APPLICATIONS INFORMATION
Device Requirements
The number of devices required to implement a given convolver
window depends on the size of the window, the required pixel
rate, and whether the pixel accuracy is to be 8 or 16 bits. In
practice the PDSP16488A supports windows requiring one, two,
four, six, or eight devices without additional logic. Table 2 gives
typical window sizes which may be obtained with the above
number of devices.
In an interlaced system the pixels from the previous field must
use the IP7:0 inputs, and the live pixels must use the L7:0 inputs.
Interlaced systems requiring extended precision pixels are non-
supported with a single device, since the L7:0 inputs are then use
for the least significant 8 bits, and the IP7:0 inputs for any more
significant bits.
If the X15 pin is left open circuit, an internal pullup will
configure the device in the host supported mode. The host must
Figs. 13 through 20 show system interconnections for these
arrangements. Other configurations are possible but may need
the support of additional pixel/line delays and/or expansion
adders. Although not necessarily shown, all configurations can
be supported by either an EPROM or a Host computer. Interlaced
or non-interlaced video may also be used, unless explicitly stated
otherwise in the text.
Expansion with 8-bit pixels is a straightforward process and
the number of devices needed is easily deduced from the window
sizes available in a single device. At pixel rates above 20MHz it
may not be practical to use more than four devices, since the full
32-bit intermediate precision is not available. The lack of expan-
sion multiplexing reduces the intermediate precision to 16 bits.
The partial sum outputs must thus not overflow these 16 bits; this
will require the coefficients to be scaled down appropriately with
a resulting loss in accuracy.
Expansion with 16-bit pixels can be achieved in several ways.
The simplest way is to use two devices, each working with 8-bit
pixels. One device handles the least significant part of the data,
and its output feeds the expansion input of a second device. This
performs the most significant half of the calculation. The least
significant half is then added to the most significant sum, after the
latter has been multiplied by 256, i.e. shifted by eight places. This
shift is done internally and controlled by Register D, bit 1. The
internal 32-bit accuracy prevents any loss in precision due the
shift and add operation.
The window size with this arrangement is restricted to that
available in a single device, at the required pixel rate but with
8-bit pixels. Thus two devices can be used, for example, to
provide an 838 window with 16-bit pixels and 10MHz rates.
If a larger extended precision window is needed, it is possible
to use four devices. Each device is then programmed to be in a
16-bit data mode, but should be restricted to rates below 20MHz,
if the 32-bit intermediate precision is to be maintained. In the 16-
bit modes, however, the output from the last line delay is not
available due to pin limitations. This is not a problem in a four
device interlaced system, since half of the devices will be fed from
an external field delay. In non interlaced systems additional
external line delays would be needed. An alternative approach would
be to configure all the devices in the appropriate 8-bit mode, do
separate least significant and most significant calculations, and then
combine the results in an external adder after a wired-in shift.
then supply a data strobe and an
control line. The X7:0 pins
R/W
must be connected to the host data bus, and are used to both load
PROG CE
and read back register values. The
connected together, and then driven by a host address decode.
REPLY
and
pins may be
The output on
, which provides a
to the host, need
PC1
not be used if the width of the data strobe is greater than the
maximum tEXP value given in Fig. 10.
The configuration bits 6:4 in register A define the window size,
maximum pixel rate, and pixel resolution. Window sizes smaller
than the maximum in any configuration are implemented by filling
in the window with zero coefficients. Bits 3:0 are irrelevant in the
Single mode, as is bit 7 if the gain control is used.
The result would be expected to lie in either the bottom 20 bits
of the 32-bit result , or possibly in the next 20-bit field displaced
by four bits. Register C, bits 5:4, must thus select one of these
fields for subsequent use by the gain control. The gain is then
adjusted such that the 16 outputs available on pins D15:0 are in
fact the 16 most significant bits of the result. The gain needed is
application specific, but if too much gain is used the OVR pin will
go high to indicate an overflow.
Register B, bits 2:1, must be set to select the required method
of defining the length of the line delays, and the use of bit 3 is
dependent on any external pixel delays before the convolver
input. No additional delays are needed on the pixel inputs in a
single device system, and register D, bits 4:2, should be reset.
The pipeline delay in the DELOP output path should match one
of those in Table 6, and is window size dependent.
Dual device configurations
Two devices, each configured with 8-bit pixels and 8W34D
windows, can be used to provide an 838 window at up to 20MHz
pixel rates. Fig. 14 shows both the non-interlaced and interlaced
arrangements.
Video lines containing up to 1024 pixels are possible in both
configurations, since each device only needs four line delays.
One device is configured as the Master by grounding
MASTER
normal way and has its
circuit.
the
pin; the other then receives control signals in the
MASTER
SINGLE
pins left open
and
The internal convolver sum, in the device producing the final
result, must be delayed by 4 pixels to match the inherent delay in
the expansion output from the other device. This is actually
achieved by delaying the pixel inputs to the line stores (register
D, bits 3:2, = 01). No additional delay in the expansion input is
needed, but the pipeline delay used to produce DELOP must be
four clocks greater than that given in Table 6 for a single device.
The DELOP output is redundant in one of the two devices.
Two devices can also be used to support systems requiring 16-
bit pixels. With this approach the 1638 multiplication is realised as two
838 operations, with the results added together after the most
significant half has been shifted by 8 places to the most significant
end. This shift operation is controlled by register D, bit 1. Both
convolvers are programmed to contain the same coefficients. The
convolved output can theoretically grow to 30 bits, and the
appropriate field must be selected before using the gain control.
Examples of this operating mode are shown in Fig. 15. Each
device must be configured in the same 8-bit pixel operating mode,
but the device producing the final result must use the 8 place shift
option on its internal sum.
Single device configuration
Fig.13 illustrates both EPROM and Host supported single
device systems, with or without interlaced video. In both cases
the Single and X15 pins must be tied low, and the
,
,
PC0 PC1
PROG
pin becomes an output
DS
and
pins are redundant. The
and indicates that a register load sequence is occurring. The first
line delay must always be bypassed in a non interlaced system,
however, since an internal pullup is not provided, the BYPASS pin
should be tied to VCC for the correct operation. With interlaced
video the BYPASS input is used to distinguish between the odd
and even fields.
CE
The
simply loaded after a power on reset signal; the latter being
RES CE
input may be used
input may be left open circuit if coefficients are to be
applied to the
input. Alternatively the
to change the coefficients at any time after power on reset; the
EPROM would then need additional address bits for the extra
sets of coefficients that are to be stored.
21
MITEL [ MITEL NETWORKS CORPORATION ]
