Analog Devices ADSP-BF53x Blackfin, Reference

Page: 1001 / 1042

ADSP-BF53x/BF56x Blackfin Processor Programming Reference D-7
Numeric Formats
Typically, block floating-point format allows you to shift out non-signifi-
cant MSBs (most significant bits), increasing the precision available in
each value. Block floating-point format can also be used to eliminate the
possibility of a data value overflowing. See
Figure D-4 . Each of the three
data samples shown has at least two non-significant, redundant sign bits.
Each data value can grow by these two bits (two orders of magnitude)
before overflowing. These bits are called guard bits.
If it is known that a process will not cause any value to grow by more than
the two guard bits, then the process can be run without loss of data. Later,
however, the block must be adjusted to replace the guard bits before the
next process.
Figure D-5 shows the data after processing but before adjustment. The
block floating-point adjustment is performed as follows.
• Assume the output of the
SIGNBITS
instruction is
SB
and
SB
is used
as an argument in the
EXPADJ
instruction. Initially, the value of
SB
is +2, corresponding to the two guard bits. During processing, each
resulting data value is inspected by the
EXPADJ
instruction, which
counts the number of redundant sign bits and adjusts
SB
if the
Figure D-4. Data With Guard Bits
Sign Bit
2 Guard Bits
0x0FFF = 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1
0x1FFF = 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1
0x07FF = 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1
To detect bit growth into two guard bits, set SB = –2