Looking for limits in branch prediction with the GTL predictor

1
Looking for limits in branch predictionwith the
GTL predictor

• André Seznec
• IRISA/INRIA/HIPEAC

2
Motivations

• Geometric history length predictors introduced in
  2004-2006
• 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 ?

3
Geometric history length predictors global
history multiple lengths
TO
T1
T2
?
L(0)
T3
L(1)
L(2)
T4
L(3)
L(4)
4
GEometric History Length predictor
The set of history lengths forms a geometric
series
Capture correlation on very long histories
0, 2, 4, 8, 16, 32, 64, 128
most of the storage for short history !!
What is important L(i)-L(i-1) is drastically
increasing
5
Combining multiple predictions

• Neural inspired predictors
• Use a (multiply)-add tree
• Partial matching
• Use tagged tables and the longest matching
  history

O-GEHL, CBP-1
TAGE, JILP 06
6
CBP-1 (2004) O-GEHL Final computation through
a sum
L(0)
PredictionSign
256Kbits 12 components 3.670 misp/KI
7
JILP 06 TAGElongest matching history
?
?
1
1
1
1
1
1
1
1
1
256Kbits 3.358 misp/KI
8
What is global history

• conditional branch history
• path confusion on short histories ?
• path history
• Direct hashing leads to path confusion ?
• Represent all branches in branch history
• Use path AND direction history

9
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

10
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

11
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

12
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

13
Will it be sufficient to win The Championship ?
GEHL history length 2,000 97 components
2.842 misp/KI
14
A step further hybrid GEHL-TAGE

• On a few benchmarks, TAGE is more accurate than
  GEHL,
• Let us try an hybrid GEHL-TAGE predictor

15
Hybrid GEHL-TAGE
Branch/path history PC
Inherit from Agree/bimode, YAGS, 2bcgskew,
16
GEHLTAGE

• 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)

17
Let us have fun !!
GEHL history length 400 TAGE history length
100,000
2.774 misp/KI
18
Might still be unsufficient ?
GEHL history length 400 TAGE history length
100,000
2.774 misp/KI
19
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

20
GTL predictor
confidence
static prediction on first occurrence
21
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
22
Geometric History Length predictorsand 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

23
The End ?