Roku Developer Program

[greubel]

I'm getting the wrong results on all box platforms.

x = Log(8192.0)/Log(2)
? "x = " Type(x) " " x
y = Int( x )
? "y = " Type(y) " " y

Int does round but on a whole number and down ?

x = Float 13
y = Integer 12

[RokuMarkn]

The result of the divide is a little less than 13, which rounds down to 12 when you convert to integer. It still prints as 13 as a float since its so close to 13. The same thing happens in C:

    float a = log(8192.0);
    float b = log(2.0);
    float x = a / b;
    int y = x;
    printf("%g %d\n", x, y);

If you printf x in %A format, it's 0X1.9FFFFEP+3 while a real 13 is 0X1.AP+3.

--Mark

[greubel]

So we never know which way int() is going to round ?
In this case, it should have been up ????

[RokuMarkn]

Int always rounds down. The problem is the number you are rounding is close to but a little less than 13.

--Mark

[RokuMarkn]

Of course if you want to round to the nearest integer rather than rounding down, you can say

y = Int(x + 0.5)

--Mark

[greubel]

That sometimes throughs the answer off.
I though that fix(x) was supposed to do that, but it doesn't work either.
There should be an exact way to do it. Either round down or round up or just chop it off.
What I'm looking for is the exponent (power of 2) of a real number as an integer.

[RokuMarkn]

I'm not sure you're understanding that the problem has nothing to do with the int conversion, it's due to the loss of precision in the log and divide operations. If you've ever taken a Numerical Analysis course, you may recall the many dangers of working with floating point numbers. In general, what you want to do is impossible for arbitrary real numbers; for example, 2^70 and (2^70)-1 are both represented by the same floating point number, so there's no way to tell that log2 of the first is 70 while log2 of the second is 69.

I would probably use something like this. It's rather ugly (and untested) but it should be more accurate than using logs.

function Log2(x as Double) as Integer
  lg = 0
  val = 1.0
  while x > val
    lg = lg + 1
    val = val * 2
  end while
  return lg
end function

--Mark

[EnTerr]

Well, hm. Python has the same precision limitations, how come it works there?

>>> from math import log
>>> 0.1 + 0.2
0.30000000000000004
>>> 2.**70, 2.**70 - 1
(1.1805916207174113e+21, 1.1805916207174113e+21)
>>> 2.**70 == 2.**70 - 1
True
>>> log(8192) / log(2)
13.0
>>> log(8192) / log(2) == 13
True

[greubel]

This appears to work for + and - floats with + and - exponents.
What I'm trying to do is to pack / unpack a "binary" real number from / to a roByteArray.
The float is done, at least it's been running for an hour, feeding the routines random values.
Now for the double.
*

function Log2(x as object) as Integer

   if Sgn(x) < 0
      x = x * -1.0
   end if
   lg = 0
   val = 1.0
   if val < x
      while x > val
         lg = lg + 1
         val = val * 2
      end while
      lg = lg -1
   elseif val > x
      while x < val
         lg = lg - 1
         val = val / 2
      end while
   end if
   return lg
   
end function

[EnTerr]

Do you really need log2 function as such?
I am asking because you are already building the mantissa (significand?) while calculating the exponent part. Intuitively feels to me the process is:

sign = sgn(x)
x = abs(x)
exp = 0
while x >= 2: x = x / 2: exp = exp - 1: end while
while x < 1: x = x * 2: exp = exp + 1 : end while
mant = int((x - 1) * 2^23 + 0.5)