DSP1620 View Datasheet(PDF) - Agere -> LSI Corporation

Part Name

Description

Manufacturer

DSP1620

DSP1627 Digital Signal Processor

Agere -> LSI Corporation 

Preliminary Data Sheet

February 1997

DSP1628 Digital Signal Processor

4 Hardware Architecture (continued)

In wideband low data rate applications, additive white

Gaussian noise (AWGN) is the principle channel impair-

ment, and Euclidean branch metric computation for

convolutional decoding is selected by resetting the

branch metric select bit to zero.

A traceback-length register is provided for programming

the traceback decode length.

A block diagram of the coprocessor and its interface to

the DSP1600 core is shown in the following figure:

EOVF

EREADY

EBUSY

ECCP

IDB

ear

edr

eir

CONTROL UNIT

ECON

BRANCH METRIC

UNIT

SiHi, i = 0, . . . ,5

ZIG10

ZQG32

G54

UPDATE UNIT

NS[63:0]

PS[63:0]

SYC

MDX

MACH

MACL

RAM4

TRACEBACK UNIT

TBLR

DSR

TBSR

5-4500 (F)

Figure 7. Error Correction Coprocessor Block

Diagram/Programming Model

The ECCP internal registers are accessed indirectly

through the address and data registers, ear and edr.

The control register, ECON, and the traceback length

register, TBLR, are used to program the operating

mode of the ECCP. The symbol registers (S0H0—

S5H5, ZIG10, ZQG32), the generating polynomial reg-

isters (ZIG10, ZQG32, G54), and the channel impulse

registers (S0H0—S5H5) are used as input to the ECCP

for MLSE or convolutional decoding. Following a Viterbi

decoding operation, the decoded symbol is read out of

the decoded symbol register, DSR. All internal states of

these memory-mapped registers are accessible and

controllable by the DSP program. During periods of si-

Lucent Technologies Inc.

multaneous DSP-ECCP activity, however, ECCP inter-

nal edr registers as well as the shared bank of RAM,

RAM4, are not accessible to the user's DSP code.

Branch Metric Unit: The branch metric unit of the

ECCP performs full-precision real and complex arith-

metic for computing 16-bit incremental branch metrics

required for MLSE equalization and convolutional de-

coding.

MLSE Branch Metric Unit: To generate the estimated

received complex signal at instance n, E(n, k) = EI(n, k)

+ j EQ(n, k), at the receiver, all possible states, k = 0 to

2C – 1 – 1, taking part in the Viterbi state transition are

convolved with the estimated channel impulse re-

sponse, H(n) = [h(n), h(n – 1), h(n – 2), . . . , h(n – C +

1)] T, where the constraint length C = {2 to 6}. Each in-

phase and quadrature-phase part of the channel tap,

h(n) = hI(n) + j hQ(n), is quantized to an 8-bit 2's com-

plement number.

The channel estimates are normalized prior to loading

into the ECCP such that the worst-case summation of

the hI(n) or hQ(n) are confined within a 10-bit 2's com-

plement number. The in-phase and quadrature-phase

parts of the received complex signal Z(n) = ZI(n) + j

ZQ(n) are also confined within a10-bit 2's complement

number. The Euclidean branch metric associated with

each of the 2C state transitions is calculated as:

BM(n, k) = XI(n, k)2 + XQ(n, k)2

where

XI(n, k) = abs{ZI(n) – EI(n, k)}

and

XQ(n, k) = abs{ZQ(n) – EQ(n, k)}

The absolute values of the difference signal are saturat-

ed at level 0xFF. The sixteen most significant bits of this

17-bit incremental branch metric are retained for the

add-compare-select operation of the Viterbi algorithm.

The in-phase and quadrature-phase parts of the re-

ceived complex signal are stored in ZIG10 and ZQG32

registers, respectively. The complex estimated channel

taps H5 through H0 are stored in S5H5 through S0H0

registers, such that the in-phase part of the channel oc-

cupies the upper byte and the quadrature-phase part of

the channel occupies the lower byte.

Convolutional Branch Metric Unit: Two types of distance

computation are implemented for convolutional decod-

ing. Convolutional decoding over a Gaussian channel is

supported with Euclidean distance measure for rate 1/1

and 1/2 convolutional encoding. Convolutional decod-

ing preceded by the MLSE equalization or other linear/

nonlinear equalization is supported with Manhattan dis-

tance measure for rate 1/1 through 1/6 convolutional

encoding.

27

Share Link: