Resource sharing

This presentation can be used for non-commercial purposes as long as this note and the copyright footers are not removed
© Giovanni De Micheli – All rights reserved
 Module 1
 Objectives
 Motivation and problem formulation
 Flat and hierarchical graphs
 Functional and memory resources
 Extension to module selection

(c) Giovanni De Micheli 2

 Allocation and binding

 Allocation:
 Number of resources available
 Binding:
 Relation between operations and resources
 Sharing:
 Many-to-one relation
 Optimum binding/sharing:
 Minimize the resource usage

(c) Giovanni De Micheli 3

 Binding

 Limiting cases:
 Dedicated resources
 One resource per operation
 No sharing

 One multi-task resource

 ALU

 One resource per type

(c) Giovanni De Micheli 4

 Optimum sharing problem
 Scheduled sequencing graphs
 Operation concurrency is well defined
 Consider operation types independently
 Problem decomposition
 Perform analysis for each resource type

(c) Giovanni De Micheli 5

 Compatibly and conflicts
 Operation compatibility: t1 x=a+b y=c+d 1 2

 Same type t2 s=x+y t=x-y 3 4

 Non concurrent t3 z=a+t 5

Compatibility graph
 Compatibility graph:
1 2
 Vertices: operations
 Edges: compatibility relation 5

3 4

 Conflict graph:
 Complement of compatibility graph Conflict graph
1 2

3 4

(c) Giovanni De Micheli 6

 Example
t1 x=a+b y=c+d 1 2

t2 s=x+y t=x-y 3 4

t3 z=a+t 5

Conflict Compatibility
1 2 1 2

5 5

3 4 3 4

Coloring Partitioning

ALU1: 1,3,5
ALU2: 2,4

(c) Giovanni De Micheli 7

 Compatibility and conflicts
 Compatibility graph:
 Partition the graph into a minimum number of cliques
 Find clique cover number k ( G+ )
 Conflict graph:
 Color the vertices by a minimum number of colors.
 Find the chromatic number х ( G_ )
 NP-complete problems:
 Heuristic algorithms

(c) Giovanni De Micheli 8

 Data-flow graphs
(flat sequencing graphs)

 The compatibility/conflict graphs have special properties:

 Compatibility
 Comparability graph
 Conflict
 Interval graph

 Polynomial time solutions:

 Golumbic’s algorithm
 Left-edge algorithm

(c) Giovanni De Micheli 9

 Perfect graphs
 Comparability graph:
 Graph G (V, E ) has an orientation G ( V, F ) with the transitive property
(vi, vj) є F and (vj, vk) є F → (vi, vk) є F
 Interval graph:
 Vertices correspond to intervals
 Edges correspond to interval intersection
 Subset of chordal graphs
 Every loop with more than three edges has a chord

(c) Giovanni De Micheli 10

 Example
NOP 0

1 2 10
TIME 1 * * +

3 6 11
TIME 2 * * <

4 7 8
TIME 3 - * *

5 9
TIME 4 - +

NOP
9

3 1 8
4 10

7 6 2
5 11

(c) Giovanni De Micheli 11

 Example

NOP 0

1 2 10
* * + 10
TIME 1 1 2
3 6 11
* 11
TIME 2 * < 3 6

4 7 8
- * *
TIME 3 4 7 8
5 9
TIME 4 - +
5 9
n

NOP

(c) Giovanni De Micheli 12

 Left-edge algorithm
 Input:
 Set of intervals with left and right edge
 Start and Stop times
 A set of colors (initially one color)
 Rationale:
 Sort intervals in a list by left edge
 Assign non overlapping intervals to first color using the list
 When possible intervals are exhausted,
increase color counter and repeat

(c) Giovanni De Micheli 13

 Example
0 1 2 3 4 5 6 7 8

1 6 1

4
7 4 3
7

3
2 5
5

Conflict graph
Intervals

1 6

0 1 2 3 4 5 6 7 8
7 4 3 1 2 3

6 7 5

2 5 4

Colored conflict Coloring

graph
(c) Giovanni De Micheli 14
 Left-edge algorithm

LEFT_EDGE(I) {
Sort elements of I in a list L in ascending order of li;
c = 0;
while (some interval has not been colored) do {
S = Ø;
r = 0;
while ( exists s є L such that ls > r ) do {
s = First element in the list L with ls > r;
S = S U {s};
r = rs;
Delete s from L;
}
c = c + 1;
Label elements of S with color c;
}
}

(c) Giovanni De Micheli 15

 Hierarchical sequencing graphs

 Hierarchical conflict/compatibility graphs:

 Easy to compute
 Prevent sharing across hierarchy
 Flatten hierarchy:
 Bigger graphs
 Destroy nice properties

(c) Giovanni De Micheli 17

 Example

a
+
TIME 1 a a
* a 2
TIME 2
2
TIME 3 2 *
*
3 3
TIME 4 * *
4 3
TIME 5 a 4 a 4
*
+ *
TIME 6 a
*
TIME 7

(a) (b) (c)

(c) Giovanni De Micheli 18

 Example

NOP NOP a
TIME 1 a a

TIME 2 BR c d c d c d
TIME 3 b b
TIME 4 b
NOP NOP

(a) (b) (c)

(c) Giovanni De Micheli 19

 Register binding problem
 Given a schedule:
 Lifetime intervals for variables
 Lifetime overlaps
 Conflict graph (interval graph):
 Vertices ↔ variables
 Edges ↔ overlaps
 Interval graph
 Compatibility graph (comparability graph):
 Complement of conflict graph

(c) Giovanni De Micheli 20

 Register sharing in data-flow graphs

 Given:
 Variable lifetime conflict graph

 Find:
 Minimum number of registers storing all the variables
 Key point:
 Interval graph
 Left-edge algorithm (polynomial-time complexity)

(c) Giovanni De Micheli 21

 Example

1 2 z1 z2
TIME 1 * * z1 z2
z1 z2
3 6
TIME 2 * * z3 z4
z3 z4
z4
z3 4 7
TIME 3 - *
z5 z6 z5 z6
z6
z5
TIME 4 5
-

(a) (b) (c)

(c) Giovanni De Micheli 22

 Register sharing
general case
 Iterative conflicts:
 Preserve values across iterations
 Circular-arc conflict graph
 Coloring is intractable
 Hierarchical graphs:
 General conflict graphs
 Coloring is intractable
 Heuristic algorithms

(c) Giovanni De Micheli 23

 Example

u
y
x
u 3 x u dx y u dx x dx x
1 2 10
TIME 1 * * +

z1 z2 3 a z1 z2 u y x
6
TIME 2 3 11
* * <
dx z3 z4
z3 z4 8
4
7 c
TIME 3 *
- *
y z7
z5 z6 z7
z5 z6
TIME 4 - 5 + 9

u y
y
u

(a) (b)

(c) Giovanni De Micheli 24

 Example
Variable-lifetimes and circular-arc conflict graph

z1
z2 u

x
z1 z2
1

u y 4 2 z4 z3
x z3 z4 y

z5 z6
z7
z6
z5
z7

(c) Giovanni De Micheli 25

 Bus sharing and binding

 Find the minimum number of busses to accommodate all data

transfer
 Find the maximum number of data transfers for a fixed number
of busses
 Similar to memory binding problem

(c) Giovanni De Micheli 30

 Example
1 2 z1 z2
TIME 1 * * z1 z2
z1 z2
3 6
TIME 2 * * z3 z4
z3 z4
z4
z3 4
7
TIME 3 - *
z5 z6 z5 z6
z6
z5
TIME 4 5
-

(a) (b) (c)

 One bus:
 3 variables can be transferred
 Two busses:
 All variables can be transferred
(c) Giovanni De Micheli 31
 Module selection problem
 Extension of resource sharing
 Library of resources:
 More than one resource per type

 Example:
 Ripple-carry adder (small and slow)
 Carry look-ahead adder (big and fast)

 Resource modeling:
 Resource subtypes with
( area, delay ) parameters

(c) Giovanni De Micheli 32

 Module selection solution
 ILP formulation:
 Decision variables
 Selectresource sub-type
 Determine ( area, delay )

 Heuristic algorithm
 Determine minimum latency with fastest resource subtypes
 Recover area by using slower resources on non-critical paths

(c) Giovanni De Micheli 33

 Example
NOP 0
(1,1) Slow multipliers
TIME 1 *
1 6
(1,2)
(2,1) save area
*

2 10
TIME 2 * +

3 8 11
* * <
TIME 3

4 7
TIME 4 - *

5 (2,2) 9
TIME 5 - +

n
NOP

 Multipliers with:
 ( Area, delay ) = ( 5,1 ) and ( 2,2 )
 Latency bound of 5

(c) Giovanni De Micheli 34

 Example 2
NOP 0

(1,1)
1 (1,2) (2,1) 10
* 2 +
TIME 1 *

3 6 11
TIME 2 * * <

4 7 * 8
TIME 3 - *

(2,2)
5
TIME 4 - + 9

n
NOP
 Latency bound of 4
 Fast multipliers for { v1 , v2 , v3 }
 Slower multiplier can be used elsewhere
 Less sharing
 Minimum-latency design uses fast multipliers only
 Impossible to use slow multipliers

(c) Giovanni De Micheli 35

 Module 2
 Objectives
 Data path generation
 Control synthesis

(c) Giovanni De Micheli 36

 Data path synthesis
 Applied after resource binding
 Connectivity synthesis:
 Connection of resources to multiplexers busses and registers
 Control unit interface
 I/O ports
 Physical data path synthesis
 Specific techniques for regular datapath design
 Regularity extraction

(c) Giovanni De Micheli 37

 Example
REGISTERS
a
3
d
x enable
y
u
r
1r
2
Mux control

ALU control (+,-,<)

AL
*

c
U

DATA-PATH CONTROL-UNIT

(c) Giovanni De Micheli 38

 Control synthesis

 Synthesis of the control unit

 Logic model:
 Synchronous FSM
 Physical implementation:
 Hard-wired or distributed FSM
 Microcode

(c) Giovanni De Micheli 39

 Example
0
NOP

1 2 6 8 1
TIME 1
* * * * + 0

7 9 1
TIME 2 3
* * + < 1
4
TIME 3 -
TIME 4 5
-
reset reset

NOP n
1,2,6,8,10 reset’ 3,7,9,11
s1 s2
reset
reset’

5
s4 reset’ 4
s3

(c) Giovanni De Micheli 40

 Summary
 Resource sharing is reducible to vertex coloring or to clique
covering:
 Simple for flat graphs
 Intractable, but still easy in practice, for other graphs

 Resource sharing has several extensions:

 Module selection
 Data path design and control synthesis are conceptually simple
but still important steps in synthesis
 Generated data path is an interconnection of blocks
 Control is one or more finite-state machines

(c) Giovanni De Micheli 41