The O-GEHL branch predictor

Download Report

Transcript The O-GEHL branch predictor

Looking for limits in branch prediction
with the GTL predictor
André Seznec
IRISA/INRIA/HIPEAC
1
André Seznec
Caps Team
IRISA/INRIA
Motivations
 Geometric history length predictors introduced in 20042006
 OGEHL, CBP-1, dec. 2004
 TAGE, JILP ’06, feb. 2006
• Storage effective
• Exploits very long global histories
• Were defined with possible implementation in mind
 What are the limits of accuracy that can be captured
with these schemes ?
 How do they compare with unconstrained prediction
schemes ?
2
André Seznec
Caps Team
Irisa
Geometric history length predictors:
global history +multiple lengths
TO
T1
T2
L(0)
L(1)
L(2)
T3
?
T4
L(3)
L(4)
3
André Seznec
Caps Team
Irisa
GEometric History Length
predictor
The set of history lengths forms a geometric series
L(0) 0
L(i)  α i  1L(1)
Capture correlation
on very long histories
{0, 2, 4, 8, 16, 32, 64, 128}
most of the storage
What
is important:
for short
history !!L(i)-L(i-1) is drastically increasing
4
André Seznec
Caps Team
Irisa
Combining multiple predictions
 Neural inspired predictors
 Use a (multiply)-add tree
O-GEHL, CBP-1
 Partial matching
 Use tagged tables and the longest matching history
TAGE, JILP’ 06
5
André Seznec
Caps Team
Irisa
CBP-1 (2004): O-GEHL
Final computation through a sum
TO
T1
T2
L(0)
L(1)
L(2)
T3
∑
T4
L(3)
Prediction=Sign
L(4)
256Kbits: 12 components 3.670 misp/KI
6
André Seznec
Caps Team
Irisa
JILP ‘06: TAGE
longest matching history
=?
1
=?
1
1
=?
1
1
1
1
1
1
256Kbits: 3.358 misp/KI
7
André Seznec
Caps Team
Irisa
What is global history
 conditional branch history:
 path confusion on short histories 
 path history:
 Direct hashing leads to path confusion 
1. Represent all branches in branch history
2. Use path AND direction history
8
André Seznec
Caps Team
Irisa
Using a kernel history and a user
history
 Traces mix user and kernel activities:
 Kernel activity after exception
• Global history pollution
 Solution: use two separate global histories
 User
history is updated only in user mode
 Kernel history is updated in both modes
9
André Seznec
Caps Team
Irisa
Accuracy limits for TAGE
 Varying the predictor size, the number of
components, the tag width, the history length.
 Allowing multiple allocations
The best accuracy on distributed traces:
3.054 misp/KI
• History length around 1,000
• 15-20 components
• No need for tags wider than 16 bits
10
André Seznec
Caps Team
Irisa
Accuracy limits for GEHL
 Varying the predictor size, the number of components,
the history length, counter width
 (slightly) improving the update policy
 and fitting in the two hours simulation rule 
on the distributed traces:
2.842 misp/KI
• 97 components
• 8 bits counter
• 2,000 bits global history
11
André Seznec
Caps Team
Irisa
GEHL vs TAGE
 Realistic implementation parameters (storage
budget, number of components)
 TAGE is more accurate than (O-)GEHL
 Unlimited budget, huge number of components
 GEHL is more accurate than TAGE
12
André Seznec
Caps Team
Irisa
Will it be sufficient
to win
The Championship ?
GEHL history length: 2,000
97 components
2.842 misp/KI
13
André Seznec
Caps Team
Irisa
A step further: hybrid GEHL-TAGE
 On a few benchmarks, TAGE is more accurate
than GEHL,
 Let
us try an hybrid GEHL-TAGE predictor
14
André Seznec
Caps Team
Irisa
Hybrid GEHL-TAGE
Branch/path history + PC
GEHL
mux
TAGE
Meta
=
egskew
Inherit from:
Agree/bimode, YAGS, 2bcgskew,
15
André Seznec
Caps Team
Irisa
GEHL+TAGE
 GEHL provides the main prediction:
 also used as the base predictor for TAGE
(YAGS inspired)
 TAGE records when GEHL fails:
{prediction, address, history}
(agree/bimode, YAGS inspired)
 Meta selects between GEHL and TAGE
(2bcgskew inspired)
16
André Seznec
Caps Team
Irisa
Let us have fun !!
GEHL history length: 400
TAGE history length: 100,000
2.774 misp/KI
17
André Seznec
Caps Team
Irisa
Might still be
unsufficient 
GEHL history length: 400
TAGE history length: 100,000
2.774 misp/KI
18
André Seznec
Caps Team
Irisa
Adding a loop predictor
 The loop predictor captures the number of
iterations of a loop
 When successively encounters 8 times the
same number of iterations, the loop
predictor provides the prediction.
 Advantage:
 Very reliable
19
André Seznec
Caps Team
Irisa
GTL predictor
Branch/path history + PC
GEHL
=
egskew
Loop
predictor
+ static prediction on first occurrence
20
confidence
Meta
mux
mux
TAGE
André Seznec
Caps Team
Irisa
Hope this will be sufficient
to win
the Championship !!
GTL
GEHL, 97 comp., 400 hist.
+ TAGE, 19 comp., 100,000 hist
+ loop predictor
2.717 misp/KI
21
André Seznec
Caps Team
Irisa
Geometric History Length predictors
and limits on branch prediction
 Unlimited budget, huge number of components
 GEHL is more accurate than TAGE
 Very old correlation can be captured:
 On two benchmarks, using 10,000 history is really
helping
 Does not seem to be a lot of potential extra benefit
from local history


We did not find any interesting extra scheme apart loop
prediction
Loop prediction, very marginal apart gzip
22
André Seznec
Caps Team
Irisa
The End 
23
André Seznec
Caps Team
Irisa