Gather bits at specific positions into a new value - c++

I have a bit-mask of N chars in size, which is statically known (i.e. can be calculated at compile time, but it's not a single constant, so I can't just write it down), with bits set to 1 denoting the "wanted" bits. And I have a value of the same size, which is only known at runtime. I want to collect the "wanted" bits from that value, in order, into the beginning of a new value. For simplicity's sake let's assume the number of wanted bits is <= 32.
Completely unoptimized reference code which hopefully has the correct behaviour:
template<int N, const char mask[N]>
unsigned gather_bits(const char* val)
{
unsigned result = 0;
char* result_p = (char*)&result;
int pos = 0;
for (int i = 0; i < N * CHAR_BIT; i++)
{
if (mask[i/CHAR_BIT] & (1 << (i % CHAR_BIT)))
{
if (val[i/CHAR_BIT] & (1 << (i % CHAR_BIT)))
{
if (pos < sizeof(unsigned) * CHAR_BIT)
{
result_p[pos/CHAR_BIT] |= 1 << (pos % CHAR_BIT);
}
else
{
abort();
}
}
pos += 1;
}
}
return result;
}
Although I'm not sure whether that formulation actually allows access to the contents of the mask at compile time. But in any case, it's available for use, maybe a constexpr function or something would be a better idea. I'm not looking here for the necessary C++ wizardry (I'll figure that out), just the algorithm.
An example of input/output, with 16-bit values and imaginary binary notation for clarity:
mask = 0b0011011100100110
val = 0b0101000101110011
--
wanted = 0b__01_001__1__01_ // retain only those bits which are set in the mask
result = 0b0000000001001101 // bring them to the front
^ gathered bits begin here
My questions are:
What's the most performant way to do this? (Are there any hardware instructions that can help?)
What if both the mask and the value are restricted to be unsigned, so a single word, instead of an unbounded char array? Can it then be done with a fixed, short sequence of instructions?

There will pext (parallel bit extract) that does exactly what you want in Intel Haswell. I don't know what the performance of that instruction will be, probably better than the alternatives though. This operation is also known as "compress-right" or simply "compress", the implementation from Hacker's Delight is this:
unsigned compress(unsigned x, unsigned m) {
unsigned mk, mp, mv, t;
int i;
x = x & m; // Clear irrelevant bits.
mk = ~m << 1; // We will count 0's to right.
for (i = 0; i < 5; i++) {
mp = mk ^ (mk << 1); // Parallel prefix.
mp = mp ^ (mp << 2);
mp = mp ^ (mp << 4);
mp = mp ^ (mp << 8);
mp = mp ^ (mp << 16);
mv = mp & m; // Bits to move.
m = m ^ mv | (mv >> (1 << i)); // Compress m.
t = x & mv;
x = x ^ t | (t >> (1 << i)); // Compress x.
mk = mk & ~mp;
}
return x;
}

Related

Asynchronous writing to a bit array

TL; DR How to safely perfom a single bit update A[n/8] |= (1<<n%8); for A being a huge array of chars (i.e., setting n's bit of A true) when computing in parallel using C++11's <thread> library?
I'm performing a computation that's easy to parallelize. I'm computing elements of a certain subset of the natural numbers, and I wanna find elements that are not in the subset. For this I create a huge array (like A = new char[20l*1024l*1024l*1024l], i.e., 20GiB). A n's bit of this array is true if n lies in my set.
When doing it in parallel and setting the bits true using A[n/8] |= (1<<n%8);, I seem to get a small loss of information, supposedly due to concurring work on the same byte of A (each thread has to first read the byte, update the single bit and write the byte back). How can I get around this? Is there a way how to do this update as an atomic operation?
The code follows. GCC version: g++ (Ubuntu 5.4.0-6ubuntu1~16.04.11) 5.4.0 20160609. The machine is an 8-core Intel(R) Xeon(R) CPU E5620 # 2.40GHz, 37GB RAM. Compiler options: g++ -std=c++11 -pthread -O3
#include <iostream>
#include <thread>
typedef long long myint; // long long to be sure
const myint max_A = 20ll*1024ll*1024ll; // 20 MiB for testing
//const myint max_A = 20ll*1024ll*1024ll*1024ll; // 20 GiB in the real code
const myint n_threads = 1; // Number of threads
const myint prime = 1543; // Tested prime
char *A;
const myint max_n = 8*max_A;
inline char getA(myint n) { return A[n/8] & (1<<(n%8)); }
inline void setAtrue(myint n) { A[n/8] |= (1<<n%8); }
void run_thread(myint startpoint) {
// Calculate all values of x^2 + 2y^2 + prime*z^2 up to max_n
// We loop through x == startpoint (mod n_threads)
for(myint x = startpoint; 1*x*x < max_n; x+=n_threads)
for(myint y = 0; 1*x*x + 2*y*y < max_n; y++)
for(myint z = 0; 1*x*x + 2*y*y + prime*z*z < max_n; z++)
setAtrue(1*x*x + 2*y*y + prime*z*z);
}
int main() {
myint n;
// Only n_threads-1 threads, as we will use the master thread as well
std::thread T[n_threads-1];
// Initialize the array
A = new char[max_A]();
// Start the threads
for(n = 0; n < n_threads-1; n++) T[n] = std::thread(run_thread, n);
// We use also the master thread
run_thread(n_threads-1);
// Synchronize
for(n = 0; n < n_threads-1; n++) T[n].join();
// Print and count all elements not in the set and n != 0 (mod prime)
myint cnt = 0;
for(n=0; n<max_n; n++) if(( !getA(n) )&&( n%1543 != 0 )) {
std::cout << n << std::endl;
cnt++;
}
std::cout << "cnt = " << cnt << std::endl;
return 0;
}
When n_threads = 1, I get the correct value cnt = 29289. When n_threads = 7, I got cnt = 29314 and cnt = 29321 on two different calls, suggesting some of the bitwise operations on a single byte were concurring.
std::atomic provides all the facilities that you need here:
std::array<std::atomic<char>, max_A> A;
static_assert(sizeof(A[0]) == 1, "Shall not have memory overhead");
static_assert(std::atomic<char>::is_always_lock_free,
"No software-level locking needed on common platforms");
inline char getA(myint n) { return A[n / 8] & (1 << (n % 8)); }
inline void setAtrue(myint n) { A[n / 8].fetch_or(1 << n % 8); }
The load in getA is atomic (equivalent to load()), and std::atomic even has built-in support for oring the stored value with another one (fetch_or), atomically of course.
When initializing A, the naive way of for (auto& a : A) a = 0; would require synchronization after every store, which you can avoid by waiving some thread-safety. std::memory_order_release only requires that what we write is visible to other threads (but not that other thread's writes are visible to us). And indeed, if you do
// Initialize the array
for (auto& a : A)
a.store(0, std::memory_order_release);
you get the safety you need without any assembly-level synchronization on x86. You could do the reverse for the loads after the threads finish, but that has no added benefit on x86 (it's just a mov either way).
Demo on the full code: https://godbolt.org/z/nLPlv1

Generate all combinations in bit version

I'd like to generate all possible combination (without repetitions) in bit representation. I can't use any library like boost or stl::next_combination - it has to be my own code (computation time is very important).
Here's my code (modified from ones StackOverflow user):
int combination = (1 << k) - 1;
int new_combination = 0;
int change = 0;
while (true)
{
// return next combination
cout << combination << endl;
// find first index to update
int indexToUpdate = k;
while (indexToUpdate > 0 && GetBitPositionByNr(combination, indexToUpdate)>= n - k + indexToUpdate)
indexToUpdate--;
if (indexToUpdate == 1) change = 1; // move all bites to the left by one position
if (indexToUpdate <= 0) break; // done
// update combination indices
new_combination = 0;
for (int combIndex = GetBitPositionByNr(combination, indexToUpdate) - 1; indexToUpdate <= k; indexToUpdate++, combIndex++)
{
if(change)
{
new_combination |= (1 << (combIndex + 1));
}
else
{
combination = combination & (~(1 << combIndex));
combination |= (1 << (combIndex + 1));
}
}
if(change) combination = new_combination;
change = 0;
}
where n - all elements, k - number of elements in combination.
GetBitPositionByNr - return position of k-th bit.
GetBitPositionByNr(13,2) = 3 cause 13 is 1101 and second bit is on third position.
It gives me correct output for n=4, k=2 which is:
0011 (3 - decimal representation - printed value)
0101 (5)
1001 (9)
0110 (6)
1010 (10)
1100 (12)
Also it gives me correct output for k=1 and k=4, but gives me wrong outpu for k=3 which is:
0111 (7)
1011 (11)
1011 (9) - wrong, should be 13
1110 (14)
I guess the problem is in inner while condition (second) but I don't know how to fix this.
Maybe some of you know better (faster) algorithm to do want I want to achieve? It can't use additional memory (arrays).
Here is code to run on ideone: IDEONE
When in doubt, use brute force. Alas, generate all variations with repetition, then filter out the unnecessary patterns:
unsigned bit_count(unsigned n)
{
unsigned i = 0;
while (n) {
i += n & 1;
n >>= 1;
}
return i;
}
int main()
{
std::vector<unsigned> combs;
const unsigned N = 4;
const unsigned K = 3;
for (int i = 0; i < (1 << N); i++) {
if (bit_count(i) == K) {
combs.push_back(i);
}
}
// and print 'combs' here
}
Edit: Someone else already pointed out a solution without filtering and brute force, but I'm still going to give you a few hints about this algorithm:
most compilers offer some sort of intrinsic population count function. I know of GCC and Clang which have __builtin_popcount(). Using this intrinsic function, I was able to double the speed of the code.
Since you seem to be working on GPUs, you can parallelize the code. I have done it using C++11's standard threading facilities, and I've managed to compute all 32-bit repetitions for arbitrarily-chosen popcounts 1, 16 and 19 in 7.1 seconds on my 8-core Intel machine.
Here's the final code I've written:
#include <vector>
#include <cstdio>
#include <thread>
#include <utility>
#include <future>
unsigned popcount_range(unsigned popcount, unsigned long min, unsigned long max)
{
unsigned n = 0;
for (unsigned long i = min; i < max; i++) {
n += __builtin_popcount(i) == popcount;
}
return n;
}
int main()
{
const unsigned N = 32;
const unsigned K = 16;
const unsigned N_cores = 8;
const unsigned long Max = 1ul << N;
const unsigned long N_per_core = Max / N_cores;
std::vector<std::future<unsigned>> v;
for (unsigned core = 0; core < N_cores; core++) {
unsigned long core_min = N_per_core * core;
unsigned long core_max = core_min + N_per_core;
auto fut = std::async(
std::launch::async,
popcount_range,
K,
core_min,
core_max
);
v.push_back(std::move(fut));
}
unsigned final_count = 0;
for (auto &fut : v) {
final_count += fut.get();
}
printf("%u\n", final_count);
return 0;
}

Compact a hex number

Is there a clever (ie: branchless) way to "compact" a hex number. Basically move all the 0s all to one side?
eg:
0x10302040 -> 0x13240000
or
0x10302040 -> 0x00001324
I looked on Bit Twiddling Hacks but didn't see anything.
It's for a SSE numerical pivoting algorithm. I need to remove any pivots that become 0. I can use _mm_cmpgt_ps to find good pivots, _mm_movemask_ps to convert that in to a mask, and then bit hacks to get something like the above. The hex value gets munged in to a mask for a _mm_shuffle_ps instruction to perform a permutation on the SSE 128 bit register.
To compute mask for _pext:
mask = arg;
mask |= (mask << 1) & 0xAAAAAAAA | (mask >> 1) & 0x55555555;
mask |= (mask << 2) & 0xCCCCCCCC | (mask >> 2) & 0x33333333;
First do bit-or on pairs of bits, then on quads. Masks prevent shifted values from overflowing to other digits.
After computing mask this way or harold's way (which is probably faster) you don't need the full power of _pext, so if targeted hardware doesn't support it you can replace it with this:
for(int i = 0; i < 7; i++) {
stay_mask = mask & (~mask - 1);
arg = arg & stay_mask | (arg >> 4) & ~stay_mask;
mask = stay_mask | (mask >> 4);
}
Each iteration moves all nibbles one digit to the right if there is some space. stay_mask marks bits that are in their final positions. This uses somewhat less operations than Hacker's Delight solution, but might still benefit from branching.
Supposing we can use _pext_u32, the issue then is computing a mask that has an F for every nibble that isn't zero. I'm not sure what the best approach is, but you can compute the OR of the 4 bits of the nibble and then "spread" it back out to F's like this:
// calculate horizontal OR of every nibble
x |= x >> 1;
x |= x >> 2;
// clean up junk
x &= 0x11111111;
// spread
x *= 0xF;
Then use that as the mask of _pext_u32.
_pext_u32 can be emulated by this (taken from Hacker's Delight, figure 7.6)
unsigned compress(unsigned x, unsigned m) {
unsigned mk, mp, mv, t;
int i;
x = x & m; // Clear irrelevant bits.
mk = ~m << 1; // We will count 0's to right.
for (i = 0; i < 5; i++) {
mp = mk ^ (mk << 1); // Parallel prefix.
mp = mp ^ (mp << 2);
mp = mp ^ (mp << 4);
mp = mp ^ (mp << 8);
mp = mp ^ (mp << 16);
mv = mp & m; // Bits to move.
m = m ^ mv | (mv >> (1 << i)); // Compress m.
t = x & mv;
x = x ^ t | (t >> (1 << i)); // Compress x.
mk = mk & ~mp;
}
return x;
}
But that's a bit of a disaster. It's probably better to just resort to branching code then.
uint32_t fun(uint32_t val) {
uint32_t retVal(0x00);
uint32_t sa(28);
for (int sb(28); sb >= 0; sb -= 4) {
if (val & (0x0F << sb)) {
retVal |= (0x0F << sb) << (sa - sb)
sa -= 4;
}
}
return retVal;
}
I think this (or something similar) is what you're looking for. Eliminating the 0 nibbles within a number. I've not debugged it, and it would only works on one side atm.
If your processor supports conditional instruction execution, you may get a benefit from this algorithm:
uint32_t compact(uint32_t orig_value)
{
uint32_t mask = 0xF0000000u; // Mask for isolating a hex digit.
uint32_t new_value = 0u;
for (unsigned int i = 0; i < 8; ++i) // 8 hex digits
{
if (orig_value & mask == 0u)
{
orig_value = orig_value << 4; // Shift the original value by 1 digit
}
new_value |= orig_value & mask;
mask = mask >> 4; // next digit
}
return new_value;
}
This looks like a good candidate for loop unrolling.
The algorithm assumes that when the original value is shifted left, zeros are shifted in, filling in the "empty" bits.
Edit 1:
On a processor that supports conditional execution of instructions, the shifting of the original value would be conditionally executed depending on the result of the ANDing of the original value and the mask. Thus no branching, only ignored instructions.
I came up with the following solution. Please take a look, maybe it will help you.
#include <iostream>
#include <sstream>
#include <algorithm>
using namespace std;
class IsZero
{
public:
bool operator ()(char c)
{
return '0' == c;
}
};
int main()
{
int a = 0x01020334; //IMPUT
ostringstream my_sstream;
my_sstream << hex << a;
string str = my_sstream.str();
int base_str_length = str.size();
cout << "Input hex: " << str << endl;
str.insert(remove_if(begin(str), end(str), IsZero()), count_if(begin(str), end(str), IsZero()), '0');
str.replace(begin(str) + base_str_length, end(str), "");
cout << "Processed hex: " << str << endl;
return 0;
}
Output:
Input hex: 1020334
Processed hex: 1233400

How to replace (or speed up) the parallel comparison when the number of pairs are extremely huge?

The run time of following code, parallel comparsion, takes forever, when the number of key in the map is huge(e.g 100000) and each of its second element have huge element(e.g 100000) as well.
Is there any possible way to speed up the the comparsion? My cpu is Xeon E5450 3.00G 4 Cores. Ram is fair enough.
// There is a map with long as its key and vector<long> as second element,
// the vector's elements are increasing sorted.
map<long, vector<long> > = aMap() ;
map<long, vector<long> >::iterator it1 = aMap.begin() ;
map<long, vector<long> >::iterator it2;
// the code need compare each key's second elements
for( ; it1 != aMap.end(); it1++ ) {
it2 = it1;
it2++;
// Parallel comparsion: THE MOST TIME CONSUMING PART
for( ; it2 != aMap.end(); it2++ ) {
unsigned long i = 0, j = 0, _union = 0, _inter = 0 ;
while( i < it1->second.size() && j < it2->second.size() ) {
if( it1->second[i] < it2->second[j] ) {
i++;
} else if( it1->second[i] > it2->second[j] ) {
j++;
} else {
i++; j++; _inter++;
}
}
_union = it1->second.size() + it2->second.size() - _inter;
if ( (double) _inter / _union > THRESH )
cout << it1->first << " might appears frequently with " << it2->first << endl;
}
}
(This is not a complete answer. It only solves part of your problem; the part about bit manipulation.)
Here's a class you might be able to use to calculate the number of intersections between two sets (the cardinality of the intersection) rather quickly.
It uses a bit vector to store the sets, which means the universe of the possible set members must be small.
#include <cassert>
#include <vector>
class BitVector
{
// IMPORTANT: U must be unsigned
// IMPORTANT: use unsigned long long in 64-bit builds.
typedef unsigned long U;
static const unsigned UBits = 8 * sizeof(U);
public:
BitVector (unsigned size)
: m_bits ((size + UBits - 1) / UBits, 0)
, m_size (size)
{
}
void set (unsigned bit)
{
assert (bit < m_size);
m_bits[bit / UBits] |= (U)1 << (bit % UBits);
}
void clear (unsigned bit)
{
assert (bit < m_size);
m_bits[bit / UBits] &= ~((U)1 << (bit % UBits));
}
unsigned countIntersect (BitVector const & that) const
{
assert (m_size == that.m_size);
unsigned ret = 0;
for (std::vector<U>::const_iterator i = m_bits.cbegin(),
j = that.m_bits.cbegin(), e = m_bits.cend(), f = that.m_bits.cend();
i != e && j != f; ++i, ++j)
{
U x = *i & *j;
// Count the number of 1 bits in x and add it to ret
// There are much better ways than this,
// e.g. using the POPCNT instruction or intrinsic
while (x != 0)
{
ret += x & 1;
x >>= 1;
}
}
return ret;
}
unsigned countUnion (BitVector const & that) const
{
assert (m_size == that.m_size);
unsigned ret = 0;
for (std::vector<U>::const_iterator i = m_bits.cbegin(),
j = that.m_bits.cbegin(), e = m_bits.cend(), f = that.m_bits.cend();
i != e && j != f; ++i, ++j)
{
U x = *i | *j;
while (x != 0)
{
ret += x & 1;
x >>= 1;
}
}
return ret;
}
private:
std::vector<U> m_bits;
unsigned m_size;
};
And here's a very small test program to see how you can use the above class. It makes two sets (each with 100K maximum elements), adds some items to them (using the set() member function) and then calculate their intersection 1 billion times. It runs in under two seconds on my machine.
#include <iostream>
using namespace std;
int main ()
{
unsigned const SetSize = 100000;
BitVector a (SetSize), b (SetSize);
for (unsigned i = 0; i < SetSize; i += 2) a.set (i);
for (unsigned i = 0; i < SetSize; i += 3) b.set (i);
unsigned x = a.countIntersect (b);
cout << x << endl;
return 0;
}
Don't forget to compile this with optimizations enabled! Otherwise it performs very badly.
POPCNT
Modern processors have an instruction to count the number of set bits in a word, called POPCNT. This is quite a lot faster than doing the naive thing written above (as a side note, there are faster ways to do it in software as well, but I didn't want to pollute the code.)
Anyways, the way to use POPCNT in C/C++ code is to use a compiler intrinsic or built-in. In MSVC, you can use __popcount() intrinsic that works on 32-bit integers. In GCC, you can use __builtin_popcountl() for 32-bit integers and __builtin_popcountll() for 64 bits. Be warned that these functions may not be available due to your compiler version, target architecture and/or compile switches.
Maybe you would like to try PPL. Or some of its analogues. I don't really understand what your code suppose to do, as it doesn't seem to have any output. But side effects free code can be effectively parallelized with tools like Parallel Patterns Library.

How to determine how many bytes an integer needs?

I'm looking for the most efficient way to calculate the minimum number of bytes needed to store an integer without losing precision.
e.g.
int: 10 = 1 byte
int: 257 = 2 bytes;
int: 18446744073709551615 (UINT64_MAX) = 8 bytes;
Thanks
P.S. This is for a hash functions which will be called many millions of times
Also the byte sizes don't have to be a power of two
The fastest solution seems to one based on tronics answer:
int bytes;
if (hash <= UINT32_MAX)
{
if (hash < 16777216U)
{
if (hash <= UINT16_MAX)
{
if (hash <= UINT8_MAX) bytes = 1;
else bytes = 2;
}
else bytes = 3;
}
else bytes = 4;
}
else if (hash <= UINT64_MAX)
{
if (hash < 72057594000000000ULL)
{
if (hash < 281474976710656ULL)
{
if (hash < 1099511627776ULL) bytes = 5;
else bytes = 6;
}
else bytes = 7;
}
else bytes = 8;
}
The speed difference using mostly 56 bit vals was minimal (but measurable) compared to Thomas Pornin answer. Also i didn't test the solution using __builtin_clzl which could be comparable.
Use this:
int n = 0;
while (x != 0) {
x >>= 8;
n ++;
}
This assumes that x contains your (positive) value.
Note that zero will be declared encodable as no byte at all. Also, most variable-size encodings need some length field or terminator to know where encoding stops in a file or stream (usually, when you encode an integer and mind about size, then there is more than one integer in your encoded object).
You need just two simple ifs if you are interested on the common sizes only. Consider this (assuming that you actually have unsigned values):
if (val < 0x10000) {
if (val < 0x100) // 8 bit
else // 16 bit
} else {
if (val < 0x100000000L) // 32 bit
else // 64 bit
}
Should you need to test for other sizes, choosing a middle point and then doing nested tests will keep the number of tests very low in any case. However, in that case making the testing a recursive function might be a better option, to keep the code simple. A decent compiler will optimize away the recursive calls so that the resulting code is still just as fast.
Assuming a byte is 8 bits, to represent an integer x you need [log2(x) / 8] + 1 bytes where [x] = floor(x).
Ok, I see now that the byte sizes aren't necessarily a power of two. Consider the byte sizes b. The formula is still [log2(x) / b] + 1.
Now, to calculate the log, either use lookup tables (best way speed-wise) or use binary search, which is also very fast for integers.
The function to find the position of the first '1' bit from the most significant side (clz or bsr) is usually a simple CPU instruction (no need to mess with log2), so you could divide that by 8 to get the number of bytes needed. In gcc, there's __builtin_clz for this task:
#include <limits.h>
int bytes_needed(unsigned long long x) {
int bits_needed = sizeof(x)*CHAR_BIT - __builtin_clzll(x);
if (bits_needed == 0)
return 1;
else
return (bits_needed + 7) / 8;
}
(On MSVC you would use the _BitScanReverse intrinsic.)
You may first get the highest bit set, which is the same as log2(N), and then get the bytes needed by ceil(log2(N) / 8).
Here are some bit hacks for getting the position of the highest bit set, which are copied from http://graphics.stanford.edu/~seander/bithacks.html#IntegerLogObvious, and you can click the URL for details of how these algorithms work.
Find the integer log base 2 of an integer with an 64-bit IEEE float
int v; // 32-bit integer to find the log base 2 of
int r; // result of log_2(v) goes here
union { unsigned int u[2]; double d; } t; // temp
t.u[__FLOAT_WORD_ORDER==LITTLE_ENDIAN] = 0x43300000;
t.u[__FLOAT_WORD_ORDER!=LITTLE_ENDIAN] = v;
t.d -= 4503599627370496.0;
r = (t.u[__FLOAT_WORD_ORDER==LITTLE_ENDIAN] >> 20) - 0x3FF;
Find the log base 2 of an integer with a lookup table
static const char LogTable256[256] =
{
#define LT(n) n, n, n, n, n, n, n, n, n, n, n, n, n, n, n, n
-1, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
LT(4), LT(5), LT(5), LT(6), LT(6), LT(6), LT(6),
LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7)
};
unsigned int v; // 32-bit word to find the log of
unsigned r; // r will be lg(v)
register unsigned int t, tt; // temporaries
if (tt = v >> 16)
{
r = (t = tt >> 8) ? 24 + LogTable256[t] : 16 + LogTable256[tt];
}
else
{
r = (t = v >> 8) ? 8 + LogTable256[t] : LogTable256[v];
}
Find the log base 2 of an N-bit integer in O(lg(N)) operations
unsigned int v; // 32-bit value to find the log2 of
const unsigned int b[] = {0x2, 0xC, 0xF0, 0xFF00, 0xFFFF0000};
const unsigned int S[] = {1, 2, 4, 8, 16};
int i;
register unsigned int r = 0; // result of log2(v) will go here
for (i = 4; i >= 0; i--) // unroll for speed...
{
if (v & b[i])
{
v >>= S[i];
r |= S[i];
}
}
// OR (IF YOUR CPU BRANCHES SLOWLY):
unsigned int v; // 32-bit value to find the log2 of
register unsigned int r; // result of log2(v) will go here
register unsigned int shift;
r = (v > 0xFFFF) << 4; v >>= r;
shift = (v > 0xFF ) << 3; v >>= shift; r |= shift;
shift = (v > 0xF ) << 2; v >>= shift; r |= shift;
shift = (v > 0x3 ) << 1; v >>= shift; r |= shift;
r |= (v >> 1);
// OR (IF YOU KNOW v IS A POWER OF 2):
unsigned int v; // 32-bit value to find the log2 of
static const unsigned int b[] = {0xAAAAAAAA, 0xCCCCCCCC, 0xF0F0F0F0,
0xFF00FF00, 0xFFFF0000};
register unsigned int r = (v & b[0]) != 0;
for (i = 4; i > 0; i--) // unroll for speed...
{
r |= ((v & b[i]) != 0) << i;
}
Find the number of bits by taking the log2 of the number, then divide that by 8 to get the number of bytes.
You can find logn of x by the formula:
logn(x) = log(x) / log(n)
Update:
Since you need to do this really quickly, Bit Twiddling Hacks has several methods for quickly calculating log2(x). The look-up table approach seems like it would suit your needs.
This will get you the number of bytes. It's not strictly the most efficient, but unless you're programming a nanobot powered by the energy contained in a red blood cell, it won't matter.
int count = 0;
while (numbertotest > 0)
{
numbertotest >>= 8;
count++;
}
You could write a little template meta-programming code to figure it out at compile time if you need it for array sizes:
template<unsigned long long N> struct NBytes
{ static const size_t value = NBytes<N/256>::value+1; };
template<> struct NBytes<0>
{ static const size_t value = 0; };
int main()
{
std::cout << "short = " << NBytes<SHRT_MAX>::value << " bytes\n";
std::cout << "int = " << NBytes<INT_MAX>::value << " bytes\n";
std::cout << "long long = " << NBytes<ULLONG_MAX>::value << " bytes\n";
std::cout << "10 = " << NBytes<10>::value << " bytes\n";
std::cout << "257 = " << NBytes<257>::value << " bytes\n";
return 0;
}
output:
short = 2 bytes
int = 4 bytes
long long = 8 bytes
10 = 1 bytes
257 = 2 bytes
Note: I know this isn't answering the original question, but it answers a related question that people will be searching for when they land on this page.
Floor((log2(N) / 8) + 1) bytes
You need exactly the log function
nb_bytes = floor(log(x)/log(256))+1
if you use log2, log2(256) == 8 so
floor(log2(x)/8)+1
You need to raise 256 to successive powers until the result is larger than your value.
For example: (Tested in C#)
long long limit = 1;
int byteCount;
for (byteCount = 1; byteCount < 8; byteCount++) {
limit *= 256;
if (limit > value)
break;
}
If you only want byte sizes to be powers of two (If you don't want 65,537 to return 3), replace byteCount++ with byteCount *= 2.
I think this is a portable implementation of the straightforward formula:
#include <limits.h>
#include <math.h>
#include <stdio.h>
int main(void) {
int i;
unsigned int values[] = {10, 257, 67898, 140000, INT_MAX, INT_MIN};
for ( i = 0; i < sizeof(values)/sizeof(values[0]); ++i) {
printf("%d needs %.0f bytes\n",
values[i],
1.0 + floor(log(values[i]) / (M_LN2 * CHAR_BIT))
);
}
return 0;
}
Output:
10 needs 1 bytes
257 needs 2 bytes
67898 needs 3 bytes
140000 needs 3 bytes
2147483647 needs 4 bytes
-2147483648 needs 4 bytes
Whether and how much the lack of speed and the need to link floating point libraries depends on your needs.
I know this question didn't ask for this type of answer but for those looking for a solution using the smallest number of characters, this does the assignment to a length variable in 17 characters, or 25 including the declaration of the length variable.
//Assuming v is the value that is being counted...
int l=0;
for(;v>>l*8;l++);
This is based on SoapBox's idea of creating a solution that contains no jumps, branches etc... Unfortunately his solution was not quite correct. I have adopted the spirit and here's a 32bit version, the 64bit checks can be applied easily if desired.
The function returns number of bytes required to store the given integer.
unsigned short getBytesNeeded(unsigned int value)
{
unsigned short c = 0; // 0 => size 1
c |= !!(value & 0xFF00); // 1 => size 2
c |= (!!(value & 0xFF0000)) << 1; // 2 => size 3
c |= (!!(value & 0xFF000000)) << 2; // 4 => size 4
static const int size_table[] = { 1, 2, 3, 3, 4, 4, 4, 4 };
return size_table[c];
}
For each of eight times, shift the int eight bits to the right and see if there are still 1-bits left. The number of times you shift before you stop is the number of bytes you need.
More succinctly, the minimum number of bytes you need is ceil(min_bits/8), where min_bits is the index (i+1) of the highest set bit.
There are a multitude of ways to do this.
Option #1.
int numBytes = 0;
do {
numBytes++;
} while (i >>= 8);
return (numBytes);
In the above example, is the number you are testing, and generally works for any processor, any size of integer.
However, it might not be the fastest. Alternatively, you can try a series of if statements ...
For a 32 bit integers
if ((upper = (value >> 16)) == 0) {
/* Bit in lower 16 bits may be set. */
if ((high = (value >> 8)) == 0) {
return (1);
}
return (2);
}
/* Bit in upper 16 bits is set */
if ((high = (upper >> 8)) == 0) {
return (3);
}
return (4);
For 64 bit integers, Another level of if statements would be required.
If the speed of this routine is as critical as you say, it might be worthwhile to do this in assembler if you want it as a function call. That could allow you to avoid creating and destroying the stack frame, saving a few extra clock cycles if it is that critical.
A bit basic, but since there will be a limited number of outputs, can you not pre-compute the breakpoints and use a case statement? No need for calculations at run-time, only a limited number of comparisons.
Why not just use a 32-bit hash?
That will work at near-top-speed everywhere.
I'm rather confused as to why a large hash would even be wanted. If a 4-byte hash works, why not just use it always? Excepting cryptographic uses, who has hash tables with more then 232 buckets anyway?
there are lots of great recipes for stuff like this over at Sean Anderson's "Bit Twiddling Hacks" page.
This code has 0 branches, which could be faster on some systems. Also on some systems (GPGPU) its important for threads in the same warp to execute the same instructions. This code is always the same number of instructions no matter what the input value.
inline int get_num_bytes(unsigned long long value) // where unsigned long long is the largest integer value on this platform
{
int size = 1; // starts at 1 sot that 0 will return 1 byte
size += !!(value & 0xFF00);
size += !!(value & 0xFFFF0000);
if (sizeof(unsigned long long) > 4) // every sane compiler will optimize this out
{
size += !!(value & 0xFFFFFFFF00000000ull);
if (sizeof(unsigned long long) > 8)
{
size += !!(value & 0xFFFFFFFFFFFFFFFF0000000000000000ull);
}
}
static const int size_table[] = { 1, 2, 4, 8, 16 };
return size_table[size];
}
g++ -O3 produces the following (verifying that the ifs are optimized out):
xor %edx,%edx
test $0xff00,%edi
setne %dl
xor %eax,%eax
test $0xffff0000,%edi
setne %al
lea 0x1(%rdx,%rax,1),%eax
movabs $0xffffffff00000000,%rdx
test %rdx,%rdi
setne %dl
lea (%rdx,%rax,1),%rax
and $0xf,%eax
mov _ZZ13get_num_bytesyE10size_table(,%rax,4),%eax
retq
Why so complicated? Here's what I came up with:
bytesNeeded = (numBits/8)+((numBits%8) != 0);
Basically numBits divided by eight + 1 if there is a remainder.
There are already a lot of answers here, but if you know the number ahead of time, in c++ you can use a template to make use of the preprocessor.
template <unsigned long long N>
struct RequiredBytes {
enum : int { value = 1 + (N > 255 ? RequiredBits<(N >> 8)>::value : 0) };
};
template <>
struct RequiredBytes<0> {
enum : int { value = 1 };
};
const int REQUIRED_BYTES_18446744073709551615 = RequiredBytes<18446744073709551615>::value; // 8
or for a bits version:
template <unsigned long long N>
struct RequiredBits {
enum : int { value = 1 + RequiredBits<(N >> 1)>::value };
};
template <>
struct RequiredBits<1> {
enum : int { value = 1 };
};
template <>
struct RequiredBits<0> {
enum : int { value = 1 };
};
const int REQUIRED_BITS_42 = RequiredBits<42>::value; // 6