Ch02 – Iteration bound

(Giới hạn lặp)

TLTK:
1. Các slide từ sách của Prof. Parhi
2. Slide của Prof. Viktor Öwall
3. Slide bài giảng của Thầy Hồ Trung Mỹ
4. KTGHK cũ

1
 Outline
 Introduction
 Loop Bound (giới hạn vòng)
 Important Definitions and Examples
 Iteration Bound (giới hạn lặp)
 Important Definitions and Examples
 Techniques to Compute Iteration Bound
 Algorithms to compute iteration bound
 Longest Path Matrix (LPM)
 Minimum Cycle Mean (MCM)

2
 Some Definitions For DSP Algorithms

 Non-terminating programs
 Execute repetitively
 Iteration
 Execution of all the computations in the algorithm once
 Iteration period
 The time required for execution of an iteration
 Iteration rate
 Number iterations executed per second

3
 Some Definitions For DSP Algorithms
 Sampling period
 The time difference between two consecutive samples
 Sampling rate (throughput)
 Number of samples processed per second
 Critical path for combinational logic circuit
 Longest path between inputs and outputs
 Critical path for sequential circuits
 Longest path between any two storage (delay)
elements
 In general, critical path is defined to be the path with
the longest computation time among all paths that
contain zero delays.
4
 Some Definitions For DSP Algorithms
 Critical path computation time determines
 Minimum feasible clock period of DSP systems
 Clock period = Clock cycle time = 1/clock rate =
1/Critical path period
 Latency
 Difference between the time an output generated and
the time at which its corresponding input was received
by the system
 Latency representations
• Absolute time units, or the number of gate delays, for
combinational logic circuit systems.
• Number of clock cycle for the sequential systems.

5
 Various Representations of DSP Systems

 DSP Algorithm Descriptions

 Behavioral Descriptions
 Graphical Representations

6
 DSP algorithm descriptions
 Mathematical formulations
 Specify functionality of DSP algorithm
 Does not specify the order & structure of the
internal operations

 Behavioral description languages or graphical

representations for architectural design

7
 Behavioral Descriptions
 Applicative languages
 A set of equations
 Silage language
 Prescriptive languages
 Specify the order of the assignment statements
 Pascal, C, SystemC,..
 Descriptive languages
 Represents the structure of DSP system
 Verilog HDL or VHDL

8
 Graphical Representations
 Efficient for investigating and analyzing data
flow properties
 Efficient for exploiting the inherent parallelism
among different subtasks
 Easy mapping DSP algorithm descriptions to
hardware structural implementations
 Provide technology-independent architecture

9
 Various graphical representations

 Block Diagram (BD)

 Signal Flow Graph (SFG)

 Data Flow Graph (DFG)

 Dependence Graph (DG)

10
Graphical representations

functional
blocks

11
 Ex: Direct Form 4-tap FIR filter

Note:
Clock speed limited by Critical Path!
TCritical = TM + (N-1)TA
N = Nr. of Taps 12
Graphical Representation Method 2:
Signal-Flow Graph (SFG)

Note:
 Source = no entering edge
 Sink = only entering edges 13
14
15
Block diagram with data broadcast

16
Graphical Representation Method 3:
Data-Flow Graph (DFG)

17
18
 Loop Bound
 Loop: a directed path that begins and ends at the
same node.
 Loop Bound of the loop
 Lower bound on the loop computation time or the time
requires to execute one iteration of the loop.
 Defined as t𝑙/w𝑙, where t𝑙 is the loop computation time
and w𝑙 is the number of delays in the loop
 E.g. 𝑦(𝑛)
(2)  A→B→A is a loop
𝑥(𝑛) (4)
A B a  t𝑙 = 2+ 4=6, w𝑙 = 1
D ⟹loop bound = t𝑙/w𝑙 =6

𝑦 𝑛 = ay 𝑛 − 1 + 𝑥 𝑛
 Loop Bound (cont’d)
 Another example
 A→B→A is a loop
y(n)  t𝑙 = 2+ 4=6, w𝑙 = 2 (since 2D)
x(n) (2) ⟹ loop bound = t𝑙/w𝑙 = 3
A (4)
B a
It means one iteration of loop can be
executed in 3 time unit.
2D

 Another example
D
(2) (4) (5) Two loops
 A→B→A : T = 6, W = 2, loop bound = 3
A B C
 A→B→C→A, T = 11, W = 1, loop bound = 11
2D
 Loop Bound (cont’d)

Same critical path

Different loop bound
Critical path:
 For combinational logic circuit: Longest path between
inputs and outputs
 For sequential circuit: Longest path between any two
storage (delay) elements.
21
Loop Bound (cont’d)

22
 Iteration Bound



 In other words, it is not possible to achieve iteration period (or

sample period) lower than iteration bound.
𝑇sample ≥ 𝑇∞
23
 Example: Iteration Bound
 Compute the loop bounds and iteration bound of the following
loops.
D  Two loops
(2) (4) (5)  A→B→A : T = 6, W = 2, loop bound = 3
A B C
 A→B→C→A, T = 11, W = 1, loop bound = 11
⇒ 𝑻∞ = 𝐦𝐚𝐱{3, 11} = 𝟏𝟏𝐮. 𝐭.
2D

 Compute the loop bounds and iteration bound of the following

loops.
 Three loops

= 12ns
24
Iteration Bound (cont’d)

25
 Critical Path
 Critical Path: the path with longest computation time
among all paths that contain zero delays.
 Critical Path: the lower bound on clock period. 𝑇clock ≥ 𝑇𝑐𝑟𝑖𝑡𝑖𝑐𝑎𝑙

 Example: Critical Paths

6→3→2→1: 5 u.t.
5→3→2→1: 5 u.t.
 Minimum clock period = 5 u.t.

 To achieve high-speed, the

length of the critical path can be
reduced by pipelining and
parallel processing (Chapter 3).
26
Algorithms to compute iteration bound

 Longest Path Matrix (LPM)

 Minimum Cycle Mean (MCM)

27
Algorithms to compute iteration bound

 Longest Path Matrix (LPM)

 Minimum Cycle Mean (MCM)

28
 Longest Path Matrix (LPM)

𝒎 𝟏 𝒎−𝟏
𝒍𝒊,𝒋 = max −𝟏, 𝒍𝒊,𝒌 + 𝒍𝒌,𝒋 𝐟𝐨𝐫 𝒌 ∈ 𝟏, 𝟐, ⋯ , 𝒅 ; 𝐦 = 𝟐, 𝟑, ⋯ , 𝒅
𝒌

𝑚
and 𝑙𝑖,𝑖 ≠ −1 29
Examples for LPM

-1 0 -1 -1
(1)
L = 4 -1 0 -1
5 -1 -1 0
5 -1 -1 -1
d i → dj
max
q q
t if at least one path exists
𝑙 (1)i,j =
-1 if no such path exists

where max tqd i → dj is the maximum of the longest

computation time between delay element di to delay
element dj

30
 𝒎 𝟏 𝒎−𝟏
𝒍𝒊,𝒋 = max −𝟏, 𝒍𝒊,𝒌 + 𝒍𝒌,𝒋 𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝒌∈ 𝟏,𝒅
𝑙 (1)1,2
• Ex. i = 1; m = 2; L(2)
(2) (1) (1) i=1
𝑙 1,1 = max( -1, 𝑙 1,k + 𝑙 k,1)
k{2}

= max( -1,0+4) = 4
(2) (1) (1)
𝑙 1,2 = max( -1, 𝑙 1,k + 𝑙 k,2)

= max( -1) = -1
j=1
𝑙 (1)2,1
𝑙 1,3(2) = max( -1, 𝑙 (1)1,k + 𝑙 (1)k,3) i=1
k{2} 4 - - -
= max(-1, 0+0) = 0 - - - -
(2)
𝑙 1,4(2) = max ( -1, 𝑙 (1)1,k + 𝑙 (1)k,4) L = - - - -
- - - -
= max(-1) = -1
31
 𝒎 𝟏 𝒎−𝟏
𝒍𝒊,𝒋 = max −𝟏, 𝒍𝒊,𝒌 + 𝒍𝒌,𝒋 𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝒌∈ 𝟏,𝒅
𝑙 (1)1,2
• Ex. i = 1; m = 2; L(2)
(2) (1) (1) i=1
𝑙 1,1 = max( -1, 𝑙 1,k + 𝑙 k,1)
k{2}

= max( -1,0+4) = 4
(2) (1) (1)
𝑙 1,2 = max( -1, 𝑙 1,k + 𝑙 k,2)

= max( -1) = -1
j=2
𝑙 1,3(2) = max( -1, 𝑙 (1)1,k + 𝑙 (1)k,3) i=1
k{2} 4 -1 - -
= max(-1, 0+0) = 0 - - - -
(2)
𝑙 1,4(2) = max ( -1, 𝑙 (1)1,k + 𝑙 (1)k,4) L = - - - -
- - - -
= max(-1) = -1
32
 𝒎 𝟏 𝒎−𝟏
𝒍𝒊,𝒋 = max −𝟏, 𝒍𝒊,𝒌 + 𝒍𝒌,𝒋 𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝒌∈ 𝟏,𝒅
𝑙 (1)1,2
• Ex. i = 1; m = 2; L(2)
(2) (1) (1) i=1
𝑙 1,1 = max( -1, 𝑙 1,k + 𝑙 k,1)
k{2}

= max( -1,0+4) = 4
(2) (1) (1)
𝑙 1,2 = max( -1, 𝑙 1,k + 𝑙 k,2)
𝑙 (1)2,3
= max( -1) = -1
j=3
𝑙 1,3(2) = max( -1, 𝑙 (1)1,k + 𝑙 (1)k,3) i=1
k{2} 4 -1 0 -
= max(-1, 0+0) = 0 - - - -
(2)
𝑙 1,4(2) = max ( -1, 𝑙 (1)1,k + 𝑙 (1)k,4) L = - - - -
- - - -
= max(-1) = -1
33
 𝒎 𝟏 𝒎−𝟏
𝒍𝒊,𝒋 = max −𝟏, 𝒍𝒊,𝒌 + 𝒍𝒌,𝒋 𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝒌∈ 𝟏,𝒅
𝑙 (1)1,2
• Ex. i = 1; m = 2; L(2)
(2) (1) (1) i=1
𝑙 1,1 = max( -1, 𝑙 1,k + 𝑙 k,1)
k{2}

= max( -1,0+4) = 4
(2) (1) (1)
𝑙 1,2 = max( -1, 𝑙 1,k + 𝑙 k,2)

= max( -1) = -1
j=4
𝑙 1,3(2) = max( -1, 𝑙 (1)1,k + 𝑙 (1)k,3) i=1
k{2} 4 -1 0 -1
= max(-1, 0+0) = 0 - - - -
(2)
𝑙 1,4(2) = max ( -1, 𝑙 (1)1,k + 𝑙 (1)k,4) L = - - - -
- - - -
= max(-1) = -1
34
 𝒎 𝒎 𝟏 + 𝒍 𝒎−𝟏 𝟏 𝒎−𝟏
𝒍𝒊,𝒋
𝒍𝒊,𝒋 ==𝒌∈
max
max −𝟏,𝒍𝒍𝒊,𝒌
−𝟏, 𝒊,𝒌 + 𝒍 𝒌,𝒋
𝒌,𝒋
𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝟏,𝒅𝒌
𝑙 (1)2,3
• Ex. i = 2; m = 2; L(2)
(2) (1) (1)
𝑙 2,1 = max( -1, 𝑙 2,k +𝑙 k,1)
k{3} i=2
= max( -1,0+5) = 5
(2) (1) (1)
𝑙 2,2 = max( -1, 𝑙 2,k + 𝑙 k,2)
k{1}

= max( -1,4+0 ) = 4 𝑙 (1)3,1

𝑙 2,3(2) = max( -1, 𝑙 (1)2,k + 𝑙 (1)k,3) i=1
4 -1 0 -1
i=2
= max(-1) = -1 5 4 -1 0
(2)
𝑙 2,4(2) = max ( -1, 𝑙 (1)2,k + 𝑙 (1)k,4) L = - - - -
k{3}
- - - -
= max(-1,0+0) = 0
35
 𝒎𝒎 𝟏 + 𝒍 𝒎−𝟏 𝟏 𝒎−𝟏
𝒍𝒊,𝒋
𝒍𝒊,𝒋 ==𝒌∈
max
max −𝟏,𝒍𝒍𝒊,𝒌
−𝟏, 𝒊,𝒌 + 𝒍 𝒌,𝒋
𝒌,𝒋
𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝟏,𝒅 𝒌

• Ex. i = 3; m = 2; L(2)
𝑙
(2) (1)
𝑙 3,k
(1)
+ 𝑙 k,1) 𝑙 (1)3,4
3,1 = max( -1,
k{4}

= max( -1,0+5) = 5 i=3

(2) (1) (1)
𝑙 3,2 = max( -1, 𝑙 3,k + 𝑙 k,2)
k{1}

= max( -1,5+0) = 5 𝑙 (1)4,1

𝑙 3,3(2) = max( -1, 𝑙 (1)3,k + 𝑙 (1)k,3) i=1
4 -1 0 -1
i=2
= max(-1) = -1 5 4 -1 0
(2) i=3
𝑙 3,4(2) = max ( -1, 𝑙 (1)3,k + 𝑙 (1)k,4) L = 5 5 -1 -1
- - - -
= max(-1) = -1
36
 𝒎 𝒎 𝟏 + 𝒍 𝒎−𝟏 𝟏 𝒎−𝟏
𝒍𝒊,𝒋
𝒍𝒊,𝒋 ==𝒌∈
max
max −𝟏,𝒍𝒍𝒊,𝒌
−𝟏, 𝒊,𝒌 + 𝒍 𝒌,𝒋
𝒌,𝒋
𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝟏,𝒅𝒌

• Ex. i = 4; m = 2; L(2)
(2) (1) (1)
𝑙 4,1 = max( -1, 𝑙 4,k +𝑙 k,1)

= max( -1) = -1
(2) (1) (1)
𝑙 4,2 = max( -1, 𝑙 4,k + 𝑙 k,2)
k{1} i=4

= max( -1,5+0) = 5 𝑙 (1)4,1

𝑙 4,3(2) = max( -1, 𝑙 (1)4,k + 𝑙 (1)k,3) i=1
4 -1 0 -1
i=2
= max(-1) = -1 5 4 -1 0
(2) i=3
𝑙 4,4(2) = max ( -1, 𝑙 (1)4,k + 𝑙 (1)k,4) L = 5 5 -1 -1
i=4
-1 5 -1 -1
= max(-1) = -1
37
 (m=1)

(m=2)

(m=3) (m=4)

38
 Examples for LPM (other solution)
𝒎 𝟏 𝒎−𝟏
𝒍𝒊,𝒋 = max −𝟏, 𝒍𝒊,𝒌 + 𝒍𝒌,𝒋
𝒌∈ 𝟏,𝒅

𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝟏
𝑳 𝑳 𝟏

+
𝑳 𝟐 = +

𝑳 𝟐 =

39
 Examples for LPM (other solution)
𝒎 𝟏 𝒎−𝟏
𝒍𝒊,𝒋 = max −𝟏, 𝒍𝒊,𝒌 + 𝒍𝒌,𝒋
𝒌∈ 𝟏,𝒅

𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝟏
𝑳 𝑳 𝟏

+
𝑳 𝟐 = +

𝑳 𝟐 =

40
 Examples for LPM (other solution)
𝒎 𝟏 𝒎−𝟏
𝒍𝒊,𝒋 = max −𝟏, 𝒍𝒊,𝒌 + 𝒍𝒌,𝒋
𝒌∈ 𝟏,𝒅

𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝟏
𝑳 𝑳 𝟏

+
𝑳 𝟐 = +

𝑳 𝟐 =

41
 Examples for LPM (other solution)
𝒎 𝟏 𝒎−𝟏
𝒍𝒊,𝒋 = max −𝟏, 𝒍𝒊,𝒌 + 𝒍𝒌,𝒋
𝒌∈ 𝟏,𝒅

𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝟏
𝑳 𝑳 𝟏

+
𝑳 𝟐 = +

𝑳 𝟐 =

42
 Examples for LPM (other solution)
𝒎 𝟏 𝒎−𝟏
𝒍𝒊,𝒋 = max −𝟏, 𝒍𝒊,𝒌 + 𝒍𝒌,𝒋
𝒌∈ 𝟏,𝒅

𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝟏
𝑳 𝑳 𝟏

+
𝑳 𝟐 = +

𝑳 𝟐 =

43
 Examples for LPM (other solution)
𝒎 𝟏 𝒎−𝟏
𝒍𝒊,𝒋 = max −𝟏, 𝒍𝒊,𝒌 + 𝒍𝒌,𝒋
𝒌∈ 𝟏,𝒅

𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝟏
𝑳 𝑳 𝟐

+
𝑳 𝟑 = +

𝑳 𝟑 =

44
 Examples for LPM (other solution)
𝒎 𝟏 𝒎−𝟏
𝒍𝒊,𝒋 = max −𝟏, 𝒍𝒊,𝒌 + 𝒍𝒌,𝒋
𝒌∈ 𝟏,𝒅

𝑙 (1)i,k ≠ −1 ; 𝑙 (m-1)k,j ≠ −1
𝟏
𝑳 𝑳 𝟑

+
𝑳 𝟒 = +

𝑳 𝟒 =

45
Examples for LPM

d1  d1: 4-5-6: 4 u.t.

d 1  d 2:
• 7: 1 u.t.
• 4-5-7: 4 u.t.
d 2  d 1:
 1-2-3-4-5-6: 8 u.t.
 1-2-3-6: 5 u.t.
d2  d2: 1-2-3-4-5-7: 8 u.t.

= max −𝟏, 𝟒 + 𝟒, 𝟒 + 𝟖
𝟏 𝟏
𝑳 𝑳

= 8 u.t.
46
TD: Đề KTGHK-HK112

47
48
49
Minimum Cycle Mean Method (MCM)

50
51
Minimum Cycle Mean Method (MCM)-Summary

52
 Examples for MCM (1/4)

Graph Gd
DFG 53
 Examples for MCM (2/4)

𝑤 2,1 𝑤 ′ 2,1

𝒘′ 𝒊, 𝒋 = −𝒘 𝒊, 𝒋
Graph Gd Graph 𝑮′𝒅
DFG 54
 Examples for MCM (3/4)

reference
node
𝒇𝟒 𝟏

𝒘′ 𝒊, 𝒋 is the weight of the edge 𝒊 → 𝒋 in 𝑮′𝒅

55
 Examples for MCM (4/4)

treated
as zero
Note that:
56
Examples for MCM (cont’d)

57
 Conclusion
When the DFG is recursive, the iteration bound
is the fundamental limit on the minimum sample
period of a hardware implementation of the DSP
program.

Two algorithms to compute iteration bound, LPM

and MCM, were explored.

58