PDSP16116

9

OSEL1 :0

The outputs from the device are selected by the OSEL0

and OSEL1 instruction bits. These controls allow selection

of the output combination during the current cycle (they are

not registered). There are four possible output configurations

that allow either complex outputs of the most or least signifi-

cant bytes, or real or imaginary outputs of the full 32-bit word

(see Table 4). OSEL0 and OSEL1 should both be tied low

when in BFP mode.

BFP MODE FFT APPLICATION

The PDSP16116 may be used as the main arithmetic unit of

the butterfly processor, which will allow the following FFT bench-

marks:

G 1024-point complex radix 2 transform in 517µs

G 512-point complex radix 2 transform in 235µs

G 256-point complex radix 2 transform in 106µs

In addition, with pin MBFP tied high, the BFP circuitry within

the PDSP16116 can be used to adaptively rescale data through-

out the course of the FFT so as to give high-resolution results.

The BFP system on the PDSP16116 can be used with any vari-

ation of the radix 2 decimation-in-time (DIT) FFT, for example,

the constant geometry algorithm, the in-place algorithm etc. An

N-point Radix 2 DIT FFT is split into log(N) passes. Each pass

consists of N/2 ‘butterflies’, each performing the operation:

A

B

Where W is the complex coefficient and A and B are the

complex data. Fig.4 illustrates how a single PDSP16116 may

be combined with two PDSP1601s and two PDSP16318s to

form a complete BFP butterfly processor. The PDSP16318s are

used to perform the complex addition and subtraction of the

butterfly operation, while the PDSP1601s are used to match

the data path of the A-word to the pipelining and shifting opera-

tions within the PDSP16116.

For more information on the theory and construction of this

butterfly processor, refer to application note AN59.

BFP MODE OPERATION

The BFP mode on the PDSP16116 is intended for use in the

FFT application described above, that is, it is intended to pre-

vent data degradation during the course of an FFT calculation.

The operation of the PDSP16116-based BFP buttertly proces-

sor (see Fig.4) is described below.

The Block Floating Point System

A block floating point system is essentially an ordinary inte-

ger arithmetic system with some additional logic, the purpose of

which is to lend the system some of the enormous dynamic

range afforded by a true floating point system without suffering

the corresponding loss in perlormance.

The initial data used by the FFT should all have the same

binary arithmetic weighting. In other words, the binary point

should occupy the same position in every data word as is nor-

mal in integer arithmetic. However, during the course of the FFT,

a variety of weightings are used in the data words to increase

the dynamic range available. This situation is similar to that within

a true floating point system, though the range of numbers rep-

resentable is more limited. In the BFP system used in the

PDSP16116, there are, within any one pass of the FFT, four

possible positions of the binary point wihin the integer words. To

record the position of its binary point, each word has a 2-bit

word tag associated with it. By way of example, in a particular

pass the following four positions of binary point may be avail-

able, each denoted by a certain value of word:

XX·XXXXXXXXXXXX word tag = 00

XXX·XXXXXXXXXXX word tag = 01

XXXX·XXXXXXXXXX word tag = 10

XXXXX·XXXXXXXXX word tag = 11

At the end of each constituent pass of the FFT, the positions

of the binary point supported may change to reflect the trend of

data increase or decreases in magnitude. Hence, in the pass

following that of the above example, the four positions of binary

point supported may be changed to:

XX·XXXXXXXXXXXX word tag = 00

XXX·XXXXXXXXXXX word tag = 01

XXXX·XXXXXXXXXX word tag = 10

XXXXX·XXXXXXXXX word tag = 11

This variation in the range of binary points supported from

pass to pass (i.e. the movement of the binary point relative to its

position in the original data) is recorded in the GWR. Thus, the

position of the binary point can be determined relative to its ini-

tial position by modifying the value of GWR by WTOUT for a

given word as shown in Table 6. As an example, if GWR=01001

and WTOUT=10 then the binary point has moved 10 places to

the right of its original position.

PDSP16116/A

XR

XI

YR

YI

BR

BI

WR

WI

WTA

WTB

EOPSS

SOBFP

PR

PI

PDSP1601/A

A

C

AI

AI15:13

PDSP1601/A

A

C

AR

AR15:13

PDSP16318/A

AB

CD

PDSP16318/A

BA

CD

WTOUT GWR

A

′RA′IB′RB′I

DAR

DAI

SFTA

SFTA

SFTR

SFTR

OER

OEI

Fig. 4 FFT butterfly processor