PDSP16510A MA
block floating point shifting scheme, which is discussed later.
Overflow can NEVER occur if the 3 bit option is chosen, but at
the expense of worse dynamic range.
When overflow does occur a flag is raised which can be
read by the user ( see later discussion on scale tag bits ), and
the results ignored. In addition all frequency bins are forced
to zero to prevent any erroneous system response.
Even with only 2 bit word growth poor dynamic range will
be obtained if the data is simply reduced to 16 bits, and
becomes worse when the incoming data does not fully occupy
all the bits in the word. These problems are overcome in the
PDSP16510, however, byablockfloatingpointschemewhich
compensates for any unnecessary word growth.
Fig. 5. RAM Organization with 1024 Point Transforms
RAM has been designed for use in a wide variety of applica-
tions. The provision of an asynchronous input strobe (DIS),
allows data to be loaded without the need for additional
externalbuffering.Anasynchronousoutputstrobe(DOS)also
allows transformed data to be dumped with the sampling
clock, this being particularly useful when the device is per-
forming the inverse transform back to the time domain. Inputs
and outputs are both supported by flag and enabling signals
which allow transfers to be properly co-ordinated with the
internal transform operation.
In many applications the DIS and DOS inputs can be tied
together and fed by the sampling clock. If the output rate must
be higher than the input rate, as with multiple devices support-
ing overlapped data samples, both strobes can still be con-
nected together. The clock supplied should then be twice or
four times the sampling clock, and an internal divider can be
usedtoprovidethecorrectlyreducedinputrate. Theprovision
of a separate DOS pin does, however, allow the output rate to
be asynchronous to the input rate, and therefore faster than
strictly needed. Further output processing at higher rates is
then possible if this is advantageous to system requirements.
The internal workspace is double buffered when 256
pointtransformsaretobeperformed. Aseparateoutputbuffer
is also provided. These resources, together with separate
input and output buses, allow new data to be loaded and old
results to be dumped, whilst the present transform is being
computed. Additional, external, input buffering is not needed
to prevent loss of incoming data whilst a transform is being
performed.
During each pass the number of sign bits in the largest
result is recorded. Before the next pass, data is shifted left
[multiplied by 2], once for every extra sign bit in this recorded
sample. At least one component in the block then fully occu-
piesthe16bitword,andmaximumdataaccuracyispreserved
Up to four shifts are possible before every pass after the
first,withatotaloffifteenforthecompletetransform.Attheend
of the transform the number of left shifts that have occurred is
indicated on S3:0. Lack of pins prevents a separate output
being available to indicate that overflow has occurred in the 2
bit word growth option. For this reason the maximum number
of compensating left shifts in this mode is restricted to 14.
State 15 is then used to indicate that overflow has occurred.
The first step in the butterfly calculation multiplies 16 bit
data values with 16 bit sine/cosine values, to give 18 bit
results. This increased word length preserves accuracy
through the following adder network, and has been shown
through simulations to be an optimum size for transform sizes
up to 1024 points. This is particularly true when the input data
is restricted to below 16 bits, as is necessary with practical A/
D converters with very high sampling rates. The bottom bit of
this 18 bit word is forced to logical one and as such is a
compromise between truncation and true rounding. It gives a
lowernoisefloorintheoutputscomparedtosimpletruncation.
To prevent any possibility of overflow during the butterfly
calculation the word length is allowed to grow by one bit
through each of the three adders. The least significant bit is
always discarded in the first two adders . Sixteen bits are then
chosen from the final adder in the manner discussed earlier,
and the number of sign bits in the largest result is recorded for
use in the following pass.
When block overlapping is required, internally stored
data will be re-used, and a proportionally smaller number of
new samples need be loaded. Note that the internal window
operator still functions correctly since it is actually applied
during the first pass, and not whilst data is being loaded. The
internal RAM organisation is shown in Fig. 4. It should be
noted that the amount of overlap between I/O transfers and
transformsiscompletelyunderthecontrolofthesystem,since
aninputenablesignal(INEN)andanoutput enable(DEN)can
be used to initiate transfers.
Fig. 3 shows one of the four internal data paths which can
computearadix-4butterflyintwelvesystemclockcycles. This
equatestocompletingthebutterfly in3cyclesforthecomplete
device.
In the 1024 point mode there is insufficient workspace for
DATA TRANSFERS
The data transfer mechanism to and from the internal
Fig. 4. RAM Organization with 256 Data Points
Fig. 6. 1024 Point Transforms with I/P Buffer
6
MITEL [ MITEL NETWORKS CORPORATION ]
