|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
ä |
Bias
is chosen as follows
|
|
|
|
ä |
Let
e be the number of bits in the exponent
|
|
|
|
ä |
bias =
2(e-1)-1
|
|
|
|
ä |
Gives
exponents in range -bias +1 to bias
|
|
|
|
ä |
e=8
=> bias is 127
|
|
|
|
=> min exponent
= -126 = 00000001
|
|
|
|
=> max exponent = 127
= 11111110
|
|
|
ä |
What
about 11111111?
|
|
|
|
ä |
Used
to represent NaN - Not a Number
|
|
|
|
ä |
Could
be caused by overflow or underflow
|
|
|
ä |
What
about 00000000?
|
|
|
|
ä |
If
mantissa is zero => the number 0
|
|
|
|
ä |
Note
this is special, since a leading one is usually implicit
|
|
|
ä |
If
mantissa is non-zero => denormalized numbers
|
|
|
|
ä |
Numbers
where leading bit is not 1
|
|
|
|
ä |
Depending
on implementation, may cause an
underflow
|
|
|
|
ä |
Too
small a number
|
|