Sami YEHIA and Olivier TEMAM LRI, Paris South University France

From Sequences of Dependent Instructions to Functions: A

Complexity-Effective Approach for Improving Performance Without ILP

or Speculation

Sami YEHIA and Olivier TEMAMLRI, Paris South University

France

2/18

Scaling Up Processors

Larger pipelines, caches, instruction windows and reservation stations

Aggressive speculation mechanisms : branch prediction, value prediction, data prefetching..

Rely on ILP exploitation What about scaling with little ILP?

3/18

Concept

264*num_registers input!

(Theoretically)

……addq r1,r2,r3subq r3,10,r4……sll r5,6,r6addq r5,r5,r4

Programr1 r2 r3 rn

r6 = f1(r1,r2,…,rn) r4 = f2(r1,r2,…,rn)

Logic circuit

r163 r162 r161 r11 r10

f163 f162 f161 f11 f10

Combinatorial Functions

A sequence of instructions is a set of functions

4/18

Principles

An « independent » Function for each output

fr3(r9,r10) = r9 + r10 – 1fr4(r9,r10) = sign_extension(r9 + r10 – 1)31:0fr5(r9,r10) = ((r9 + r10 – 1)<<1) >> 1fbr(r9,r10) = (r9 + r10 – 1) ((r9 + r10 – 1)<< 1)>>1)

DFG

5/18

Hardware Operator

+

+

a b

out

c

f1

f1i = f’(ai,bi,cout1i-1)cout1i =f’c(ai,bi,cout1i-1)outi = f’’(f1i,ci,cout2i-1) = f’’(ai,bi ,ci,cout1i-1,cout2i-1)cout2i = f’’c(ai,bi ,ci,cout1i-1,cout2i-1)

Eliminate dependencies to calculate a+b+c

r10 + r9 –1 to hardware operators

6/18

Complexity Effectiveness

Scalability of ILP Vs. Functions

Complexity

Performance

ILP exploitation

Functions

7/18

Related Work

ASIC General-Purpose context

• 3-1 Interlock Collapsing ALU [Y. Sazeides, S. Vassiliadis and J. Smith, Micro’ 29, 1996]

• Chimaera [Z. YE et al., ISCA’ 27, 2000]

• Grid Processors [R. Nagarajan et al., MICRO’ 34, 2001]

• Cascade one or more hardware operators to execute specific functions

AND OR XOR

AND OR XOR

Adder

8/18

Building Functions

From traces of instructions to configuration macros compilation toolchain to study:• Potential of the approach• Performance analysis on a superscalar processor

Traces

9/18

Potential of the Approach

Cuts : limits to DFG collapsing (height)• Number of inputs• Non-collapsable instructions• Load instructions (27,7 %)• Carries from upper significant bits

Theoretical speedup

The lower the ILP the higher speedup

op

op

LD

op

op

memF2

mem

F1@

op

Cut

@

10/18

Theoretical Speedup

1

1,2

1,4

1,6

1,8

2

2,2

2,4

2,6

gzip vpr gcc mcf crafty parser eon perlbmk gap vortex bzip2 twolf AVG

Th

eore

tica

l S

pee

du

p

16 inst.

32 inst.

64 inst.

128 inst.

256 inst.

512 inst.

1024 inst.

11/18

Number of Inputs

0

10

20

30

40

50

60

70

80

90

100

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40Number of inputs

Per

cent

age

of N

umbe

r of

Exe

cute

d F

unct

ions

All inputs

Register inputs

1

1,2

1,4

1,6

1,8

2

2,2

2,4

2,6

gzip vpr gcc mcf crafty parser eon perlbmk gap vortex bzip2 twolf AVG

Th

eo

reti

ca

l Sp

ee

du

p

5 inputs

10 inputs

20 inputs

30 inputs

40 inputs

12/18

Non Collapsable Instructions

0

10

20

30

40

50

60

70

80

90

100

gzip vpr gcc mcf crafty parser eon perlbmk gap vortex bzip2 twolf AVG

Pe

rce

nta

ge

of T

ota

l Exe

cute

d In

stru

ctio

ns

13/18

Implementation

rePlay Framework

14/18

Performance Evaluation

1

1,1

1,2

1,3

1,4

1,5

1,6

1,7

gzip vpr gcc mcf crafty parser eon perlbmk gap vortex bzip2 twolf AVG

Spe

edup

Global Speedup

Local Speedup

15/18

RePlay Optimization Engine Delay

0,9

1

1,1

1,2

1,3

1,4

1,5

1,6

1,7

gzip vpr gcc mcf crafty parser eon perlbmk gap vortex bzip2 twolf AVG

Loca

l Spe

edup

10 cycles

1000 cycles

10000 cycles

Function built “offline”

16/18

Latency of Function units

0,8

0,9

1

1,1

1,2

1,3

1,4

1,5

1,6

1,7

gzip vpr gcc mcf crafty parser eon perlbmk gap vortex bzip2 twolf AVG

Loca

l Spe

edup

1 cycle2 cycles3 cycles

17/18

Future Work

Address prediction to overcome Load cuts

Address Prediction& Cache Preloadingop

op

LD

op

op

mem

F2

mem

F1@

op

op

op

LD

op

op

mem

@

op

@’

F1

@

LD

@’

F2

mem

18/18

Q & A

Carries from Upper Significant Bits

1

1,2

1,4

1,6

1,8

2

2,2

2,4

2,6

2,8

gzip vpr gcc mcf crafty parser eon perlbmk gap vortex bzip2 twolf AVG

Th

eo

retic

al S

pe

ed

up

Cuts due toUpperSignificantCarries

IgnoringCuts due toUpperSignificantCarries

Optimization Engine Delay

0,9

1

1,1

1,2

1,3

1,4

1,5

1,6

1,7

gzip vpr gcc mcf crafty parser eon perlbmk gap vortex bzip2 twolf AVG

Glo

ba

l Sp

ee

du

p

10 cycles

1000 cycles

10000 cycles

Latency of Function units

0,8

0,9

1

1,1

1,2

1,3

1,4

1,5

1,6

1,7

gzip vpr gcc mcf crafty parser eon perlbmk gap vortex bzip2 twolf AVG

Glo

bal S

peed

up

1 cycle2 cycles3 cycles