VALUE PREDICTION WITH PERCEPTRONS

1
VALUE PREDICTION WITH PERCEPTRONS

• Anita Thomas

2
Objectives

• Speculative Execution of Dependent Instructions
• Prediction of instruction output values using
  perceptrons
• Forwarding of predicted values to dependent
  instructions
• Low Misprediction Rates High IPC

3
Instruction Dependences

• Control flow dependences overcome by predicting
  direction of branches
• Data flow dependences
• -WAW, WAR Removed by static dynamic renaming
  
• E.g. MOV R1,R2 MOV R3,R2
• SUB R1,R2,8 SUB R2,R1,8
• -RAW
• E.g. MOV R1,R2
• SUB R3,R1,8
• Value prediction overcomes RAW dependences by
  predicting the result of an instruction

4
Value Prediction

• Prediction
• Predict the result of an instruction using local
  history global correlations
• Dependent instructions execute with speculative
  input operands
• Verification
• After execution, parent instruction compares the
  result with its predicted value
• In case of a misprediction, the dependent
  instructions re-execute with correct input
  operands.
• Misprediction penalty is 1 cycle more than
  non-speculative execution is required for
  verification of the predicted value

5
Perceptrons
...
X 1
X i

1
X n
wi
w1
w0
wn
Y

• Learns a target boolean function t(x1.xn) of n
  inputs
• The output y of a perceptron
• y w0 Sxiwi
• Inputs x1-xn are bipolar
• w0 is the bias weight of the perceptron. It
  learns the bias of the target output

6
Perceptions (Cont.)

• Weights w1wn represent degree of correlation
  between the inputs and the output
• (Output lt threshold OR Output gt -threshold )
  ?further training required. Prediction can be
  used only for update of weights cannot be used
  in actual execution
• Output gt threshold ? prediction can be used
  weights updated depending on their correlation
  with the actual value
• Can use longer history length since perceptron
  size scales linearly with number of inputs

7
Last-2 Value Predictor

• Predictor has 2 levels Value History Table (VHT)
  Table of Perceptrons
• VHT 4-way set associative
• Last 2 values, value0 value1 maintained by LRU
• Value History Pattern (vhp1,vhp2,vhp3,..,vhp n)
• Pattern of 0 1 for last n instances of the
  instruction
• Vhp bit 0 ? value0 Vhp bit 1 ? value1
• n is the length of history to be considered
• Table Of Perceptrons
• Weights0..n - w0 is the bias weight
• - w1..n represent correlation weights
  where n is the length of history considered

8
Last-2 Value Prediction
Table Of Perceptrons

• Value History Table

Data Values
Tag
LRU
VHP
Decoder
Decoder
Instruction Address
Hash function
-
Predicted Date Value
w
21 MUX
Prediction Valid
9
Predictor Lookup

• Index for VHT PC Vht-size
• vht-size number of sets in the VHT
• Entries in the set are scanned till tag PC
• Index for the Perceptron Table (PC xor vhp)
  Pht-size
• Perceptron Lookup
• Output y w0
• For all history bits, i 1n
• if vhp i 1
• y y w i
• else
• y y w i
• y gt 0 ? Prediction value0 y lt 0 ?
  Prediction value1
• Prediction is forwarded to dependent
  instructions only if
• y gt Threshold

10
Predictor Update

• If Resolved Output is one among existing 2 values
  in VHT
• Bias weight w0 captures only Instruction Bias
• Resolved Output value0 ? w0 decremented
• Resolved Output value1 ? w0 incremented
• Remaining Weights w1-wn capture Prediction
  correctness For all i 1- n
• Vhp i 0 Resolved value value0 or
• Vhp i 1 Resolved value value1
• w i is positively correlated ? w i
  incremented
• Vhp i 0 Resolved value value1 or
• Vhp i 1 Resolved value value0
• w i is negatively correlated ? w i
  decremented
• LRU Vhp are updated in the VHT

11
Predictor Update (Cont.)

• If Resolved Output is a new value
• LRU value in VHT is replaced with this new value
• If w0 indicates bias to the replaced value, bias
  is no longer valid has to be reset
• (Lru 0 w0 lt 0) OR (Lru 1 w0 gt 0) ? w0
  0
• All remaining weights are uncorrelated
  decremented
• LRU Vhp are updated in the VHT

12
Stride-based Value Prediction
Any Stride/ Update value stride
Same Stride/ Update value
VHT miss/Update value

• Switches from Transient to Steady State if the
  same stride occurs twice consecutively
• Perceptrons can provide additional confidence
  estimation in this change from Transient to
  Steady state
• Results are predicted only in the steady state.
  Predicts results like memory addresses that vary
  by a constant stride

Init Dont Predict
Steady Predict
Same Stride/ Update value
Transient Dont Predict
Different Stride/ Update value stride
Different Stride/ Update value stride
13
Stride-based Value Prediction
Value History Table (VHT)
Tag State Value Stride
Decoder
Instruction Address
Hash function
Predicted Date Value


w
(i)
Prediction Valid
14
Thinking Ahead

• Use Stride Predictor for instructions that
  perform memory address computation
• Use Last-2 value predictor for predicting memory
  register values
• Scale Last-2 value predictor to Last-4 values by
  defining additional thresholds
• Implement perceptrons to provide better
  confidence estimation for Stride Prediction when
  switching from Transient to Steady state
• Considering global correlations along with local
  history for predictions

15

• Questions???
• Comments.
• Suggestions.

16

• THANK YOU !!