Motorola DSP96002 User Manual

Page 734

MOTOROLA

DSP96002 USER’S MANUAL

C-11

The registers d8 and d9 are auxiliary registers which can be used for temporary data storage. Their main

purpose is to allow a fast, four-cycle radix-2, decimation in time FFT butterfly kernel, though their use is cer-

tainly not limited to this application. d8 and d9 can be used as source operands in multiply operations and

MOVE instructions, but can only be written as destinations of MOVE instructions.

The format conversion unit provides automatic format conversion from/to the SP and DP memory storage

formats to/from the DP storage format in the data ALUs register file. The conversion is depicted in Figure

C-10 and is done in a transparent fashion.

When moving SP numbers into the data ALU (see Figure C-10a), the 52-bit fraction of the DP internal for-

mat is written with the 23-bit fraction of the source in its most significant bits, and the implicit integer bit is

made explicit. The remaining bits of the fraction are set equal to zero. If the number in question is denor-

malized (exponent = e

min

and the first bit of the mantissa = 0), the U tag is set. In the non-IEEE "flush to

zero" mode (indicated by the FZ bit in the Status Register), the number is considered zero when used as

an operand for floating-point operations, although the contents of the register are not changed. In the IEEE

31 30 29

Fraction

23 22

95 94 74 73 72 71

63 62

32 31

11 10

(3)

Fraction

X or Y Data Memory

i = 1 when normalized
i = 0 when unnormalized

(2)

Figure C-10a. Automatic Format Conversion – Single Precision

40 39

31 30 29

Fraction

23 22

X or Y Data Memory

† †

†

(1)

Notes: * –

† – When NaN, bits 71, 72, 73 = 1
When not NaN Bit 74

↔

Bit 30

Bits 73, 72, 71 are complement

of Bit 74.

(1) – Bits 32-39 are nonzero when the register
contains a SEP floating point result or a
DP floating point number.
Bits 32-39 are zero when the register
contains a SP floating point number.

(2) – Bits 11-31 are only nonzero when the

(3) – Bits 0-10 are always zero when
representing a floating point number.