Is there an efficient way to calculate ceiling of log_b(a)? - algorithm

Question

I need to accurately calculate where a and b
are both integers. If I simply use typical change of base formula with floating point math functions I wind up with errors due to rounding error.

Answer 1

You can use this identity:
b^logb(a) = a
So binary search x = logb(a) so the result of b^x is biggest integer which is still less than a and afterwards just increment the final result.
Here small C++ example for 32 bits:
//---------------------------------------------------------------------------
DWORD u32_pow(DWORD a,DWORD b) // = a^b
{
int i,bits=32;
DWORD d=1;
for (i=0;i<bits;i++)
{
d*=d;
if (DWORD(b&0x80000000)) d*=a;
b<<=1;
}
return d;
}
//---------------------------------------------------------------------------
DWORD u32_log2(DWORD a) // = ceil(log2(a))
{
DWORD x;
for (x=32;((a&0x80000000)==0)&&(x>1);x--,a<<=1);
return x;
}
//---------------------------------------------------------------------------
DWORD u32_log(DWORD b,DWORD a) // = ceil(logb(a))
{
DWORD x,m,bx;
// edge cases
if (b< 2) return 0;
if (a< 2) return 0;
if (a<=b) return 1;
m=1<<(u32_log2(a)-1); // max limit for b=2, all other bases lead to smaller exponents anyway
for (x=0;m;m>>=1)
{
x|=m;
bx=u32_pow(b,x);
if (bx>=a) x^=m;
}
return x+1;
}
//---------------------------------------------------------------------------
Where DWORD is any unsigned 32bit int type... for more info about pow,log,exp and bin search see:
Power by squaring for negative exponents
Note that u32_log2 is not really needed (unless you want bigints) you can use constant bitwidth instead, also some CPUs like x86 has single asm instruction returning the same much faster than for loop...
Now the next step is exploit the fact that the u32_pow bin search is the same as the u32_log bin search so we can merge the two functions and get rid of one nested for loop completely improving complexity considerably like this:
//---------------------------------------------------------------------------
DWORD u32_pow(DWORD a,DWORD b) // = a^b
{
int i,bits=32;
DWORD d=1;
for (i=0;i<bits;i++)
{
d*=d;
if (DWORD(b&0x80000000)) d*=a;
b<<=1;
}
return d;
}
//---------------------------------------------------------------------------
DWORD u32_log2(DWORD a) // = ceil(log2(a))
{
DWORD x;
for (x=32;((a&0x80000000)==0)&&(x>1);x--,a<<=1);
return x;
}
//---------------------------------------------------------------------------
DWORD u32_log(DWORD b,DWORD a) // = ceil(logb(a))
{
const int _bits=32; // DWORD bitwidth
DWORD bb[_bits]; // squares of b LUT for speed up b^x
DWORD x,m,bx,bx0,bit,bits;
// edge cases
if (b< 2) return 0;
if (a< 2) return 0;
if (a<=b) return 1;
// max limit for x where b=2, all other bases lead to smaller x
bits=u32_log2(a);
// compute bb LUT
bb[0]=b;
for (bit=1;bit< bits;bit++) bb[bit]=bb[bit-1]*bb[bit-1];
for ( ;bit<_bits;bit++) bb[bit]=1;
// bin search x and b^x at the same time
for (bx=1,x=0,bit=bits-1,m=1<<bit;m;m>>=1,bit--)
{
x|=m; bx0=bx; bx*=bb[bit];
if (bx>=a){ x^=m; bx=bx0; }
}
return x+1;
}
//---------------------------------------------------------------------------
The only drawback is that we need LUT for squares of b so: b,b^2,b^4,b^8... up to bits number of squares
Beware squaring will double the number of bits so you should also handle overflow if b or a are too big ...
[Edit2] more optimization
As benchmark on normal ints (on bigints the bin search is much much faster) revealed bin search version is the same speed as naive version (because of many subsequent operations except multiplications):
DWORD u32_log_naive(DWORD b,DWORD a) // = ceil(logb(a))
{
int x,bx;
if (b< 2) return 0;
if (a< 2) return 0;
if (a<=b) return 1;
for (x=2,bx=b;bx*=b;x++)
if (bx>=a) break;
return x;
}
We can optimize more:
we can comment out computation of unused squares:
//for ( ;bit<_bits;bit++) bb[bit]=1;
with this bin search become faster also on ints but not by much
we can use faster log2 instead of naive one
see: Fastest implementation of log2(int) and log2(float)
putting all together (x86 CPUs):
DWORD u32_log(DWORD b,DWORD a) // = ceil(logb(a))
{
const int _bits=32; // DWORD bitwidth
DWORD bb[_bits]; // squares of b LUT for speed up b^x
DWORD x,m,bx,bx0,bit,bits;
// edge cases
if (b< 2) return 0;
if (a< 2) return 0;
if (a<=b) return 1;
// max limit for x where b=2, all other bases lead to smaller x
asm {
bsr eax,a; // bits=u32_log2(a);
mov bits,eax;
}
// compute bb LUT
bb[0]=b;
for (bit=1;bit< bits;bit++) bb[bit]=bb[bit-1]*bb[bit-1];
// for ( ;bit<_bits;bit++) bb[bit]=1;
// bin search x and b^x at the same time
for (bx=1,x=0,bit=bits-1,m=1<<bit;m;m>>=1,bit--)
{
x|=m; bx0=bx; bx*=bb[bit];
if (bx>=a){ x^=m; bx=bx0; }
}
return x+1;
}
however the speed up is just slight for example naive 137 ms bin search 133 ms ... note that faster log2 did almost no change but that is because how my compiler is handling inline asm (not sure why BDS2006 and BCC32 is very slow on switching between asm and C++ but its true that is why in older C++ builders inline asm functions where not a good choice for speed optimizations unless a major speedup was expected) ...