PDSP16510
spaced frequencies are to be detected, and one is of smaller
magnitude than the other. It does, however, reduce the actual
frequency resolution, and the Hamming window may then be
preferable.
Data in and out of the device is represented by 16 bit real
and imaginary components, with 16 bit sine and cosine values
contained in an internal ROM. Conditional scaling, coupled
with word growth through the butterfly data path, gives in-
creased dynamic range. Transforms can be computed with
sample sizes of either 256 or 1024 data points. The 256 point
option can alternatively be used to simultaneously execute
either four 64 point transforms, or sixteen 16 point transforms.
The 16 point mode can only be used with a rectangular
window, and no overlapping of data blocks is possible.
The device can be configured, either, to perform continu-
ous transforms in a real time application, or as slave processor
to a more general purpose signal processing system. In the
continuous mode, with transform sizes of 256 points or less,
it contains three internal control units which simultaneously
allow new data to be loaded, present data to be transformed,
and previous results to be dumped. Additional, external, input/
output buffering is not needed. The internal input buffer also
allows data blocks to be overlapped by either 50% or 75%,
apart from the mode with no overlaps.
When 1024 point transforms are to be calculated, without
loss of incoming data during the transform time, it is necessary
to use an input buffer. This requirement is satisfied by a single
PDSP16540 support device.
In any of the real or complex modes it is possible to obtain
higher performance by connecting devices in parallel. It is then
possible to increase the sampling rate to that of the system
clock used for internal operations.
The mode of operation of the device is controlled by 16
bits in a control register. These are loaded through the
AUX15:0 port when a control signal DEF is active low. This
port is also used to provide the imaginary component of
complex input data, and, if complex transforms are to be
performed, an external tristate buffer will be needed to isolate
the control information. This should only be enabled when
DEF is active. DEF is also used to initialise the internal
circuitry, and can be a simple power on reset if control
parameters need not be subsequently changed.
INPUT
SELECT
RAM
SIN / COS
ROM
Shift left until largest point
has one sign bit.
16
16
MULTIPLIER
S
S
29 - 14 13 - 0
"1"
18
16
FIRST ADDER
19Bit Result
18 - 1
0
REGISTER FILE
SECOND ADDER
19Bit Result
18 - 1
0
REGISTER FILE
THIRD ADDER
19Bit Result
18 - 3
17 - 2
DATA PRECISION
During each pass of a radix-4 fast Fourier transform it is
possible for either component of a particular result to grow by
a factor of up to four in the first pass, and 5.242 in subsequent
passes. This is between two and three bits in each pass and
the data path must allow for this word growth to avoid any
possibility of overflow. At the end of the data path the word is
again reduced to 16 bits by discarding least significant bits.
Any un-necessary word growth to prevent overflow thus
results in loss of arithmetic precision, and has a detrimental
effect on the dynamic range achievable.
In practice these large word growths only occur when
bipolar complex square waves are transformed, and even
then will not occur on every pass. The PDSP16510 compro-
mises by allowing a 2 bit word growth during the butterfly
calculation in the first pass. This is equivalent to ignoring the
most significant bit of the 19 bit final result, which is assumed
to be an extra sign bit, and then selecting the next 16 bits for
CR
BIT3
SELECT
Fig. 3 One of Four Data Paths
storage. In subsequent passes a Control Register Bit allows
the user to continue to select these 16 bits, or instead to use
the 16 most significant bits. The latter option is equivalent to
a 3 bit word growth. The 2 or 3 bit word growth option applies
to ALL subsequent passes and is not a per pass option.
If the 2 bit option is selected there is a possibility of
overflow occurring in one of the passes. The prediction of
overflow is mathematically difficult, and only occurs with
specific complex square waves. Scaling down the inputs
cannot be guaranteed to prevent overflow because of the
4
MITEL [ MITEL NETWORKS CORPORATION ]
