Number of digits of float value

[Marc Schellens]

Is there a standard way of determining the number of digits
a float needs (base 10) to be printed (the integer part)?

log10 is too expensive.
As is a loop dividing by 10.

I could think of something dividing by 10^numeric_limits::max_exponent10/2,
then dividing by
10^numeric_limits::max_exponent10/4,
....

but I don't want to reinvent the wheel as this seems to be
a very basic/often needed function.

thanks,
marc

[Ivan Vecerina]

"Marc Schellens" <m_*********@hotmail.com> wrote in message
news:c6**********@news.riken.go.jp...

Is there a standard way of determining the number of digits
a float needs (base 10) to be printed (the integer part)?

log10 is too expensive.
As is a loop dividing by 10.

I could think of something dividing by 10^numeric_limits::max_exponent10/2, then dividing by
10^numeric_limits::max_exponent10/4,
...

but I don't want to reinvent the wheel as this seems to be
a very basic/often needed function.

Not needed that often if you use e.g. stringstream.

In C, many will simply use sprintf and strlen: first print
it to a temporary buffer, then copy it to a final destination.

Are log10 or a loop really too expensive?

Be careful as long as you are working with a floating point
value, subtle rounding errors could happen in rare cases
and lead to nasty bugs.
If you know the range of values that can occur, and can
convert the float to an integer, and performance really is
critical, you could use something like (for int val in 0..9999):
int numDigits = (val>=100) ? ( val>=1000 ? 4 : 3 )
: ( val>=10 ? 2 : 1 );

hth,
Ivan
--
http://ivan.vecerina.com/contact/?subject=NG_POST <- e-mail contact form

[Marc Schellens]

Ivan Vecerina wrote:

"Marc Schellens" <m_*********@hotmail.com> wrote in message
news:c6**********@news.riken.go.jp...

Is there a standard way of determining the number of digits
a float needs (base 10) to be printed (the integer part)?

log10 is too expensive.
As is a loop dividing by 10.

I could think of something dividing by
10^numeric_limits::max_exponent10/2,

then dividing by
10^numeric_limits::max_exponent10/4,
...

but I don't want to reinvent the wheel as this seems to be
a very basic/often needed function.

Not needed that often if you use e.g. stringstream.

In C, many will simply use sprintf and strlen: first print
it to a temporary buffer, then copy it to a final destination.

Are log10 or a loop really too expensive?

Its for an interpreter (GDL - gnudatalanguage)
Large arrays are put out this way.
Therefore I think log10 and a loop is to expensive, ie.
as there are faster ways (like in your example) I would like to
implement these.
Be careful as long as you are working with a floating point
value, subtle rounding errors could happen in rare cases
and lead to nasty bugs.
If you know the range of values that can occur, and can
convert the float to an integer, and performance really is
critical, you could use something like (for int val in 0..9999):
int numDigits = (val>=100) ? ( val>=1000 ? 4 : 3 )
: ( val>=10 ? 2 : 1 );

Anyway, I thought there might be a standard way.
Right now I implemented it with ostringstream, but that is still
expensive I guess.

Thanks anyway,
marc

[Karl Heinz Buchegger]

Marc Schellens wrote:

Anyway, I thought there might be a standard way.
Right now I implemented it with ostringstream, but that is still
expensive I guess.

I guess you are right.
Thats simply becase the ostringstream still has to do a loop
dividing by 10. But it has to do other things as well. While
your loop only counts how often it has been executed, the
loop in the stream formatting has to format the individual
digits also.

But to be 100% sure, some testing would be needed.

--
Karl Heinz Buchegger
kb******@gascad.at

[Ivan Vecerina]

"Marc Schellens" <m_*********@hotmail.com> wrote in message
news:c6**********@news.riken.go.jp...

Ivan Vecerina wrote:

Are log10 or a loop really too expensive?

Its for an interpreter (GDL - gnudatalanguage)
Large arrays are put out this way.
Therefore I think log10 and a loop is to expensive, ie.
as there are faster ways (like in your example) I would like to
implement these.

Things can be made faster, especially once you have an integer.
But would this step really be the limiting factor?

Be careful as long as you are working with a floating point
value, subtle rounding errors could happen in rare cases
and lead to nasty bugs.
If you know the range of values that can occur, and can
convert the float to an integer, and performance really is
critical, you could use something like (for int val in 0..9999):
int numDigits = (val>=100) ? ( val>=1000 ? 4 : 3 )
: ( val>=10 ? 2 : 1 );

Anyway, I thought there might be a standard way.
Right now I implemented it with ostringstream, but that is still
expensive I guess.

Note that on many platforms, C++ streams still are significantly
slower than sprintf. And this will have a much greater impact
than the digit counting.
An integer-to-digits routine is also easy enough to implement,
and can be made faster by going backwards (from the end of the
buffer upwards).

If you provide more context around what you are trying to optimize,
I am sure that we can find more effective ways to improve
performance (than by working on digit counting).
Kind regards,
Ivan
--
http://ivan.vecerina.com/contact/?subject=NG_POST <- e-mail contact form
C++ MVP for Brainbench - www.brainbench.com