableImpl {
public Packed64(long[] blocks, int valueCount, int bitsPerValue) {
super(valueCount, bitsPerValue);
this.blocks = blocks;
updateCached();
maskRight = ~0L << (BLOCK_SIZE-bitsPerValue) >>> (BLOCK_SIZE-bitsPerValue);
bpvMinusBlockSize = bitsPerValue - BLOCK_SIZE;
}
/**
@ -161,12 +103,12 @@ class Packed64 extends PackedInts.MutableImpl {
throws IOException {
super(valueCount, bitsPerValue);
int size = size(valueCount, bitsPerValue);
blocks = new long[size+1]; // +1 due to non-conditional tricks
// TODO: find a faster way to bulk-read longs...
blocks = new long[size]; // Previously +1 due to non-conditional tricks
for(int i=0;i<size;i++) {
blocks[i] = in.readLong();
}
updateCached();
maskRight = ~0L << (BLOCK_SIZE-bitsPerValue) >>> (BLOCK_SIZE-bitsPerValue);
bpvMinusBlockSize = bitsPerValue - BLOCK_SIZE;
}
private static int size(int valueCount, int bitsPerValue) {
@ -174,48 +116,57 @@ class Packed64 extends PackedInts.MutableImpl {
return (int)(totBitCount/64 + ((totBitCount % 64 == 0 ) ? 0:1));
}
private void updateCached() {
readMasks = MASKS[bitsPerValue];
shifts = SHIFTS[bitsPerValue];
writeMasks = WRITE_MASKS[bitsPerValue];
maxPos = (int)((((long)blocks.length) * BLOCK_SIZE / bitsPerValue) - 2);
}
/**
* @param index the position of the value.
* @return the value at the given index.
*/
@Override
public long get(final int index) {
assert index >= 0 && index < size();
// The abstract index in a bit stream
final long majorBitPos = (long)index * bitsPerValue;
final int elementPos = (int)(majorBitPos >>> BLOCK_BITS); // / BLOCK_SIZE
final int bitPos = (int)(majorBitPos & MOD_MASK); // % BLOCK_SIZE);
// The index in the backing long-array
final int elementPos = (int)(majorBitPos >>> BLOCK_BITS);
// The number of value-bits in the second long
final long endBits = (majorBitPos & MOD_MASK) + bpvMinusBlockSize;
final int base = bitPos * FAC_BITPOS;
assert elementPos < blocks.length : "elementPos: " + elementPos + "; blocks.len: " + blocks.length;
return ((blocks[elementPos] << shifts[base]) >>> shifts[base+1]) |
((blocks[elementPos+1] >>> shifts[base+2]) & readMasks[bitPos]);
if (endBits <= 0) { // Single block
return (blocks[elementPos] >>> -endBits) & maskRight;
}
// Two blocks
return ((blocks[elementPos] << endBits)
| (blocks[elementPos+1] >>> (BLOCK_SIZE - endBits)))
& maskRight;
}
@Override
public void set(final int index, final long value) {
// The abstract index in a contiguous bit stream
final long majorBitPos = (long)index * bitsPerValue;
// The index in the backing long-array
final int elementPos = (int)(majorBitPos >>> BLOCK_BITS); // / BLOCK_SIZE
final int bitPos = (int)(majorBitPos & MOD_MASK); // % BLOCK_SIZE);
final int base = bitPos * FAC_BITPOS;
// The number of value-bits in the second long
final long endBits = (majorBitPos & MOD_MASK) + bpvMinusBlockSize;
blocks[elementPos ] = (blocks[elementPos ] & writeMasks[base])
| (value << shifts[base + 1] >>> shifts[base]);
blocks[elementPos+1] = (blocks[elementPos+1] & writeMasks[base+1])
| ((value << shifts[base + 2]) & writeMasks[base+2]);
if (endBits <= 0) { // Single block
blocks[elementPos] = blocks[elementPos] & ~(maskRight << -endBits)
| (value << -endBits);
return;
}
// Two blocks
blocks[elementPos] = blocks[elementPos] & ~(maskRight >>> endBits)
| (value >>> endBits);
blocks[elementPos+1] = blocks[elementPos+1] & (~0L >>> endBits)
| (value << (BLOCK_SIZE - endBits));
}
@Override
public String toString() {
return "Packed64(bitsPerValue=" + bitsPerValue + ", size="
+ size() + ", maxPos=" + maxPos
+ ", elements.length=" + blocks.length + ")";
+ size() + ", elements.length=" + blocks.length + ")";
}
@Override
public long ramBytesUsed() {
return RamUsageEstimator.sizeOf(blocks);
}
@ -226,7 +177,7 @@ class Packed64 extends PackedInts.MutableImpl {
assert fromIndex <= toIndex;
// minimum number of values that use an exact number of full blocks
final int nAlignedValues = 64 / pgcd(64, bitsPerValue);
final int nAlignedValues = 64 / gcd(64, bitsPerValue);
final int span = toIndex - fromIndex;
if (span <= 3 * nAlignedValues) {
// there needs be at least 2 * nAlignedValues aligned values for the
@ -270,6 +221,17 @@ class Packed64 extends PackedInts.MutableImpl {
}
}
private static int gcd(int a, int b) {
if (a < b) {
return gcd(b, a);
} else if (b == 0) {
return a;
} else {
return gcd(b, a % b);
}
}
@Override
public void clear() {
Arrays.fill(blocks, 0L);
}