Lecture 7 Behavioral Synthesis for LowPower

1
Lecture 7Behavioral Synthesis forLow-Power

• Behavioral Level Transforms
• Potential for large power reduction
• Summary

• Michael L. Bushnell
• CAIP Center and WINLAB
• ECE Dept., Rutgers U., Piscataway, NJ

2
Motivation

• Conventional automated layout synthesis method
• Describe design at RTL or higher level
• Generate technology-independent realization
• Map logic-level circuit to technology library
• Optimization goal shifting from low-area to
  low-power and higher performance
• Need accurate signal probability/activity
  estimates
• Consider low-power needs at all design levels

3
Behavioral-Level Transformations
4
Algorithm-Level Power Reductions vs. Other Levels
5
Differential Coefficients for Finite Impulse
Response (FIR) Filters

• Discrete-time Linear Time-Invariant FIR system
• Ci are the filter coefficients
• N is taps or filter length
• Differential Coefficients Method (DCM) reduces
  computations to save power
• Uses differences between coefficients rather than
  direct-form computation
• Uses various orders of differences
• Requires more storage devices and storage
  accesses

6
First-Order Differences

• 3 Consecutive outputs Y
• Rewrite product terms
• Except for C0, can express each coefficient as
  the sum of preceding coefficient and difference
  between it and the preceding coefficient

7
First-Order Differences (contd)

• Store product terms and reuse them for next
  output time period
• Need only 1 extra addition per product term and 1
  storage element
• Store C0 and ds
• Trade off long multiplier for a short one and
  storage overheads
• d1k-1/k is first-order difference between Ck and
  Ck-1

8
Orders of Differences
9
Second-Order Differences

• Coefficient
  expressions
• Needs just 2 extra storage variables and 2 extra
  additions per product to compute FIR output with
  2nd-order differences compared with direct form
  computation

10
Generalized mth-Order and Negative Differences

• mth-order differences require storage of m
  intermediate results for each product term, of
  size N, so need mN storage variables and m
  additions per product term compared with direct
  form
• Differences can be positive or negative
• Possible to get absolute value of partial product
  with negative differences

11
Sorted Recursive Differences (SRD)

• DCM only applicable to systems where envelope
  generated by coefficient sequences (or
  differences) is a smoothly-varying continuous
  function
• Mainly for low-pass FIR filters
• Recursively sort coefficients and use various
  orders of differences to reduce computation
• Use transposed direct form of FIR output
  computation
• No restriction on applicable coefficient sequence
• Word length reduction not the same for each
  coefficient

12
Transposed Direct-Form (TDF) Computation

• Compute all N product terms for particular data
  before computing terms for next sequential data
• Same throughput as for direct-form computation
• Signal-flow graph

13
Signal Flow Graph for TDM Realization
14
Maximum Savings in Adds Using SRD
15
Frequency Response of SRD Low-Pass Filter and
Hamming Window
16
Savings in Adds for Low-Pass Filter

• Black N 201, Grey N 101, White N 51

17
Savings in Shifts Using SRD
18
Least-Squares Coefficient Optimization for Filters

• Find coefficient closest to desired coefficient,
  but with fewest of 1 bits -- 1s called a
• Goal is to reduce the of additions
• Use sign-magnitude coefficient representation
• k maximum code class allowed per coefficient
• Use branch-and-bound method to solve integer
  programming problem of selecting coefficient
  approximations
• Shown to reduce addition computations in
  low-power filters by more than 40

19
Activity-Driven Architectural Transformations

• Basic idea Power consumption in digital filters
  depends on order of addition operators
• Restructure addition tree to move adders with
  higher-coefficient multiplications towards the
  output
• Higher-activity circuitry is moved closest to
  root of addition tree
• Definition of average signal activity over N
  consecutive time frames

20
Data Flow Graph of IIR Filter
21
Perfectly Balanced Addition Tree
22
Filter Implementations

• Can be bit-serial or word-parallel arithmetic
• W bits fed in parallel to adders and multipliers
• At time t 1, z of W bits change from time t
  values
• Activity b (t) z / W
• b (t) is a random variable stochastic process
  strict sense stationary
• Average power dissipation proportional to
• In bit-serial implementations, intra-word bit
  differences, and not inter-word bit differences,
  cause node activity

23
Architectural Transforms on DSP Filters

• I inputs to word-parallel computation tree
• I 2 l 1, l levels in tree
• y S aj bj
• Obtain minimum average value of qi over all
  balanced adder nodes when
• a1 a2 aI or a1 a2 aI
• Minimum average value in a linear array of adders
  when
• a1 a2 aI

j I j 1

24
Linear Array of Adders

• Assume mutually independent
  inputs, but method works even
  when signals correlate due to
  reconvergent fanouts

25
Power Optimization Algorithm

• Simulate circuit at functional level
• Using random, mutually-independent input values
• Note signal activities at all adder inputs
• Restructure adder trees using above 2 hypotheses
• Move additions with high activity closer to root
  of computation tree
• Recompute average activities
• Iterate until no additional power is saved
• Method shown to save up to 23 of power

26
Architecture-Driven Voltage Scaling

• Scale down VDD to save power, but increases
  circuit delay
• Reduce delay by scaling down device sizes (less
  C)
• But interconnect C becomes dominant, not device C
• Need architectural transformation to introduce
  more parallelism to compensate for increased
  delay
• Introduce parallel or pipelined architecture

27
Example Original Data Path Operator
28
Redesigned Parallel Implementation gt 2 X Area
Increase
29
Redesigned Pipelined Implementation
30
Operation Reduction Methods

• Reduce operators in data flow graph
• Computes X3 AX2 BX C
• Reduces C, but may slow down critical paths
• Reduction maintaining throughput

31
Example

• Reduction with less throughput

32
Operation Substitution Methods

• Multiplication uses more energy than addition
• Replace multipliers with adders in high-level
  synthesis

33
Method and Results

• Transformations
• Common sub-expression utilization
• Apply distributive law
• Replace multiplication with repeated shifting and
  adding
• On an 11-tap FIR filter
• Saved 62 of dissipated power

34
Precomputation-Based Optimization

• Basic idea Precompute (with low-overhead
  hardware) circuit output logic values 1 cycle
  before they are needed
• Use precomputed information in next clock cycle
  to disable unneeded hardware, reduces switching
  activity
• Must be careful Precomputation hardware can add
  to area and lengthen clock period

35
Precomputation Architecture
36
Explanation

• When both functions are 0, indicates that no
  prediction of output value is possible
• When prediction happens (we know definitely that
  output is 1 or 0), we turn off R2
• Reduces activity in combinational logic block
• R1 still is active, so Comb still computes
  correct output
• More effective if P (f1 f2) is large
• For comparator, saves 50 of the power

37
Specific Comparator Example
38
Can Precompute Outputs Needed 2 or More Clocks
Later

• Can reduce switching activity by 12.5

39
Example Adder-Comparator
40
Precomputation with Shannons Expansion Theorem
41
Summary

• Behavioral or architectural level synthesis
• Resynthesize state variable equations to save
  power
• Scale down supply voltage, and introduce
  parallelism and pipelining to make up for
  slow-down of hardware

42
Future Research Directions

• Formal methods for behavioral or data flow level
  to explore power reduction design space
• Behavioral level power estimation algorithms
  needed
• Synthesis scheduling and data path allocation
  algorithms should incorporate power tradeoffs