Develop Reference

Generate CRC in Java and C [closed]

Closed. This question needs debugging details. It is not currently accepting answers.
Edit the question to include desired behavior, a specific problem or error, and the shortest code necessary to reproduce the problem. This will help others answer the question.
Closed 7 years ago.
Improve this question
Here my code in C.
unsigned int crc32b(unsigned char *message) {
int i, j;
unsigned int byte, crc, mask;
i = 0;
crc = 0xFFFFFFFF;
while (message[i] != 0) {
byte = message[i]; // Get next byte.
crc = crc ^ byte;
for (j = 7; j >= 0; j--) { // Do eight times.
mask = -(crc & 1);
crc = (crc >> 1) ^ (0xEDB88320 & mask);
}
i = i + 1;
}
return ~crc;
}
But in Java, unsign int is not supported. so I try to get unsign by &0xFFFFFF
But and it wrong. How to fix it in java

You can easily emulate most unsigned arithmetic with signed integers in Java (some things are more annoying, such as division, but it is not used here).
This in particular is very easy, 0xFFFFFFFF is just itself, also known as -1.
The >> should be replaced by its unsigned counterpart, >>>.
And the byte from the message has to be masked to undo the sign-extension.
In total (not tested)
int crc32b(byte[] message) {
int i, j;
int b, crc, mask;
i = 0;
crc = -1;
while (i < message.length) {
b = message[i] & 0xFF; // Get next byte.
crc = crc ^ b;
for (j = 7; j >= 0; j--) { // Do eight times.
mask = -(crc & 1);
crc = (crc >>> 1) ^ (0xEDB88320 & mask);
}
i = i + 1;
}
return ~crc;
}
That's still quite C-ish though.

How to get CRC64 distributed calculation (use its linearity property)?

I need hash over pretty large files which is stored on distributed FS. I'm able to process parts of file with much more better performance than whole file so I'd like to be able to calculate hash over parts and then sum it.
I'm thinking about CRC64 as hashing algorithm but I have no clue how to use its theoretical 'linear function' property so I can sum CRC over parts of file. Any recommendation? Anything I missed here?
Additional notes why I'm looking at CRC64:
I can control file blocks but because of application nature they need to have different size (up to 1 byte, no any fixed blocks are possible).
I know about CRC32 implementation (zlib) which includes way to sum CRC over parts but I'd like something more wider. 8 bytes look nice for me.
I know CRC is pretty fast. I'd like to get profit from this as file can be really huge (up to few Gb).

Decided that this was generally useful enough to write and make available:
/* crc64.c -- compute CRC-64
* Copyright (C) 2013 Mark Adler
* Version 1.4 16 Dec 2013 Mark Adler
*/
/*
This software is provided 'as-is', without any express or implied
warranty. In no event will the author be held liable for any damages
arising from the use of this software.
Permission is granted to anyone to use this software for any purpose,
including commercial applications, and to alter it and redistribute it
freely, subject to the following restrictions:
1. The origin of this software must not be misrepresented; you must not
claim that you wrote the original software. If you use this software
in a product, an acknowledgment in the product documentation would be
appreciated but is not required.
2. Altered source versions must be plainly marked as such, and must not be
misrepresented as being the original software.
3. This notice may not be removed or altered from any source distribution.
Mark Adler
madler#alumni.caltech.edu
*/
/* Compute CRC-64 in the manner of xz, using the ECMA-182 polynomial,
bit-reversed, with one's complement pre and post processing. Provide a
means to combine separately computed CRC-64's. */
/* Version history:
1.0 13 Dec 2013 First version
1.1 13 Dec 2013 Fix comments in test code
1.2 14 Dec 2013 Determine endianess at run time
1.3 15 Dec 2013 Add eight-byte processing for big endian as well
Make use of the pthread library optional
1.4 16 Dec 2013 Make once variable volatile for limited thread protection
*/
#include <stdio.h>
#include <inttypes.h>
#include <assert.h>
/* The include of pthread.h below can be commented out in order to not use the
pthread library for table initialization. In that case, the initialization
will not be thread-safe. That's fine, so long as it can be assured that
there is only one thread using crc64(). */
#include <pthread.h> /* link with -lpthread */
/* 64-bit CRC polynomial with these coefficients, but reversed:
64, 62, 57, 55, 54, 53, 52, 47, 46, 45, 40, 39, 38, 37, 35, 33, 32,
31, 29, 27, 24, 23, 22, 21, 19, 17, 13, 12, 10, 9, 7, 4, 1, 0 */
#define POLY UINT64_C(0xc96c5795d7870f42)
/* Tables for CRC calculation -- filled in by initialization functions that are
called once. These could be replaced by constant tables generated in the
same way. There are two tables, one for each endianess. Since these are
static, i.e. local, one should be compiled out of existence if the compiler
can evaluate the endianess check in crc64() at compile time. */
static uint64_t crc64_little_table[8][256];
static uint64_t crc64_big_table[8][256];
/* Fill in the CRC-64 constants table. */
static void crc64_init(uint64_t table[][256])
{
unsigned n, k;
uint64_t crc;
/* generate CRC-64's for all single byte sequences */
for (n = 0; n < 256; n++) {
crc = n;
for (k = 0; k < 8; k++)
crc = crc & 1 ? POLY ^ (crc >> 1) : crc >> 1;
table[0][n] = crc;
}
/* generate CRC-64's for those followed by 1 to 7 zeros */
for (n = 0; n < 256; n++) {
crc = table[0][n];
for (k = 1; k < 8; k++) {
crc = table[0][crc & 0xff] ^ (crc >> 8);
table[k][n] = crc;
}
}
}
/* This function is called once to initialize the CRC-64 table for use on a
little-endian architecture. */
static void crc64_little_init(void)
{
crc64_init(crc64_little_table);
}
/* Reverse the bytes in a 64-bit word. */
static inline uint64_t rev8(uint64_t a)
{
uint64_t m;
m = UINT64_C(0xff00ff00ff00ff);
a = ((a >> 8) & m) | (a & m) << 8;
m = UINT64_C(0xffff0000ffff);
a = ((a >> 16) & m) | (a & m) << 16;
return a >> 32 | a << 32;
}
/* This function is called once to initialize the CRC-64 table for use on a
big-endian architecture. */
static void crc64_big_init(void)
{
unsigned k, n;
crc64_init(crc64_big_table);
for (k = 0; k < 8; k++)
for (n = 0; n < 256; n++)
crc64_big_table[k][n] = rev8(crc64_big_table[k][n]);
}
/* Run the init() function exactly once. If pthread.h is not included, then
this macro will use a simple static state variable for the purpose, which is
not thread-safe. The init function must be of the type void init(void). */
#ifdef PTHREAD_ONCE_INIT
# define ONCE(init) \
do { \
static pthread_once_t once = PTHREAD_ONCE_INIT; \
pthread_once(&once, init); \
} while (0)
#else
# define ONCE(init) \
do { \
static volatile int once = 1; \
if (once) { \
if (once++ == 1) { \
init(); \
once = 0; \
} \
else \
while (once) \
; \
} \
} while (0)
#endif
/* Calculate a CRC-64 eight bytes at a time on a little-endian architecture. */
static inline uint64_t crc64_little(uint64_t crc, void *buf, size_t len)
{
unsigned char *next = buf;
ONCE(crc64_little_init);
crc = ~crc;
while (len && ((uintptr_t)next & 7) != 0) {
crc = crc64_little_table[0][(crc ^ *next++) & 0xff] ^ (crc >> 8);
len--;
}
while (len >= 8) {
crc ^= *(uint64_t *)next;
crc = crc64_little_table[7][crc & 0xff] ^
crc64_little_table[6][(crc >> 8) & 0xff] ^
crc64_little_table[5][(crc >> 16) & 0xff] ^
crc64_little_table[4][(crc >> 24) & 0xff] ^
crc64_little_table[3][(crc >> 32) & 0xff] ^
crc64_little_table[2][(crc >> 40) & 0xff] ^
crc64_little_table[1][(crc >> 48) & 0xff] ^
crc64_little_table[0][crc >> 56];
next += 8;
len -= 8;
}
while (len) {
crc = crc64_little_table[0][(crc ^ *next++) & 0xff] ^ (crc >> 8);
len--;
}
return ~crc;
}
/* Calculate a CRC-64 eight bytes at a time on a big-endian architecture. */
static inline uint64_t crc64_big(uint64_t crc, void *buf, size_t len)
{
unsigned char *next = buf;
ONCE(crc64_big_init);
crc = ~rev8(crc);
while (len && ((uintptr_t)next & 7) != 0) {
crc = crc64_big_table[0][(crc >> 56) ^ *next++] ^ (crc << 8);
len--;
}
while (len >= 8) {
crc ^= *(uint64_t *)next;
crc = crc64_big_table[0][crc & 0xff] ^
crc64_big_table[1][(crc >> 8) & 0xff] ^
crc64_big_table[2][(crc >> 16) & 0xff] ^
crc64_big_table[3][(crc >> 24) & 0xff] ^
crc64_big_table[4][(crc >> 32) & 0xff] ^
crc64_big_table[5][(crc >> 40) & 0xff] ^
crc64_big_table[6][(crc >> 48) & 0xff] ^
crc64_big_table[7][crc >> 56];
next += 8;
len -= 8;
}
while (len) {
crc = crc64_big_table[0][(crc >> 56) ^ *next++] ^ (crc << 8);
len--;
}
return ~rev8(crc);
}
/* Return the CRC-64 of buf[0..len-1] with initial crc, processing eight bytes
at a time. This selects one of two routines depending on the endianess of
the architecture. A good optimizing compiler will determine the endianess
at compile time if it can, and get rid of the unused code and table. If the
endianess can be changed at run time, then this code will handle that as
well, initializing and using two tables, if called upon to do so. */
uint64_t crc64(uint64_t crc, void *buf, size_t len)
{
uint64_t n = 1;
return *(char *)&n ? crc64_little(crc, buf, len) :
crc64_big(crc, buf, len);
}
#define GF2_DIM 64 /* dimension of GF(2) vectors (length of CRC) */
static uint64_t gf2_matrix_times(uint64_t *mat, uint64_t vec)
{
uint64_t sum;
sum = 0;
while (vec) {
if (vec & 1)
sum ^= *mat;
vec >>= 1;
mat++;
}
return sum;
}
static void gf2_matrix_square(uint64_t *square, uint64_t *mat)
{
unsigned n;
for (n = 0; n < GF2_DIM; n++)
square[n] = gf2_matrix_times(mat, mat[n]);
}
/* Return the CRC-64 of two sequential blocks, where crc1 is the CRC-64 of the
first block, crc2 is the CRC-64 of the second block, and len2 is the length
of the second block. */
uint64_t crc64_combine(uint64_t crc1, uint64_t crc2, uintmax_t len2)
{
unsigned n;
uint64_t row;
uint64_t even[GF2_DIM]; /* even-power-of-two zeros operator */
uint64_t odd[GF2_DIM]; /* odd-power-of-two zeros operator */
/* degenerate case */
if (len2 == 0)
return crc1;
/* put operator for one zero bit in odd */
odd[0] = POLY; /* CRC-64 polynomial */
row = 1;
for (n = 1; n < GF2_DIM; n++) {
odd[n] = row;
row <<= 1;
}
/* put operator for two zero bits in even */
gf2_matrix_square(even, odd);
/* put operator for four zero bits in odd */
gf2_matrix_square(odd, even);
/* apply len2 zeros to crc1 (first square will put the operator for one
zero byte, eight zero bits, in even) */
do {
/* apply zeros operator for this bit of len2 */
gf2_matrix_square(even, odd);
if (len2 & 1)
crc1 = gf2_matrix_times(even, crc1);
len2 >>= 1;
/* if no more bits set, then done */
if (len2 == 0)
break;
/* another iteration of the loop with odd and even swapped */
gf2_matrix_square(odd, even);
if (len2 & 1)
crc1 = gf2_matrix_times(odd, crc1);
len2 >>= 1;
/* if no more bits set, then done */
} while (len2 != 0);
/* return combined crc */
crc1 ^= crc2;
return crc1;
}
/* Test crc64() on vector[0..len-1] which should have CRC-64 crc. Also test
crc64_combine() on vector[] split in two. */
static void crc64_test(void *vector, size_t len, uint64_t crc)
{
uint64_t crc1, crc2;
/* test crc64() */
crc1 = crc64(0, vector, len);
if (crc1 ^ crc)
printf("mismatch: %" PRIx64 ", should be %" PRIx64 "\n", crc1, crc);
/* test crc64_combine() */
crc1 = crc64(0, vector, (len + 1) >> 1);
crc2 = crc64(0, vector + ((len + 1) >> 1), len >> 1);
crc1 = crc64_combine(crc1, crc2, len >> 1);
if (crc1 ^ crc)
printf("mismatch: %" PRIx64 ", should be %" PRIx64 "\n", crc1, crc);
}
/* Test vectors. */
#define TEST1 "123456789"
#define TESTLEN1 9
#define TESTCRC1 UINT64_C(0x995dc9bbdf1939fa)
#define TEST2 "This is a test of the emergency broadcast system."
#define TESTLEN2 49
#define TESTCRC2 UINT64_C(0x27db187fc15bbc72)
int main(void)
{
crc64_test(TEST1, TESTLEN1, TESTCRC1);
crc64_test(TEST2, TESTLEN2, TESTCRC2);
return 0;
}

OK, my contribution to this. Ported to Java.
I cannot win from 8-byte blocks without doing unsafe thing so I removed block calculation.
I stay with ECMA polynom - ISO one looks too transparent as for me.
Of course in final version I will move test code under JUnit.
So here is code:
package com.test;
import java.util.Arrays;
/**
* CRC-64 implementation with ability to combine checksums calculated over different blocks of data.
**/
public class CRC64 {
private final static long POLY = (long) 0xc96c5795d7870f42L; // ECMA-182
/* CRC64 calculation table. */
private final static long[] table;
/* Current CRC value. */
private long value;
static {
table = new long[256];
for (int n = 0; n < 256; n++) {
long crc = n;
for (int k = 0; k < 8; k++) {
if ((crc & 1) == 1) {
crc = (crc >>> 1) ^ POLY;
} else {
crc = (crc >>> 1);
}
}
table[n] = crc;
}
}
public CRC64() {
this.value = 0;
}
public CRC64(long value) {
this.value = value;
}
public CRC64(byte [] b, int len) {
this.value = 0;
update(b, len);
}
/**
* Construct new CRC64 instance from byte array.
**/
public static CRC64 fromBytes(byte [] b) {
long l = 0;
for (int i = 0; i < 4; i++) {
l <<= 8;
l ^= (long) b[i] & 0xFF;
}
return new CRC64(l);
}
/**
* Get 8 byte representation of current CRC64 value.
**/
public byte[] getBytes() {
byte [] b = new byte[8];
for (int i = 0; i < 8; i++) {
b[7 - i] = (byte) (this.value >>> (i * 8));
}
return b;
}
/**
* Get long representation of current CRC64 value.
**/
public long getValue() {
return this.value;
}
/**
* Update CRC64 with new byte block.
**/
public void update(byte [] b, int len) {
int idx = 0;
this.value = ~this.value;
while (len > 0) {
this.value = table[((int) (this.value ^ b[idx])) & 0xff] ^ (this.value >>> 8);
idx++;
len--;
}
this.value = ~this.value;
}
private static final int GF2_DIM = 64; /* dimension of GF(2) vectors (length of CRC) */
private static long gf2MatrixTimes(long [] mat, long vec)
{
long sum = 0;
int idx = 0;
while (vec != 0) {
if ((vec & 1) == 1)
sum ^= mat[idx];
vec >>>= 1;
idx++;
}
return sum;
}
private static void gf2MatrixSquare(long [] square, long [] mat)
{
for (int n = 0; n < GF2_DIM; n++)
square[n] = gf2MatrixTimes(mat, mat[n]);
}
/*
* Return the CRC-64 of two sequential blocks, where summ1 is the CRC-64 of the
* first block, summ2 is the CRC-64 of the second block, and len2 is the length
* of the second block.
*/
static public CRC64 combine(CRC64 summ1, CRC64 summ2, long len2)
{
// degenerate case.
if (len2 == 0)
return new CRC64(summ1.getValue());
int n;
long row;
long [] even = new long[GF2_DIM]; // even-power-of-two zeros operator
long [] odd = new long[GF2_DIM]; // odd-power-of-two zeros operator
// put operator for one zero bit in odd
odd[0] = POLY; // CRC-64 polynomial
row = 1;
for (n = 1; n < GF2_DIM; n++) {
odd[n] = row;
row <<= 1;
}
// put operator for two zero bits in even
gf2MatrixSquare(even, odd);
// put operator for four zero bits in odd
gf2MatrixSquare(odd, even);
// apply len2 zeros to crc1 (first square will put the operator for one
// zero byte, eight zero bits, in even)
long crc1 = summ1.getValue();
long crc2 = summ2.getValue();
do {
// apply zeros operator for this bit of len2
gf2MatrixSquare(even, odd);
if ((len2 & 1) == 1)
crc1 = gf2MatrixTimes(even, crc1);
len2 >>>= 1;
// if no more bits set, then done
if (len2 == 0)
break;
// another iteration of the loop with odd and even swapped
gf2MatrixSquare(odd, even);
if ((len2 & 1) == 1)
crc1 = gf2MatrixTimes(odd, crc1);
len2 >>>= 1;
// if no more bits set, then done
} while (len2 != 0);
// return combined crc.
crc1 ^= crc2;
return new CRC64(crc1);
}
private static void test(byte [] b, int len, long crcValue) throws Exception {
/* Test CRC64 default calculation. */
CRC64 crc = new CRC64(b, len);
if (crc.getValue() != crcValue) {
throw new Exception("mismatch: " + String.format("%016x", crc.getValue())
+ " should be " + String.format("%016x", crcValue));
}
/* test combine() */
CRC64 crc1 = new CRC64(b, (len + 1) >>> 1);
CRC64 crc2 = new CRC64(Arrays.copyOfRange(b, (len + 1) >>> 1, b.length), len >>> 1);
crc = CRC64.combine(crc1, crc2, len >>> 1);
if (crc.getValue() != crcValue) {
throw new Exception("mismatch: " + String.format("%016x", crc.getValue())
+ " should be " + String.format("%016x", crcValue));
}
}
public static void main(String [] args) throws Exception {
final byte[] TEST1 = "123456789".getBytes();
final int TESTLEN1 = 9;
final long TESTCRC1 = 0x995dc9bbdf1939faL; // ECMA.
test(TEST1, TESTLEN1, TESTCRC1);
final byte[] TEST2 = "This is a test of the emergency broadcast system.".getBytes();
final int TESTLEN2 = 49;
final long TESTCRC2 = 0x27db187fc15bbc72L; // ECMA.
test(TEST2, TESTLEN2, TESTCRC2);
final byte[] TEST3 = "IHATEMATH".getBytes();
final int TESTLEN3 = 9;
final long TESTCRC3 = 0x3920e0f66b6ee0c8L; // ECMA.
test(TEST3, TESTLEN3, TESTCRC3);
}
}

What do "& 0xFF" and ">>>" shifting do?

I am trying to understand the below code.
The method getKey() returns a string, and getDistance() returns a double. The code is a taken from a class which is meant to hold String (the key) and Double (the distance) pairs.
To be more specific I am unsure as to what the lines that do the shifting do.
public void serialize (byte[] outputArray) {
// write the length of the string out
byte[] data = getKey().getBytes ();
for (int i = 0; i < 2; i++) {
outputArray[i] = (byte) ((data.length >>> ((1 - i) * 8)) & 0xFF);
}
// write the key out
for (int i = 0; i < data.length; i++) {
outputArray[i + 2] = data[i];
}
// now write the distance out
long bits = Double.doubleToLongBits (getDistance());
for (int i = 0; i < 8; i++) {
outputArray[i + 2 + data.length] = (byte) ((bits >>> ((7 - i) * 8)) & 0xFF);
}
}
Any help would be very appreciated.

>>> is unsigned shift to right operator. It shifts the sign bit too.
& 0xFF retains bits to make a 8-bit (byte) value, otherwise you may have some garbage.

Start by reading Java's tutorial on bitwise operators. In short:
>>> is an unsigned right shift
& 0xFF is ANDing the outcome of (bits >>> ((7 - i) * 8)) with 0xFF

Bit manipulation C source in Java

I try to calculate the checksum of a Sega Genesis rom file in Java. For this i want to port a code snipped from C into Java:
static uint16 getchecksum(uint8 *rom, int length)
{
int i;
uint16 checksum = 0;
for (i = 0; i < length; i += 2)
{
checksum += ((rom[i] << 8) + rom[i + 1]);
}
return checksum;
}
I understand what the code does. It sums all 16bit numbers (combined from two 8 bit ones). But what i didn't understand is what's happening with the overflow of the uint16 and how this transfers to Java code?
Edit:
This code seems to work, thanks:
int calculatedChecksum = 0;
int bufferi1=0;
int bufferi2=0;
bs = new BufferedInputStream(new FileInputStream(this.file));
bufferi1 = bs.read();
bufferi2 = bs.read();
while(bufferi1 != -1 && bufferi2 != -1){
calculatedChecksum += (bufferi1*256 + bufferi2);
calculatedChecksum = calculatedChecksum % 0x10000;
bufferi1 = bs.read();
bufferi2 = bs.read();
}

Simply put, the overflow is lost.
A more correct approach (imho) is to use uint32 for summation, and then you have the sum in the lower 16 bits, and the overflow in the upper 16 bits.

static int checksum(final InputStream in) throws IOException {
short v = 0;
int c;
while ((c = in.read()) >= 0) {
v += (c << 8) | in.read();
}
return v & 0xffff;
}
This should work equivalently; by using & 0xffff, we get to treat the value in v as if it were unsigned the entire time, since arithmetic overflow is identical w.r.t. bits.

You want addition modulo 216, which you can simply spell out manually:
checksum = (checksum + ((rom[i] << 8) + rom[i + 1])) % 0x10000;
// ^^^^^^^^^

Extract bit sequences of arbitrary length from byte[] array efficiently

I'm looking for the most efficient way of extracting (unsigned) bit sequences of arbitrary length (0 <= length <= 16) at arbitrary position. The skeleton class show how my current implementation essentially handles the problem:
public abstract class BitArray {
byte[] bytes = new byte[2048];
int bitGet;
public BitArray() {
}
public void readNextBlock(int initialBitGet, int count) {
// substitute for reading from an input stream
for (int i=(initialBitGet>>3); i<=count; ++i) {
bytes[i] = (byte) i;
}
prepareBitGet(initialBitGet, count);
}
public abstract void prepareBitGet(int initialBitGet, int count);
public abstract int getBits(int count);
static class Version0 extends BitArray {
public void prepareBitGet(int initialBitGet, int count) {
bitGet = initialBitGet;
}
public int getBits(int len) {
// intentionally gives meaningless result
bitGet += len;
return 0;
}
}
static class Version1 extends BitArray {
public void prepareBitGet(int initialBitGet, int count) {
bitGet = initialBitGet - 1;
}
public int getBits(int len) {
int byteIndex = bitGet;
bitGet = byteIndex + len;
int shift = 23 - (byteIndex & 7) - len;
int mask = (1 << len) - 1;
byteIndex >>= 3;
return (((bytes[byteIndex] << 16) |
((bytes[++byteIndex] & 0xFF) << 8) |
(bytes[++byteIndex] & 0xFF)) >> shift) & mask;
}
}
static class Version2 extends BitArray {
static final int[] mask = { 0x0, 0x1, 0x3, 0x7, 0xF, 0x1F, 0x3F, 0x7F, 0xFF,
0x1FF, 0x3FF, 0x7FF, 0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF };
public void prepareBitGet(int initialBitGet, int count) {
bitGet = initialBitGet;
}
public int getBits(int len) {
int offset = bitGet;
bitGet = offset + len;
int byteIndex = offset >> 3; // originally used /8
int bitIndex = offset & 7; // originally used %8
if ((bitIndex + len) > 16) {
return ((bytes[byteIndex] << 16 |
(bytes[byteIndex + 1] & 0xFF) << 8 |
(bytes[byteIndex + 2] & 0xFF)) >> (24 - bitIndex - len)) & mask[len];
} else if ((offset + len) > 8) {
return ((bytes[byteIndex] << 8 |
(bytes[byteIndex + 1] & 0xFF)) >> (16 - bitIndex - len)) & mask[len];
} else {
return (bytes[byteIndex] >> (8 - offset - len)) & mask[len];
}
}
}
static class Version3 extends BitArray {
int[] ints = new int[2048];
public void prepareBitGet(int initialBitGet, int count) {
bitGet = initialBitGet;
int put_i = (initialBitGet >> 3) - 1;
int get_i = put_i;
int buf;
buf = ((bytes[++get_i] & 0xFF) << 16) |
((bytes[++get_i] & 0xFF) << 8) |
(bytes[++get_i] & 0xFF);
do {
buf = (buf << 8) | (bytes[++get_i] & 0xFF);
ints[++put_i] = buf;
} while (get_i < count);
}
public int getBits(int len) {
int bit_idx = bitGet;
bitGet = bit_idx + len;
int shift = 32 - (bit_idx & 7) - len;
int mask = (1 << len) - 1;
int int_idx = bit_idx >> 3;
return (ints[int_idx] >> shift) & mask;
}
}
static class Version4 extends BitArray {
int[] ints = new int[1024];
public void prepareBitGet(int initialBitGet, int count) {
bitGet = initialBitGet;
int g = initialBitGet >> 3;
int p = (initialBitGet >> 4) - 1;
final byte[] b = bytes;
int t = (b[g] << 8) | (b[++g] & 0xFF);
final int[] i = ints;
do {
i[++p] = (t = (t << 16) | ((b[++g] & 0xFF) <<8) | (b[++g] & 0xFF));
} while (g < count);
}
public int getBits(final int len) {
final int i;
bitGet = (i = bitGet) + len;
return (ints[i >> 4] >> (32 - len - (i & 15))) & ((1 << len) - 1);
}
}
public void benchmark(String label) {
int checksum = 0;
readNextBlock(32, 1927);
long time = System.nanoTime();
for (int pass=1<<18; pass>0; --pass) {
prepareBitGet(32, 1927);
for (int i=2047; i>=0; --i) {
checksum += getBits(i & 15);
}
}
time = System.nanoTime() - time;
System.out.println(label+" took "+Math.round(time/1E6D)+" ms, checksum="+checksum);
try { // avoid having the console interfere with our next measurement
Thread.sleep(369);
} catch (InterruptedException e) {}
}
public static void main(String[] argv) {
BitArray test;
// for the sake of getting a little less influence from the OS for stable measurement
Thread.currentThread().setPriority(Thread.MAX_PRIORITY);
while (true) {
test = new Version0();
test.benchmark("no implementaion");
test = new Version1();
test.benchmark("Durandal's (original)");
test = new Version2();
test.benchmark("blitzpasta's (adapted)");
test = new Version3();
test.benchmark("MSN's (posted)");
test = new Version4();
test.benchmark("MSN's (half-buffer modification)");
System.out.println("--- next pass ---");
}
}
}
This works, but I'm looking for a more efficient solution (performance wise). The byte array is guaranteed to be relatively small, between a few bytes up to a max of ~1800 bytes. The array is read exactly once (completely) between each call to the read method. There is no need for any error checking in getBits(), such as exceeding the array etc.
It seems my initial question above isn't clear enough. A "bit sequence" of N bits forms an integer of N bits, and I need to extract those integers with minimal overhead. I have no use for strings, as the values are either used as lookup indices or are directly fed into some computation. So basically, the skeleton shown above is a real class and getBits() signature shows how the rest of the code interacts with it.
Extendet the example code into a microbenchmark, included blitzpasta's solution (fixed missing byte masking). On my old AMD box it turns out as ~11400ms vs ~38000ms. FYI: Its the divide and modulo operations that kill the performance. If you replace /8 with >>3 and %8 with &7, both solutions are pretty close to each other (jdk1.7.0ea104).
There seemed to be a bit confusion about how and what to work on. The first, original post of the example code included a read() method to indicate where and when the byte buffer was filled. This got lost when the code was turned into the microbench. I re-introduced it to make this a little clearer.
The idea is to beat all existing versions by adding another subclass of BitArray which need to implement getBits() and prepareBitGet(), the latter may be empty. Do not change the benchmarking to give your solution an advantage, the same could be done for all the existing solutions, making this a completely moot optimization! (really!!)
I added a Version0, which does nothing but increment the bitGet state. It always returns 0 to get a rough idea how big the benchmark overhead is. Its only there for comparison.
Also, an adaption on MSN's idea was added (Version3). To keep things fair and comparable for all competitors, the byte array filling is now part of the benchmark, as well as a preparatory step (see above). Originally MSN's solution did not do so well, there was lots of overhead in preparing the int[] buffer. I took the liberty of optimizing the step a little, which turned it into a fierce competitor :)
You might also find that I de-convoluted your code a little. Your getBit() could be condensed into a 3-liner, probably shaving off one or two percent. I deliberately did this to keep the code readable and because the other versions aren't as condensed as possible either (again for readability).
Conclusion (code example above update to include versions based on all applicable contributions). On my old AMD box (Sun JRE 1.6.0_21), they come out as:
V0 no implementaion took 5384 ms
V1 Durandal's (original) took 10283 ms
V2 blitzpasta's (adapted) took 12212 ms
V3 MSN's (posted) took 11030 ms
V4 MSN's (half-buffer modification) took 9700 ms
Notes: In this benchmark an average of 7.5 bits is fetched per call to getBits(), and each bit is only read once. Since V3/V4 have to pay a high initialization cost, they tend to show better runtime behavior with more, shorter fetches (and consequently worse the closer to the maximum of 16 the average fetch size gets). Still, V4 stays slightly ahead of all others in all scenarios.
In an actual application, the cache contention must be taken into account, since the extra space needed for V3/v4 may increase cache misses to a point where V0 would be a better choice. If the array is to be traversed more than once, V4 should be favored, since it fetches faster than every other and the costly initialization is amortized after the fist pass.

If you just want the unsigned bit sequence as an int.
static final int[] lookup = {0x0, 0x1, 0x3, 0x7, 0xF, 0x1F, 0x3F, 0x7F, 0xFF, 0x1FF, 0x3FF, 0x7FF, 0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF };
/*
* bytes: byte array, with the bits indexed from 0 (MSB) to (bytes.length * 8 - 1) (LSB)
* offset: index of the MSB of the bit sequence.
* len: length of bit sequence, must from range [0,16].
* Not checked for overflow
*/
static int getBitSeqAsInt(byte[] bytes, int offset, int len){
int byteIndex = offset / 8;
int bitIndex = offset % 8;
int val;
if ((bitIndex + len) > 16) {
val = ((bytes[byteIndex] << 16 | bytes[byteIndex + 1] << 8 | bytes[byteIndex + 2]) >> (24 - bitIndex - len)) & lookup[len];
} else if ((offset + len) > 8) {
val = ((bytes[byteIndex] << 8 | bytes[byteIndex + 1]) >> (16 - bitIndex - len)) & lookup[len];
} else {
val = (bytes[byteIndex] >> (8 - offset - len)) & lookup[len];
}
return val;
}
If you want it as a String (modification of Margus' answer).
static String getBitSequence(byte[] bytes, int offset, int len){
int byteIndex = offset / 8;
int bitIndex = offset % 8;
int count = 0;
StringBuilder result = new StringBuilder();
outer:
for(int i = byteIndex; i < bytes.length; ++i) {
for(int j = (1 << (7 - bitIndex)); j > 0; j >>= 1) {
if(count == len) {
break outer;
}
if((bytes[byteIndex] & j) == 0) {
result.append('0');
} else {
result.append('1');
}
++count;
}
bitIndex = 0;
}
return result.toString();
}

Well, depending on how far you want to go down the time vs. memory see-saw, you can allocate a side table of every 32-bits at every 16-bit offset and then do a mask and shift based on the 16-bit offset:
byte[] bytes = new byte[2048];
int bitGet;
unsigned int dwords[] = new unsigned int[2046];
public BitArray() {
for (int i=0; i<bytes.length; ++i) {
bytes[i] = (byte) i;
}
for (int i= 0; i<dwords.length; ++i) {
dwords[i]=
(bytes[i ] << 24) |
(bytes[i + 1] << 16) |
(bytes[i + 2] << 8) |
(bytes[i + 3]);
}
}
int getBits(int len)
{
int offset= bitGet;
int offset_index= offset>>4;
int offset_offset= offset & 15;
return (dwords[offset_index] >> offset_offset) & ((1 << len) - 1);
}
You avoid the branching (at the cost of quadrupling your memory footprint). And is looking up the mask really that much faster than (1 << len) - 1?

Just wondering why can't you use java.util.BitSet;
Basically what you can do, is to read the whole data as byte[], convert it to binary in string format and use string utilities like .substring() to do the work. This will also work bit sequences > 16.
Lets say you have 3 bytes: 1, 2, 3 and you want to extract bit sequence from 5th to 16th bit.
Number Binary
1 00000001
2 00000010
3 00000011
Code example:
public static String getRealBinary(byte[] input){
StringBuilder sb = new StringBuilder();
for (byte c : input) {
for (int n = 128; n > 0; n >>= 1){
if ((c & n) == 0)
sb.append('0');
else sb.append('1');
}
}
return sb.toString();
}
public static void main(String[] args) {
byte bytes[] = new byte[]{1,2,3};
String sbytes = getRealBinary(bytes);
System.out.println(sbytes);
System.out.println(sbytes.substring(5,16));
}
Output:
000000010000001000000011
00100000010
Speed:
I did a testrun for 1m times and on my computer it took 0.995s, so its reasonably very fast:
Code to repeat the test yourself:
public static void main(String[] args) {
Random r = new Random();
byte bytes[] = new byte[4];
long start, time, total=0;
for (int i = 0; i < 1000000; i++) {
r.nextBytes(bytes);
start = System.currentTimeMillis();
getRealBinary(bytes).substring(5,16);
time = System.currentTimeMillis() - start;
total+=time;
}
System.out.println("It took " +total + "ms");
}

You want at most 16 bits, taken from an array of bytes. 16 bits can span at most 3 bytes.
Here's a possible solution:
int GetBits(int bit_index, int bit_length) {
int byte_offset = bit_index >> 3;
return ((((((byte_array[byte_offset]<<8)
+byte_array[byte_offset+1])<<8)
+byte_array[byte_offset+2]))
>>(24-(bit_index&7)+bit_length))))
&((1<<bit_length)-1);
}
[Untested]
If you call this a lot you should precompute the 24-bit values for the 3 concatenated bytes, and store those into an int array.
I'll observe that if you are coding this in C on an x86, you don't even need to precompute the 24 bit array; simply access the by te array at the desire offset as a 32 bit value. The x86 will do unaligned fetches just fine. [commenter noted that endianess mucks this up, so it isn't an answer, OK, do the 24 bit version.]

Since Java 7 BitSet has the toLongArray method, which I believe will do exactly what the question asks for:
int subBits = (int) bitSet.get(lowBit, highBit).toLongArray()[0];
This has the advantage that it works with sequences larger than ints or longs. It has the performance disadvantage that a new BitSet object must be allocated, and a new array object to hold the result.
It would be really interesting to see how this compares with the other methods in the benchmark.