Related
My task is to design a function that fulfils those requirements:
Function shall sum members of given one-dimensional array. However, it should sum only members whose number of ones in the binary representation is higher than defined threshold (e.g. if the threshold is 4, number 255 will be counted and 15 will not)
The array length is arbitrary
The function shall utilize as little memory as possible and shall be written in an efficient way
The production function code (‘sum_filtered(){..}’) shall not use any standard C library functions (or any other libraries)
The function shall return 0 on success and error code on error
The array elements are of a type 16-bit signed integer and an overflow during calculation shall be regarded as a failure
Use data types that ensure portability between different CPUs (so the calculations will be the same on 8/16/32-bit MCU)
The function code should contain a reasonable amount of comments in doxygen annotation
Here is my solution:
#include <iostream>
using namespace std;
int sum_filtered(short array[], int treshold)
{
// return 1 if invalid input parameters
if((treshold < 0) || (treshold > 16)){return(1);}
int sum = 0;
int bitcnt = 0;
for(int i=0; i < sizeof(array); i++)
{
// Count one bits of integer
bitcnt = 0;
for (int pos = 0 ; pos < 16 ; pos++) {if (array[i] & (1 << pos)) {bitcnt++;}}
// Add integer to sum if bitcnt>treshold
if(bitcnt>treshold){sum += array[i];}
}
return(0);
}
int main()
{
short array[5] = {15, 2652, 14, 1562, -115324};
int result = sum_filtered(array, 14);
cout << result << endl;
short array2[5] = {15, 2652, 14, 1562, 15324};
result = sum_filtered(array2, -2);
cout << result << endl;
}
However I'm not sure whether this code is portable between different CPUs.
And I don't how can an overflow occur during calculation and what can be other errors during processing of arrays with this function.
Can somebody more experienced give me his opinion?
Well, I can foresee one problem:
for(int i=0; i < sizeof(array); i++)
array in this context is a pointer, so will likely be 4 on 32bit systems, or 8 on 64bit systems. You really do want to be passing a count variable (in this case 5) into the sum_filtered function (and then you can pass the count as sizeof(array) / sizeof(short)).
Anyhow, this code:
// Count one bits of integer
bitcnt = 0;
for (int pos = 0 ; pos < 16 ; pos++) {if (array[i] & (1 << pos)) {bitcnt++;}}
Effectively you are doing a popcount here (which can be done using __builtin_popcount on gcc/clang, or __popcnt on MSVC. They are compiler specific, but usually boil down to a single popcount CPU instruction on most CPUs).
If you do want to do this the slow way, then an efficient approach is to treat the computation as a form of bitwise SIMD operation:
#include <cstdint> // or stdint.h if you have a rubbish compiler :)
uint16_t popcount(uint16_t s)
{
// perform 8x 1bit adds
uint16_t a0 = s & 0x5555;
uint16_t b0 = (s >> 1) & 0x5555;
uint16_t s0 = a0 + b0;
// perform 4x 2bit adds
uint16_t a1 = s0 & 0x3333;
uint16_t b1 = (s0 >> 2) & 0x3333;
uint16_t s1 = a1 + b1;
// perform 2x 4bit adds
uint16_t a2 = s1 & 0x0F0F;
uint16_t b2 = (s1 >> 4) & 0x0F0F;
uint16_t s2 = a2 + b2;
// perform 1x 8bit adds
uint16_t a3 = s2 & 0x00FF;
uint16_t b3 = (s2 >> 8) & 0x00FF;
return a3 + b3;
}
I know it says you can't use stdlib functions (your 4th point), but that shouldn't apply to the standardised integer types surely? (e.g. uint16_t) If it does, well then there is no way to guarantee portability across platforms. You're out of luck.
Personally I'd just use a 64bit integer for the sum. That should reduce the risk of any overflows *(i.e. if the threshold is zero, and all the values are -128, then you'd overflow if the array size exceeded 0x1FFFFFFFFFFFF elements (562,949,953,421,311 in decimal).
#include <cstdint>
int64_t sum_filtered(int16_t array[], uint16_t threshold, size_t array_length)
{
// changing the type on threshold to be unsigned means we don't need to test
// for negative numbers.
if(threshold > 16) { return 1; }
int64_t sum = 0;
for(size_t i=0; i < array_length; i++)
{
if (popcount(array[i]) > threshold)
{
sum += array[i];
}
}
return sum;
}
Input is a bitarray stored in contiguous memory with 1 bit of the bitarray per 1 bit of memory.
Output is an array of the indices of set bits of the bitarray.
Example:
bitarray: 0000 1111 0101 1010
setA: {4,5,6,7,9,11,12,14}
setB: {2,4,5,7,9,10,11,12}
Getting either set A or set B is fine.
The set is stored as an array of uint32_t so each element of the set is an unsigned 32 bit integer in the array.
How to do this about 5 times faster on a single cpu core?
current code:
#include <iostream>
#include <vector>
#include <time.h>
using namespace std;
template <typename T>
uint32_t bitarray2set(T& v, uint32_t * ptr_set){
uint32_t i;
uint32_t base = 0;
uint32_t * ptr_set_new = ptr_set;
uint32_t size = v.capacity();
for(i = 0; i < size; i++){
find_set_bit(v[i], ptr_set_new, base);
base += 8*sizeof(uint32_t);
}
return (ptr_set_new - ptr_set);
}
inline void find_set_bit(uint32_t n, uint32_t*& ptr_set, uint32_t base){
// Find the set bits in a uint32_t
int k = base;
while(n){
if (n & 1){
*(ptr_set) = k;
ptr_set++;
}
n = n >> 1;
k++;
}
}
template <typename T>
void rand_vector(T& v){
srand(time(NULL));
int i;
int size = v.capacity();
for (i=0;i<size;i++){
v[i] = rand();
}
}
template <typename T>
void print_vector(T& v, int size_in = 0){
int i;
int size;
if (size_in == 0){
size = v.capacity();
} else {
size = size_in;
}
for (i=0;i<size;i++){
cout << v[i] << ' ';
}
cout << endl;
}
int main(void){
const int test_size = 6000;
vector<uint32_t> vec(test_size);
vector<uint32_t> set(test_size*sizeof(uint32_t)*8);
rand_vector(vec);
//for (int i; i < 64; i++) vec[i] = -1;
//cout << "input" << endl;
print_vector(vec);
//cout << "calculate result" << endl;
int i;
int rep = 10000;
uint32_t res_size;
struct timespec tp_start, tp_end;
clock_gettime(CLOCK_MONOTONIC, &tp_start);
for (i=0;i<rep;i++){
res_size = bitarray2set(vec, set.data());
}
clock_gettime(CLOCK_MONOTONIC, &tp_end);
double timing;
const double nano = 0.000000001;
timing = ((double)(tp_end.tv_sec - tp_start.tv_sec )
+ (tp_end.tv_nsec - tp_start.tv_nsec) * nano) /(rep);
cout << "timing per cycle: " << timing << endl;
cout << "print result" << endl;
//print_vector(set, res_size);
}
result (compiled with icc -O3 code.cpp -lrt)
...
timing per cycle: 0.000739613 (7.4E-4).
print result
0.0008 seconds to convert 768000 bits to set. But there are at least 10,000 arrays of 768,000 bits in each cycle. That is 8 seconds per cycle. That is slow.
The cpu has popcnt instruction and sse4.2 instruction set.
Thanks.
Update
template <typename T>
uint32_t bitarray2set(T& v, uint32_t * ptr_set){
uint32_t i;
uint32_t base = 0;
uint32_t * ptr_set_new = ptr_set;
uint32_t size = v.capacity();
uint32_t * ptr_v;
uint32_t * ptr_v_end = &(v[size]);
for(ptr_v = v.data(); ptr_v < ptr_v_end; ++ptr_v){
while(*ptr_v) {
*ptr_set_new++ = base + __builtin_ctz(*ptr_v);
(*ptr_v) &= (*ptr_v) - 1; // zeros the lowest 1-bit in n
}
base += 8*sizeof(uint32_t);
}
return (ptr_set_new - ptr_set);
}
This updated version uses the inner loop provided by rhashimoto. I don't know if the inlining actually makes the function slower (i never thought that can happen!). The new timing is 1.14E-5 (compiled by icc -O3 code.cpp -lrt, and benchmarked on random vector).
Warning:
I just found that reserving instead of resizing a std::vector, and then write directly to the vector's data through raw pointing is a bad idea. Resizing first and then use raw pointer is fine though. See Robᵩ's answer at Resizing a C++ std::vector<char> without initializing data I am going to just use resize instead of reserve and stop worrying about the time that resize wastes by calling constructor of each element of the vector... at least vectors actually uses contiguous memory, like a plain array (Are std::vector elements guaranteed to be contiguous?)
I notice that you use .capacity() when you probably mean to use .size(). That could make you do extra unnecessary work, as well as giving you the wrong answer.
Your loop in find_set_bit() iterates over all 32 bits in the word. You can instead iterate only over each set bit and use the BSF instruction to determine the index of the lowest bit. GCC has an intrinsic function __builtin_ctz() to generate BSF or the equivalent - I think that the Intel compiler also supports it (you can inline assembly if not). The modified function would look like this:
inline void find_set_bit(uint32_t n, uint32_t*& ptr_set, uint32_t base){
// Find the set bits in a uint32_t
while(n) {
*ptr_set++ = base + __builtin_ctz(n);
n &= n - 1; // zeros the lowest 1-bit in n
}
}
On my Linux machine, compiling with g++ -O3, replacing that function drops the reported time from 0.000531434 to 0.000101352.
There are quite a few ways to find a bit index in the answers to this question. I do think that __builtin_ctz() is going to be the best choice for you, though. I don't believe that there is a reasonable SIMD approach to your problem, as each input word produces a variable amount of output.
As suggested by #davidbak, you could use a table lookup to process 4 elements of the bitmap at once.
Each lookup produces a variable-sized chunk of set members, which we can handle by using popcnt.
#rhashimoto's scalar ctz-based suggestion will probably do better with sparse bitsets that have lots of zeros, but this should be better when there are a lot of set bits.
I'm thinking something like
// a vector of 4 elements for every pattern of 4 bits.
// values range from 0 to 3, and will have a multiple of 4 added to them.
alignas(16) static const int LUT[16*4] = { 0,0,0,0, ... };
// mostly C, some pseudocode.
unsigned int bitmap2set(int *set, int input) {
int *set_start = set;
__m128i offset = _mm_setzero_si128();
for (nibble in input[]) { // pseudocode for the actual shifting / masking
__m128i v = _mm_load_si128(&LUT[nibble]);
__m128i vpos = _mm_add_epi32(v, offset);
_mm_store((__m128i*)set, vpos);
set += _mm_popcount_u32(nibble); // variable-length store
offset = _mm_add_epi32(offset, _mm_set1_epi32(4)); // increment the offset by 4
}
return set - set_start; // set size
}
When a nibble isn't 1111, the next store will overlap, but that's fine.
Using popcnt to figure out how much to increment a pointer is a useful technique in general for left-packing variable-length data into a destination array.
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;
}
I am using Anthony Williams' fixed point library described in the Dr Dobb's article "Optimizing Math-Intensive Applications with Fixed-Point Arithmetic" to calculate the distance between two geographical points using the Rhumb Line method.
This works well enough when the distance between the points is significant (greater than a few kilometers), but is very poor at smaller distances. The worst case being when the two points are equal or near equal, the result is a distance of 194 meters, while I need precision of at least 1 metre at distances >= 1 metre.
By comparison with a double precision floating-point implementation, I have located the problem to the fixed::sqrt() function, which performs poorly at small values:
x std::sqrt(x) fixed::sqrt(x) error
----------------------------------------------------
0 0 3.05176e-005 3.05176e-005
1e-005 0.00316228 0.00316334 1.06005e-006
2e-005 0.00447214 0.00447226 1.19752e-007
3e-005 0.00547723 0.0054779 6.72248e-007
4e-005 0.00632456 0.00632477 2.12746e-007
5e-005 0.00707107 0.0070715 4.27244e-007
6e-005 0.00774597 0.0077467 7.2978e-007
7e-005 0.0083666 0.00836658 1.54875e-008
8e-005 0.00894427 0.00894427 1.085e-009
Correcting the result for fixed::sqrt(0) is trivial by treating it as a special case, but that will not solve the problem for small non-zero distances, where the error starts at 194 metres and converges toward zero with increasing distance. I probably need at least an order of maginitude improvement in precision toward zero.
The fixed::sqrt() algorithim is briefly explained on page 4 of the article linked above, but I am struggling to follow it let alone determine whether it is possible to improve it. The code for the function is reproduced below:
fixed fixed::sqrt() const
{
unsigned const max_shift=62;
uint64_t a_squared=1LL<<max_shift;
unsigned b_shift=(max_shift+fixed_resolution_shift)/2;
uint64_t a=1LL<<b_shift;
uint64_t x=m_nVal;
while(b_shift && a_squared>x)
{
a>>=1;
a_squared>>=2;
--b_shift;
}
uint64_t remainder=x-a_squared;
--b_shift;
while(remainder && b_shift)
{
uint64_t b_squared=1LL<<(2*b_shift-fixed_resolution_shift);
int const two_a_b_shift=b_shift+1-fixed_resolution_shift;
uint64_t two_a_b=(two_a_b_shift>0)?(a<<two_a_b_shift):(a>>-two_a_b_shift);
while(b_shift && remainder<(b_squared+two_a_b))
{
b_squared>>=2;
two_a_b>>=1;
--b_shift;
}
uint64_t const delta=b_squared+two_a_b;
if((2*remainder)>delta)
{
a+=(1LL<<b_shift);
remainder-=delta;
if(b_shift)
{
--b_shift;
}
}
}
return fixed(internal(),a);
}
Note that m_nVal is the internal fixed point representation value, it is an int64_t and the representation uses Q36.28 format (fixed_resolution_shift = 28). The representation itself has enough precision for at least 8 decimal places, and as a fraction of equatorial arc is good for distances of around 0.14 metres, so the limitation is not the fixed-point representation.
Use of the rhumb line method is a standards body recommendation for this application so cannot be changed, and in any case a more accurate square-root function is likely to be required elsewhere in the application or in future applications.
Question: Is it possible to improve the accuracy of the fixed::sqrt() algorithm for small non-zero values while still maintaining its bounded and deterministic convergence?
Additional Information
The test code used to generate the table above:
#include <cmath>
#include <iostream>
#include "fixed.hpp"
int main()
{
double error = 1.0 ;
for( double x = 0.0; error > 1e-8; x += 1e-5 )
{
double fixed_root = sqrt(fixed(x)).as_double() ;
double std_root = std::sqrt(x) ;
error = std::fabs(fixed_root - std_root) ;
std::cout << x << '\t' << std_root << '\t' << fixed_root << '\t' << error << std::endl ;
}
}
Conclusion
In the light of Justin Peel's solution and analysis, and comparison with the algorithm in "The Neglected Art of Fixed Point Arithmetic", I have adapted the latter as follows:
fixed fixed::sqrt() const
{
uint64_t a = 0 ; // root accumulator
uint64_t remHi = 0 ; // high part of partial remainder
uint64_t remLo = m_nVal ; // low part of partial remainder
uint64_t testDiv ;
int count = 31 + (fixed_resolution_shift >> 1); // Loop counter
do
{
// get 2 bits of arg
remHi = (remHi << 2) | (remLo >> 62); remLo <<= 2 ;
// Get ready for the next bit in the root
a <<= 1;
// Test radical
testDiv = (a << 1) + 1;
if (remHi >= testDiv)
{
remHi -= testDiv;
a += 1;
}
} while (count-- != 0);
return fixed(internal(),a);
}
While this gives far greater precision, the improvement I needed is not to be achieved. The Q36.28 format alone just about provides the precision I need, but it is not possible to perform a sqrt() without loss of a few bits of precision. However some lateral thinking provides a better solution. My application tests the calculated distance against some distance limit. The rather obvious solution in hindsight is to test the square of the distance against the square of the limit!
Given that sqrt(ab) = sqrt(a)sqrt(b), then can't you just trap the case where your number is small and shift it up by a given number of bits, compute the root and shift that back down by half the number of bits to get the result?
I.e.
sqrt(n) = sqrt(n.2^k)/sqrt(2^k)
= sqrt(n.2^k).2^(-k/2)
E.g. Choose k = 28 for any n less than 2^8.
The original implementation obviously has some problems. I became frustrated with trying to fix them all with the way the code is currently done and ended up going at it with a different approach. I could probably fix the original now, but I like my way better anyway.
I treat the input number as being in Q64 to start which is the same as shifting by 28 and then shifting back by 14 afterwards (the sqrt halves it). However, if you just do that, then the accuracy is limited to 1/2^14 = 6.1035e-5 because the last 14 bits will be 0. To remedy this, I then shift a and remainder correctly and to keep filling in digits I do the loop again. The code can be made more efficient and cleaner, but I'll leave that to someone else. The accuracy shown below is pretty much as good as you can get with Q36.28. If you compare the fixed point sqrt with the floating point sqrt of the input number after it has been truncated by fixed point(convert it to fixed point and back), then the errors are around 2e-9(I didn't do this in the code below, but it requires one line of change). This is right in line with the best accuracy for Q36.28 which is 1/2^28 = 3.7529e-9.
By the way, one big mistake in the original code is that the term where m = 0 is never considered so that bit can never be set. Anyway, here is the code. Enjoy!
#include <iostream>
#include <cmath>
typedef unsigned long uint64_t;
uint64_t sqrt(uint64_t in_val)
{
const uint64_t fixed_resolution_shift = 28;
const unsigned max_shift=62;
uint64_t a_squared=1ULL<<max_shift;
unsigned b_shift=(max_shift>>1) + 1;
uint64_t a=1ULL<<(b_shift - 1);
uint64_t x=in_val;
while(b_shift && a_squared>x)
{
a>>=1;
a_squared>>=2;
--b_shift;
}
uint64_t remainder=x-a_squared;
--b_shift;
while(remainder && b_shift)
{
uint64_t b_squared=1ULL<<(2*(b_shift - 1));
uint64_t two_a_b=(a<<b_shift);
while(b_shift && remainder<(b_squared+two_a_b))
{
b_squared>>=2;
two_a_b>>=1;
--b_shift;
}
uint64_t const delta=b_squared+two_a_b;
if((remainder)>=delta && b_shift)
{
a+=(1ULL<<(b_shift - 1));
remainder-=delta;
--b_shift;
}
}
a <<= (fixed_resolution_shift/2);
b_shift = (fixed_resolution_shift/2) + 1;
remainder <<= (fixed_resolution_shift);
while(remainder && b_shift)
{
uint64_t b_squared=1ULL<<(2*(b_shift - 1));
uint64_t two_a_b=(a<<b_shift);
while(b_shift && remainder<(b_squared+two_a_b))
{
b_squared>>=2;
two_a_b>>=1;
--b_shift;
}
uint64_t const delta=b_squared+two_a_b;
if((remainder)>=delta && b_shift)
{
a+=(1ULL<<(b_shift - 1));
remainder-=delta;
--b_shift;
}
}
return a;
}
double fixed2float(uint64_t x)
{
return static_cast<double>(x) * pow(2.0, -28.0);
}
uint64_t float2fixed(double f)
{
return static_cast<uint64_t>(f * pow(2, 28.0));
}
void finderror(double num)
{
double root1 = fixed2float(sqrt(float2fixed(num)));
double root2 = pow(num, 0.5);
std::cout << "input: " << num << ", fixed sqrt: " << root1 << " " << ", float sqrt: " << root2 << ", finderror: " << root2 - root1 << std::endl;
}
main()
{
finderror(0);
finderror(1e-5);
finderror(2e-5);
finderror(3e-5);
finderror(4e-5);
finderror(5e-5);
finderror(pow(2.0,1));
finderror(1ULL<<35);
}
with the output of the program being
input: 0, fixed sqrt: 0 , float sqrt: 0, finderror: 0
input: 1e-05, fixed sqrt: 0.00316207 , float sqrt: 0.00316228, finderror: 2.10277e-07
input: 2e-05, fixed sqrt: 0.00447184 , float sqrt: 0.00447214, finderror: 2.97481e-07
input: 3e-05, fixed sqrt: 0.0054772 , float sqrt: 0.00547723, finderror: 2.43815e-08
input: 4e-05, fixed sqrt: 0.00632443 , float sqrt: 0.00632456, finderror: 1.26255e-07
input: 5e-05, fixed sqrt: 0.00707086 , float sqrt: 0.00707107, finderror: 2.06055e-07
input: 2, fixed sqrt: 1.41421 , float sqrt: 1.41421, finderror: 1.85149e-09
input: 3.43597e+10, fixed sqrt: 185364 , float sqrt: 185364, finderror: 2.24099e-09
I'm not sure how you're getting the numbers from fixed::sqrt() shown in the table.
Here's what I do:
#include <stdio.h>
#include <math.h>
#define __int64 long long // gcc doesn't know __int64
typedef __int64 fixed;
#define FRACT 28
#define DBL2FIX(x) ((fixed)((double)(x) * (1LL << FRACT)))
#define FIX2DBL(x) ((double)(x) / (1LL << FRACT))
// De-++-ified code from
// http://www.justsoftwaresolutions.co.uk/news/optimizing-applications-with-fixed-point-arithmetic.html
fixed sqrtfix0(fixed num)
{
static unsigned const fixed_resolution_shift=FRACT;
unsigned const max_shift=62;
unsigned __int64 a_squared=1LL<<max_shift;
unsigned b_shift=(max_shift+fixed_resolution_shift)/2;
unsigned __int64 a=1LL<<b_shift;
unsigned __int64 x=num;
unsigned __int64 remainder;
while(b_shift && a_squared>x)
{
a>>=1;
a_squared>>=2;
--b_shift;
}
remainder=x-a_squared;
--b_shift;
while(remainder && b_shift)
{
unsigned __int64 b_squared=1LL<<(2*b_shift-fixed_resolution_shift);
int const two_a_b_shift=b_shift+1-fixed_resolution_shift;
unsigned __int64 two_a_b=(two_a_b_shift>0)?(a<<two_a_b_shift):(a>>-two_a_b_shift);
unsigned __int64 delta;
while(b_shift && remainder<(b_squared+two_a_b))
{
b_squared>>=2;
two_a_b>>=1;
--b_shift;
}
delta=b_squared+two_a_b;
if((2*remainder)>delta)
{
a+=(1LL<<b_shift);
remainder-=delta;
if(b_shift)
{
--b_shift;
}
}
}
return (fixed)a;
}
// Adapted code from
// http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Digit-by-digit_calculation
fixed sqrtfix1(fixed num)
{
fixed res = 0;
fixed bit = (fixed)1 << 62; // The second-to-top bit is set
int s = 0;
// Scale num up to get more significant digits
while (num && num < bit)
{
num <<= 1;
s++;
}
if (s & 1)
{
num >>= 1;
s--;
}
s = 14 - (s >> 1);
while (bit != 0)
{
if (num >= res + bit)
{
num -= res + bit;
res = (res >> 1) + bit;
}
else
{
res >>= 1;
}
bit >>= 2;
}
if (s >= 0) res <<= s;
else res >>= -s;
return res;
}
int main(void)
{
double testData[] =
{
0,
1e-005,
2e-005,
3e-005,
4e-005,
5e-005,
6e-005,
7e-005,
8e-005,
};
int i;
for (i = 0; i < sizeof(testData) / sizeof(testData[0]); i++)
{
double x = testData[i];
fixed xf = DBL2FIX(x);
fixed sqf0 = sqrtfix0(xf);
fixed sqf1 = sqrtfix1(xf);
double sq0 = FIX2DBL(sqf0);
double sq1 = FIX2DBL(sqf1);
printf("%10.8f: "
"sqrtfix0()=%10.8f / err=%e "
"sqrt()=%10.8f "
"sqrtfix1()=%10.8f / err=%e\n",
x,
sq0, fabs(sq0 - sqrt(x)),
sqrt(x),
sq1, fabs(sq1 - sqrt(x)));
}
printf("sizeof(double)=%d\n", (int)sizeof(double));
return 0;
}
And here's what I get (with gcc and Open Watcom):
0.00000000: sqrtfix0()=0.00003052 / err=3.051758e-05 sqrt()=0.00000000 sqrtfix1()=0.00000000 / err=0.000000e+00
0.00001000: sqrtfix0()=0.00311279 / err=4.948469e-05 sqrt()=0.00316228 sqrtfix1()=0.00316207 / err=2.102766e-07
0.00002000: sqrtfix0()=0.00445557 / err=1.656955e-05 sqrt()=0.00447214 sqrtfix1()=0.00447184 / err=2.974807e-07
0.00003000: sqrtfix0()=0.00543213 / err=4.509667e-05 sqrt()=0.00547723 sqrtfix1()=0.00547720 / err=2.438148e-08
0.00004000: sqrtfix0()=0.00628662 / err=3.793423e-05 sqrt()=0.00632456 sqrtfix1()=0.00632443 / err=1.262553e-07
0.00005000: sqrtfix0()=0.00701904 / err=5.202484e-05 sqrt()=0.00707107 sqrtfix1()=0.00707086 / err=2.060551e-07
0.00006000: sqrtfix0()=0.00772095 / err=2.501943e-05 sqrt()=0.00774597 sqrtfix1()=0.00774593 / err=3.390476e-08
0.00007000: sqrtfix0()=0.00836182 / err=4.783859e-06 sqrt()=0.00836660 sqrtfix1()=0.00836649 / err=1.086198e-07
0.00008000: sqrtfix0()=0.00894165 / err=2.621519e-06 sqrt()=0.00894427 sqrtfix1()=0.00894409 / err=1.777289e-07
sizeof(double)=8
EDIT:
I've missed the fact that the above sqrtfix1() won't work well with large arguments. It can be fixed by appending 28 zeroes to the argument and essentially calculating the exact integer square root of that. This comes at the expense of doing internal calculations in 128-bit arithmetic, but it's pretty straightforward:
fixed sqrtfix2(fixed num)
{
unsigned __int64 numl, numh;
unsigned __int64 resl = 0, resh = 0;
unsigned __int64 bitl = 0, bith = (unsigned __int64)1 << 26;
numl = num << 28;
numh = num >> (64 - 28);
while (bitl | bith)
{
unsigned __int64 tmpl = resl + bitl;
unsigned __int64 tmph = resh + bith + (tmpl < resl);
tmph = numh - tmph - (numl < tmpl);
tmpl = numl - tmpl;
if (tmph & 0x8000000000000000ULL)
{
resl >>= 1;
if (resh & 1) resl |= 0x8000000000000000ULL;
resh >>= 1;
}
else
{
numl = tmpl;
numh = tmph;
resl >>= 1;
if (resh & 1) resl |= 0x8000000000000000ULL;
resh >>= 1;
resh += bith + (resl + bitl < resl);
resl += bitl;
}
bitl >>= 2;
if (bith & 1) bitl |= 0x4000000000000000ULL;
if (bith & 2) bitl |= 0x8000000000000000ULL;
bith >>= 2;
}
return resl;
}
And it gives pretty much the same results (slightly better for 3.43597e+10) than this answer:
0.00000000: sqrtfix0()=0.00003052 / err=3.051758e-05 sqrt()=0.00000000 sqrtfix2()=0.00000000 / err=0.000000e+00
0.00001000: sqrtfix0()=0.00311279 / err=4.948469e-05 sqrt()=0.00316228 sqrtfix2()=0.00316207 / err=2.102766e-07
0.00002000: sqrtfix0()=0.00445557 / err=1.656955e-05 sqrt()=0.00447214 sqrtfix2()=0.00447184 / err=2.974807e-07
0.00003000: sqrtfix0()=0.00543213 / err=4.509667e-05 sqrt()=0.00547723 sqrtfix2()=0.00547720 / err=2.438148e-08
0.00004000: sqrtfix0()=0.00628662 / err=3.793423e-05 sqrt()=0.00632456 sqrtfix2()=0.00632443 / err=1.262553e-07
0.00005000: sqrtfix0()=0.00701904 / err=5.202484e-05 sqrt()=0.00707107 sqrtfix2()=0.00707086 / err=2.060551e-07
0.00006000: sqrtfix0()=0.00772095 / err=2.501943e-05 sqrt()=0.00774597 sqrtfix2()=0.00774593 / err=3.390476e-08
0.00007000: sqrtfix0()=0.00836182 / err=4.783859e-06 sqrt()=0.00836660 sqrtfix2()=0.00836649 / err=1.086198e-07
0.00008000: sqrtfix0()=0.00894165 / err=2.621519e-06 sqrt()=0.00894427 sqrtfix2()=0.00894409 / err=1.777289e-07
2.00000000: sqrtfix0()=1.41419983 / err=1.373327e-05 sqrt()=1.41421356 sqrtfix2()=1.41421356 / err=1.851493e-09
34359700000.00000000: sqrtfix0()=185363.69654846 / err=5.097361e-06 sqrt()=185363.69655356 sqrtfix2()=185363.69655356 / err=1
.164153e-09
Many many years ago I worked on a demo program for a small computer our outfit had built. The computer had a built-in square-root instruction, and we built a simple program to demonstrate the computer doing 16-bit add/subtract/multiply/divide/square-root on a TTY. Alas, it turned out that there was a serious bug in the square root instruction, but we had promised to demo the function. So we created an array of the squares of the values 1-255, then used a simple lookup to match the value typed in to one of the array values. The index was the square root.
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