Elenco Electronics MO-1251 Manual de usuario descargar pdf (Pagina 79)

100

Pipelining

Pipelining gets its name

from

the fact that several processors

are connected serially, like sec-

tions

pipe, with

the output of

the first tied to

the

input

of the

second,

whose

output is tied to

the input of

the third. and so

forth.

Data

that enters

the

pipeline is acted upon by each

processor

as it travels through

the

pipeline.

Analogous to

pipelining

is the

operation of an assembly line. An

assembly line is divided into sev-

eral

stages, at each of which a

particular operation takes place.

One station might install the en-

gine into the body, for example,

while the next puts on

door.

the

partially

assembled car

travels down the

line, more and

more items are added until, fi-

nally,

a finished product emerges

at the end. By

advancing the

ve-

hicle as each step is completed

and replacing it with

a vehicle

from

the previous station, no as-

sembly

station is

idle.

Therefore,

it is possible

to turn out a steady

stream of finished cars at a rate

that is much faster

than

each

had been individually

crafted.

The same

strategy can be ap-

plied

to many computer func-

tions-

floating -point arithmetic,

for example. 1b

see how pipelin-

ing

can improve

performance,

let's

compare

the methods used

by a microprocessor

and a math

coprocessor

to multiply two

numbers.

Typically, the CPU accom-

plishes the Job

(through soft-

ware)

by (1) splitting each

number into

an exponent and a

mantissa, (2) adding

the expo-

nents,

(3) storing the result in

temporary memory, (4)

multiply-

ing

the

mantissas, (5) fetching

the exponent, and (6)

combining

the two into

final

answer. Only

after the entire

sequence has run

its course

can the CPU

start on

another

calculation.

The

math coprocessor, on

the

other hand, multiplies

the two

numbers

using a multi -stage

pipeline, as shown

in Fig. 3. The

two

numbers

enter the pipeline's

first

stage,

where

they are identi-

fied

as to exponent

and man-

tissa.

The elements

are then

shifted to

the second stage,

INPUT

COMPARATOR

AND SELECTOR

LARGER

EXPONENT

FACTOR

RIGHT

SHIFTER

MULIIPLIER

SUPPRESS

ZEROS

ADDER

CORRECTED

EXPONENT

LEFT

SHIFTER

NORMALIZED

FRACTION

EXPONENT

DIFFERENCE

Fig.

3. BLOCK

DIAGRAM of a plpelined math coprocessor.

Coprocessor

Or Accelerator?

One term loosely bandied

about the PC

industry is the

coprocessor

accelerator board,

a name generally

given to adapt-

er cards

that contain high-

speed CPU chips

increase

sys-

tem

throughput. Where the

term

originated

is anybody's

guess, but most of

those boards

are not

coprocessors;

they're

simply accelerators.

The

difference is in the way an

accelerator

board operates. To

install an

accelerator card in

most cases you must

remove the

system's old

CPU and run

jumper

cable from

the acceler-

ator card

to the now -empty CPU

socket.

Notice

that the two

pro-

cessors don't work together

si-

multaneously;

at best you

may

switch between them to

accom-

modate

temperamental soft-

ware applications.

By contrast.

to achieve true

parallel process-

ing, both

CPU's must be accessi-

ble by software at

the all times -

something the typical

acceler-

ator card simply

cannot do.

where

the mantissas are weight-

ed for

processing.

The third stage

multiplies the mantissas

and

the

fourth strips the

fraction

un-

necessary zeros; the

fifth and

final stage adds the exponents

and converts the results to con-

ventional

scientific notation.

The coprocessor really picks

speed, because after the first pair

numbers is

shifted

from the

first stage to the second, another

pair can enter the now -empty

first stage. In fact, with each suc-

cessive

movement

data be-

tween stages, another pair of

numbers can enter the pipeline.

causing it to fill. After the fifth

shift, results begin pouring out

of the pipeline.

Of course, pipelining is not

limited to Intel's math

coprocessors; a pipeline can

also

be created using several trans -

puters. However, pipelining

only efficient when the number of

similar computations to be per-

formed is large, as in floating -

point arithmetic. Instructions

are still executed one step at a

time,

and

problems that

rely on

the results of one operation be-

fore invoking another are at a dis-

advantage. For that reason,

designers developed

number

other architectures for increas-

ing performance.

1 2 ... 74 75 76 77 78 79 80 81 82 83 84 ... 99 100

Comentarios a estos manuales

Sin comentarios

Elenco Electronics MO-1251 Manual de usuario Pagina 79

Comentarios a estos manuales