Appunti per l'esame di Advanced Computer Architectures - parte 2

ISSUE EXECUTE WB

Structure -1

-

structural

cen cheeks

RSis FUs

and deen

Waw war Struct

renoming CDB

on

delaxing

Reu p

I

Tomaselo averlay iterations due

logs to:

of

con

Register Reneming

- Resolution

Branch

Fost

- UW-BASED

7

. SPECLATION

It ideas

combines 3 : execte

Prediction

Branch choose

1

. to

Dynamic which to

I

:

suppatime

Scheduling commit

Dynamic art-of execution

.

2 onder oder

but in

: ,

Speculation catal

to execute

3

. before solved

dependences

Is

: ore

speculation

idea

Basic nw-bord in

: execute dependent branch before

and the

to I branch

issue a

on

- ,

outcome in known

allow

to them

ander

execute to

to but commit

out of to farce

Is

- in ader

-

to until

mydote)

action commits

ony invocable (live I

- prevent on

the

dote

speculative allow it

When to RF

lorger we memory

or

is

I

a no ,

Outcome branches

of executed

speculated and speculation

if

program in as

is was

dynamic

(without speculation

Conet ,

decade

fetch

scheduling

the would

simple and

only

,

execute

not .

TO MASULO WITH ROB

REORDER BUFFER register results

that

structure hold

No Fifo

in

= food

the

uncommitted value the

Once of

ader commits

for in

result raß

I

Is .

leaves the it

ROB and register

put into

is

↑ exceptions

misprediction

specedent brandes

undo

to

cory Is an or

Not stapes but i

a

anyone :

,

Strate

EXECUTION

ISSUE COMMIT

WRITE

-

structural des monder

ben

Roo to,

Waw commit

Struct

and fee

war CDA

on

moming delagine

raw *

UW RENAMING

REMISTER structure register

that the

data physical

supplies the

RENARING Cimple number of

MAP = that

existen register

requested architectural

currently to the

coregends table

update permanently to indicate the physical

the that

instruction commit renoming

= register the architectural

destination actual

holding corspends

value the

to

register

Use in-ader

to commit

Rob enforce

· EXPLICIT RENAMING

RECUSTER

Allocate destination wites

register

physical that

for evers

· new .

a

Removes all chance hotads

of

· war waw

or

Physical not

register by

used

· becomes active

when Es

fee any

Renome register

all decode

file

physical register during

into

· single

a reation

registers

committed in

FUs holding

write and

rad

· file

and temporary

from register

Commit data

to (mo

register

redates

only architectural physical

of

· maggine

movement)

When have have

commit

· remefing

we a

we

a

It with

be scareboard

used

· can registen than

plasial by

have ta

needed (layer nwl

>

- ISA su e

REGISTER ROB

RENAMING

S .

the struction then

commit with

simpler roß

registes

-Decollocativa complex

mare

architectural registers

Dynamic design

complicates

to physical and

of

- magina

debuggieq brand prediction

and

NB needed

exceptions

for precise

: is

roß

STATC SCHEDULING METHODS

reschedule cade

the

Simple code motion : unrolled

gets

the execute

to

lovs ader

Loop immer

unrolung more

in

: itmations at ance

langer cade

- starting closing

to cost

due

losing and

performance of

-

- each itmations (startug and overhead

wind down

PIPELINE the reduce

SOFWARE to

be operations and

pipelined

: in hase

program can

the staty

cost iteration

wind

and from

down phases

of ance per

to loof branche

ance have

per con &

GLOBAL schedules

Technique that

Code trases

TRACE SCHEDU4NG :

SCHEDUNG : blows that

bosic

12009-fee of

of

sequence sequence

ore

probability

straight order decessing of

Is)

branch

man of

in

executed Algorithm

being follows

wors as :

· frequent

(most branch

blos

of path

1 basic

picc sequence

a

.

2 poths

find

profiling branch

feedback to

Use commen

. the

schedule of

troe

i ance

lode branches

to aut

handle

add troe

of

sumping

# daes supert

not loops unwelling

loof to

>

- needed avencore

this limitations

Restrictions motion

cade trace

in

on a

datofla must charpe data/control

mot dependences

(maintain

- exception behavior must be possed solutions

- *

(haperblo scheduling)

predicate

of

1 use Is

. speculative Ispeculative scheduling)

of and

2

. speck

Is

use the

lative brach

before

I

more on

USPERBLOC SCHEDULING

SUPERBLOCH SCHEDUCING

SPECULATIVE SCHEALING

TRACE

VLIW ARCHITECTURE

(Very (6-16)

VLIW mode

architecture

words to

Long I time

fetch at

Is

mar

= a

There central flow

/simple

PC

only 1

· is not

parallelism af

Explicit (padblism found time)

time

compile

· un

,

mordering not

comple

(fors

performed by

No real

nw

I

· on on

,

poroped

at into

Operations time

that the

begin

must simple I

vaw

· some or a

The schedules (exploit

execution

parallel (P)

to

compiler

· maximite

+ dependency

Is home con

very

complexity

low

+ no

food parallelism

detect

+ compiler con

Easy

+ incas the

to #FU

Huge #registe

- data transport

Large capacity

- Cade

lange size

- EXECUTION

SUPERSCALAR

Fetching multiple (energe insefficient

das sch

Is inse

- =

per

mar

Sequential pipelined) CP171

(no scheduline

Pipelining dynamic

and (P 2

=

1

sugardor CP

20122 -

>

- enc ISSUE-width

MODERN SUPER SCALAR completion

multiple

A

MODERN MICROARCHITECTURES SCHEDULING MULTIPLE ISSUE+ SPECULATION

BINAMIC

:

Ansian pipeline ECC

mydate

.

1 RSis and in Is/

= 2

logic

2

. handle

to

Design between

possible dependencies Is

ony

Bottleneen logic

Issue

= superscolar

EXAmple Tomeslo

=

LIMITTS TO ILP realized

be

pr

a no can

Assumptions ideal machine

for on :

Register

.

1 vintal

infinite register (no wawtwar

renoming :

prediction at

Branch un time

it's

perfect

can't because

misprediction

2 done

be

-

: no

. (mot compile time)

Juny perfectly

prediction predicted

3

. Jump ore

:

4 dies

Menay-address esic

and

, lotency Unlimited #Is

.I

3 insed

all

for

cycle I per

,:

All Catencies depending

and noces

Fus of izame

sequence

ony

1 Is con on successive

=

Evagh replicated all ready

to allow

Fus to

Is issue

Today Window size

:

between and

32 ,

200

to 2400

own comparisons

up

Limits todasi

of CPUs : define

#FUs that

(exeute) the rected

#Is

3 be

· maximum isue

can or

,

#busses committed

=> the

(committ than the

smaller

· much

in ce

some is

# rister

posts the

· Window

fil

for size

CURRENT SUPERSCALAR dynamically

state-of-the-ort

Commercial aled scheduled supersdor

for purpose= processor

Guant implementations Pouerpa Alpha

Intel Coll supersdor

cor Mips Spars

i

= , ...

, ,

,

↑ embedded

VLIC successful os

processes de electronic

media for

pocessor consumer devices

.

8 HANDLING

EXCEPTION events requesting

INTERRUPT that the execution

moud

special by

alter frequen

=

attention

the They be

the

of :

can

processer

.

Assmahona -

extend event

by

- : an interna

Samchenaus by event

exceptions)

- : an

PRESSE there instruction

for before

which all

single

If I

INTERRUPT is a

= that their

have state

committed following have

and Is

one no

restart at internet

modified state point

the

-

any

Exceptions divided into classes

are : invoking

Asanchanaus Internet Handler

the

Suchamau I

>

User requested corred the interat

addresses

processor

-

User User the

Non-moskoble

roskable & PC

some

u

~

- Between register

status

Within rad a

- Som

Resume Terminate special indirect register

use jump

a

-

techniques temet Handler

to speculate

3 exceptions ·

on

>

- :

prediction mechanism

- cheek prediction mechanism

- mechanism

recovery

-

GPU

It's elaborations

quaphics

accelator for

& an

Giu accelerates functions

computino having

do general-purpose similar

· characteristics computing

to graphic :

Extensive parallelism

data

- control

Few Is

- operations

moth

Cong

- Feu sanchronization

- parallelism

Some tosk

- accelente

GPUs designed rendering

to

· 3D

ar

Loc be scheduled

Multiple theods

& simple

can on car

a

Throoh str

the come

we t different strom

to I

simple ren

I

it fetch decade unit

has and

and multiple Al

cor own ,

,

Conge register file

SIMT execution

simultaneous

ARCHITECTURE for the

Hw responsible of

is I

some

= data

different

stream ALs

an sind

an dato elements

the to

When opplies