package com.kamikaze.pfordelta; import java.util.Arrays; /** * This is a version of the kamikaze PForDelta library that * was slightly cleaned up by D. Lemire. It is included in the * JavaFastPFOR library for comparison purposes. As the original */ /** * Implementation of the optimized PForDelta algorithm for sorted integer * arrays. The basic ideas are based on * * 1. Original algorithm from * http://homepages.cwi.nl/~heman/downloads/msthesis.pdf * * 2. Optimization and variation from * http://www2008.org/papers/pdf/p387-zhangA.pdf * * 3. Further optimization http://www2009.org/proceedings/pdf/p401.pdf * * As a part of the PForDelta implementation, Simple16 is used to compress * exceptions. The original Simple16 algorithm can also be found in the above * literatures. * * This implementation overcomes the problem that Simple16 cannot deal with * more than 2^28 numbers. * * Author: hao yan hyan2008@gmail.com */ public class PForDelta { // NOTE: we expect the blockSize is always < (1<<(31-POSSIBLE_B_BITS)). // For example, in the current default settings, // the blockSize < (1<<(31-5)), that is, < 2^27, the commonly used block // size is 128 or 256. // All possible values of b in the PForDelta algorithm private static final int[] POSSIBLE_B = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 16, 20, 28 }; // Max number of bits to store an uncompressed value private static final int MAX_BITS = 32; // Header records the value of b and the number of exceptions in the // block private static final int HEADER_NUM = 2; // Header size in bits private static final int HEADER_SIZE = MAX_BITS * HEADER_NUM; private static final int[] MASK = { 0x00000000, 0x00000001, 0x00000003, 0x00000007, 0x0000000f, 0x0000001f, 0x0000003f, 0x0000007f, 0x000000ff, 0x000001ff, 0x000003ff, 0x000007ff, 0x00000fff, 0x00001fff, 0x00003fff, 0x00007fff, 0x0000ffff, 0x0001ffff, 0x0003ffff, 0x0007ffff, 0x000fffff, 0x001fffff, 0x003fffff, 0x007fffff, 0x00ffffff, 0x01ffffff, 0x03ffffff, 0x07ffffff, 0x0fffffff, 0x1fffffff, 0x3fffffff, 0x7fffffff, 0xffffffff }; /** * Compress one block of blockSize integers using PForDelta with the * optimal parameter b * * @param inBlock * the block to be compressed * @param blockSize * the block size * @return the compressed block */ public static int[] compressOneBlockOpt(final int[] inBlock, int blockSize) { // find the best b that may lead to the smallest overall // compressed size int currentB = POSSIBLE_B[0]; int[] outBlock = null; int tmpB = currentB; // deal with the large exception cases boolean hasBigNum = checkBigNumbers(inBlock, POSSIBLE_B[POSSIBLE_B.length - 1], blockSize); if (hasBigNum) { currentB = 4; System.out.println("has big num and the currentB is: " + currentB); } else { int optSize = estimateCompressedSize(inBlock, tmpB, blockSize); for (int i = 1; i < POSSIBLE_B.length; ++i) { tmpB = POSSIBLE_B[i]; int curSize = estimateCompressedSize(inBlock, tmpB, blockSize); if (curSize < optSize) { currentB = tmpB; optSize = curSize; } } } // compress the block using the above best b outBlock = compressOneBlock(inBlock, currentB, blockSize); return outBlock; } /** * Decompress one block using PForDelta * * @param outBlock * the block that was decompressed * @param inBlock * the block to be decompressed * @param blockSize * the number of elements in the decompressed block * @return the compressed size in bits */ public static int decompressOneBlock(int[] outBlock, int[] inBlock, int blockSize) { int[] expAux = new int[blockSize * 2]; int expNum = inBlock[0] & 0x3ff; int bits = (inBlock[0] >>> 10) & (0x1f); // decompress the b-bit slots int offset = HEADER_SIZE; int compressedBits = 0; if (bits == 0) { Arrays.fill(outBlock, 0); } else { compressedBits = decompressBBitSlots(outBlock, inBlock, blockSize, bits); // Note that blocksize must be ==128 in order to use // decompressBBitSlotsWithHardCodes // compressedBits = // decompressBBitSlotsWithHardCodes(outBlock, inBlock, // blockSize, bits); } offset += compressedBits; // decompress exceptions if (expNum > 0) { compressedBits = decompressBlockByS16(expAux, inBlock, offset, expNum * 2); offset += compressedBits; for (int i = 0; i < expNum; i++) { int curExpPos = expAux[i]; int curHighBits = expAux[i + expNum]; outBlock[curExpPos] = (outBlock[curExpPos] & MASK[bits]) | ((curHighBits & MASK[32 - bits]) << bits); } } return offset; } /** * Estimate the compressed size in ints of a block * * @param inputBlock * the block to be compressed * @param bits * the value of the parameter b * @param blockSize * the block size * @return the compressed size in ints */ public static int estimateCompressedSize(int[] inputBlock, int bits, int blockSize) { int maxNoExp = (1 << bits) - 1; // Size of the header and the bits-bit slots int outputOffset = HEADER_SIZE + bits * blockSize; int expNum = 0; for (int i = 0; i < blockSize; ++i) { if (inputBlock[i] > maxNoExp) { expNum++; } } outputOffset += (expNum << 5); return outputOffset; } /** * Check if the block contains big numbers that is greater than ((1<< * bits)-1) * * @param inputBlock * the block to be compressed * @param bits * the numbers of bits to decide whether a number is a * big number * @param blockSize * the block size * @return true if there is any big numbers in the block */ public static boolean checkBigNumbers(int[] inputBlock, int bits, int blockSize) { int maxNoExp = (1 << bits) - 1; for (int i = 0; i < blockSize; ++i) { if (inputBlock[i] > maxNoExp) return true; } return false; } /** * The core implementation of compressing a block with blockSize * integers using PForDelta with the given parameter b * * @param inputBlock * the block to be compressed * @param bits * the the value of the parameter b * @param blockSize * the block size * @return the compressed block */ public static int[] compressOneBlock(int[] inputBlock, int bits, int blockSize) { int[] expAux = new int[blockSize * 2]; int maxCompBitSize = HEADER_SIZE + blockSize * (MAX_BITS + MAX_BITS + MAX_BITS) + 32; int[] tmpCompressedBlock = new int[(maxCompBitSize >>> 5)]; int outputOffset = HEADER_SIZE; int expUpperBound = 1 << bits; int expNum = 0; for (int elem : inputBlock) { if (elem >= expUpperBound) { expNum++; } } int expIndex = 0; // compress the b-bit slots for (int i = 0; i < blockSize; ++i) { if (inputBlock[i] < expUpperBound) { writeBits(tmpCompressedBlock, inputBlock[i], outputOffset, bits); } else // exp { // store the lower bits-bits of the exception writeBits(tmpCompressedBlock, inputBlock[i] & MASK[bits], outputOffset, bits); // write the position of exception expAux[expIndex] = i; // write the higher 32-bits bits of the // exception expAux[expIndex + expNum] = (inputBlock[i] >>> bits) & MASK[32 - bits]; expIndex++; } outputOffset += bits; } // the first int in the compressed block stores the value of b // and the number of exceptions // tmpCompressedBlock[0] = ((bits & MASK[POSSIBLE_B_BITS]) << // (31-POSSIBLE_B_BITS)) | (expNum & MASK[31-POSSIBLE_B_BITS]); tmpCompressedBlock[0] = ((bits & MASK[10]) << 10) | (expNum & 0x3ff); tmpCompressedBlock[1] = inputBlock[blockSize - 1]; // compress exceptions if (expNum > 0) { int compressedBitSize = compressBlockByS16( tmpCompressedBlock, outputOffset, expAux, expNum * 2); outputOffset += compressedBitSize; } // discard the redundant parts in the tmpCompressedBlock int compressedSizeInInts = (outputOffset + 31) >>> 5; int[] compBlock; compBlock = new int[compressedSizeInInts]; System.arraycopy(tmpCompressedBlock, 0, compBlock, 0, compressedSizeInInts); return compBlock; } /** * Decompress b-bit slots * * @param outDecompSlots * decompressed block which is the output * @param inCompBlock * the compressed block which is the input * @param blockSize * the block size * @param bits * the value of the parameter b * @return the compressed size in bits of the data that has been * decompressed */ public static int decompressBBitSlots(int[] outDecompSlots, int[] inCompBlock, int blockSize, int bits) { int compressedBitSize = 0; int offset = HEADER_SIZE; for (int i = 0; i < blockSize; i++) { outDecompSlots[i] = readBits(inCompBlock, offset, bits); offset += bits; } compressedBitSize = bits * blockSize; return compressedBitSize; } /** * Compress a block of blockSize integers using Simple16 algorithm * * @param outCompBlock * the compressed block which is the output * @param outStartOffsetInBits * the start offset in bits of the compressed block * @param inBlock * the block to be compressed * @param blockSize * the block size * @return the compressed size in bits */ private static int compressBlockByS16(int[] outCompBlock, int outStartOffsetInBits, int[] inBlock, int blockSize) { int outOffset = (outStartOffsetInBits + 31) >>> 5; int num, inOffset = 0, numLeft; for (numLeft = blockSize; numLeft > 0; numLeft -= num) { num = Simple16.s16Compress(outCompBlock, outOffset, inBlock, inOffset, numLeft, blockSize); outOffset++; inOffset += num; } int compressedBitSize = (outOffset << 5) - outStartOffsetInBits; return compressedBitSize; } /** * Decompress a block of blockSize integers using Simple16 algorithm * * @param outDecompBlock * the decompressed block which is the output * @param inCompBlock * the compressed block which is the input * @param blockSize * the block size * @param inStartOffsetInBits * the start offset in bits of the compressed block * @return the compressed size in bits of the data that has been * decompressed */ public static int decompressBlockByS16(int[] outDecompBlock, int[] inCompBlock, int inStartOffsetInBits, int blockSize) { int inOffset = (inStartOffsetInBits + 31) >>> 5; int num, outOffset = 0, numLeft; for (numLeft = blockSize; numLeft > 0; numLeft -= num) { num = Simple16.s16Decompress(outDecompBlock, outOffset, inCompBlock, inOffset, numLeft); outOffset += num; inOffset++; } int compressedBitSize = (inOffset << 5) - inStartOffsetInBits; return compressedBitSize; } /** * Write a certain number of bits of an integer into an integer array * starting from the given start offset * * @param out * the output array * @param val * the integer to be written * @param outOffset * the start offset in bits in the output array * @param bits * the number of bits to be written (bits greater or equal to 0) */ public static final void writeBits(int[] out, int val, int outOffset, int bits) { if (bits == 0) return; final int index = outOffset >>> 5; final int skip = outOffset & 0x1f; val &= (0xffffffff >>> (32 - bits)); out[index] |= (val << skip); if (32 - skip < bits) { out[index + 1] |= (val >>> (32 - skip)); } } /** * Decompress the b-bit slots using hardcoded unpack methods * * @param decompressedSlots * the decompressed output * @param compBlock * the compressed input block * @param blockSize * the block size which is 256 by default * @param bits * the value of b * @return the processed data size (the number of bits in the compressed * form) */ static int decompressBBitSlotsWithHardCodes(int[] decompressedSlots, int[] compBlock, int blockSize, int bits) { int compressedBitSize = 0; PForDeltaUnpack128.unpack(decompressedSlots, compBlock, bits); compressedBitSize = bits * blockSize; return compressedBitSize; } /** * Read a certain number of bits of an integer into an integer array * starting from the given start offset * * @param in * the input array * @param inOffset * the start offset in bits in the input array * @param bits * the number of bits to be read, unlike writeBits(), * readBits() does not deal with bits==0 and thus bits * must be greater than 0. When bits ==0, the calling functions will * just skip the entire bits-bit slots without decoding * them * @return the bits bits of the input */ public static final int readBits(int[] in, final int inOffset, final int bits) { final int index = inOffset >>> 5; final int skip = inOffset & 0x1f; int val = in[index] >>> skip; if (32 - skip < bits) { val |= (in[index + 1] << (32 - skip)); } return val & (0xffffffff >>> (32 - bits)); } }