basic loop parallelization

Basic Loop ParallelizationBasic Loop Parallelization

Taewook Eom

RTOS Lab., School of EECS, SNU

2

OutlineOutline

Data dependences

Loop parallelization

How to extract data dependence information

3

Loop Parallelization GoalLoop Parallelization Goal

DoAll loopsLoops whose iterations can execute in parallel

Example

For (i = 0; i < n; i++)

A[i] = B[i] + C[i]

Statically assign each iteration to a processor

4

Data Dependence of ScalarsData Dependence of Scalars

Data DependencesTrue dependence: write read

Anti dependence: read write

Output dependence: write write

Data dependence exists from a dynamic instance i to i’ iff

either i or i’ is a write operation

i and i’ refer to the same variable

i executes before i’

5

Data Dependences for Arrays (1)Data Dependences for Arrays (1)

DoAll loop vs. DoAcross loop

Loop-carried vs. Loop-independent dependenceDependence exists across iterations vs. within one iteration

Recognize DoAll loops (intuitively)Find data dependences in loops

No dependences crossing iteration boundaries ⇒ Parallelizable

for (i = 2; i < 5; i++)A[i] = A[i-2] + 1

for (i = 0; i < 3; i++)A[i] = A[i] + 1

6

Data Dependences for Arrays (2)Data Dependences for Arrays (2)

AssumptionsA loop is in canonical form

• Index runs from 1 to N by increments of 1

Perfectly nested loops• Only the innermost loop has statements other than for

statements

ExampleFor (i1 = 1; i1 <= n1; i1++)

For (i2 = 1; i2 <= n2; i2++)

For (i3 = 1; i3 <= n3; i3++)

A[i1, i2, i3] = A[i1, i2, i3] + 1;

7

Iteration Space (1)Iteration Space (1)

Iteration is represented as coordinates in iteration space

n-dimensional discrete space for n-deep loops

Iteration space diagrams show dependences between all iterations of a loop

Vs. Dependence graph shows all dependences between two statements in a loop

Example

i

j

Iteration Space Traversal

FOR (i = 0; i < 5; i++) FOR (j = 0; j < 7; j++) A[i, j] = A[i-1, j-1] + 1

8

Iteration Space (2)Iteration Space (2)

Sequential execution order of iterationLexicographic order

[0,0], [0,1], … , [0,6], [1,1], … , [1,6], [2,2], …

Iteration is lexicographically less than , iff

Examples• (1,6,0) < (2, 0, 0)

• (0,3,1) < (0,5,0)

cccc iiandiiiitsc )..,()..,.(. 1111

i

i

9

Distance Vectors (1)Distance Vectors (1)

A loop has a distance vectorIf there exists data dependence from a node to a later node and

Indicates how many iterations apart are source and sink of dependence

Example

Distance vector = [1,1]

FOR (i = 0; i < 5; i++) FOR (j = 0; j < 7; j++) A[i, j] = A[i-1, j-1] + 1

iid

i

i

d

i

j

Iteration Space

10

Distance Vectors (2)Distance Vectors (2)

Since then is lexicographically

greater than or equal to 0

All dependence must point forward in time

Distance vector (infinitely large set)[0,0], [0,1], … , [0,n], [1,-n], … , [1,0], … , [1,n], … , [n,-n], … , [n,0], … , [n,n]

Summary: Direction vectors0 or lexicographically positive

[0,0], [0,+], [+,-], [+,0], [+,+]

ii

d 0d

11

Direction Vectors (1)Direction Vectors (1)

The sign of the distance

There are different notations< or +: dependence from earlier to later iteration

= or 0: dependence within the same iteration

> or -: dependence from later to earlier iteration

ExampleFor (i = 0; i < n; i++) For (j = 0; j < n; j++) A[i, j] = A[i-2, j+3]

i = (1, 4)i’ = (3, 1)d = i’ - i = (2, -3)Direction = (<, >) or (+, -)

12

Direction Vectors (2)Direction Vectors (2)

What can the direction vectors for 3-D loops be like?

[? ? ?]

[0 ? ?]

[0 0 ?]

[0 0 0]

[0 0 － ]

[0 0 +]

[0 + ?]

[+ ? ?]

[0 － ?]

[ － ? ?]

Backwarddirection

…

…

13

Dependence VectorsDependence Vectors

Dependence VectorSet of distance vectors

ExampleFor (I = 0; i < n; i++)

For (j = 0; j < n; j++) {

A[i][j] = A[i][j] + 1;

B[i][j] = B[i][j-1] + 1;

}

Test for parallelization:A loop with data dependence is parallelizable if for all = (0)d

Dd

Distance vectors : [0, 0], [0, 1]Dependence vectors : [0,[0, 1]]

14

n-Dimensional Loopsn-Dimensional Loops

1

0FOR (i = 0; i < n; i++) FOR (j = 0; j < n; j++) A[i, j] = A[i, j-1] + 1

FOR (i = 0; i < n; i++) FOR (j = 0; j < n; j++) A[i, j] = A[i-1, j] + 1

FOR (i = 0; i < n; i++) FOR (j = 0; j < n; j++) A[i, j] = A[i-1, j-1] + 1

0

1

1

1

j

i

j

i

j

i

Loop i isparallelizable

Loop j isparallelizable

Loop j isparallelizable

15

Test for ParallelizationTest for Parallelization

is loop-carried at level iif di is the first non-zero element

⇒ does not affect the parallelizability of inner loop

Ex) (0,0,2,0,1, …) : level 3 dependence

The i th loop of an n-dimensional loop is parallelizable if there does not exist any level i dependences

The i th loop of an n-dimensional loop is parallelizable for all dependences , either

(d1, … ,di-1) > 0 (= lexicographically forward)

(d1, … ,di) = 0

)...,,( 1 nddd

),...,,( 21 ndddd

16

Improving ParallelizabilityImproving Parallelizability

Scalar Privatization

Loop-carried dependence is an anti dependence• No value is really carried across the loop

Data is local to each loop iteration• Eliminating loop carried dependences

Scalar privatization should be applied before any serious parallelization attempt

float t;for (i = 0; i < n; i++) { t = A[i]; b[i] = t * t;}

for (i = 0; i < n; i++) { float t; t = A[i]; b[i] = t * t;}

True dependence(loop-indep.)

Anti dependence(loop-carried)

Not parallelizable!! Parallelizable!!

17

SummarySummary

Data Dependenceread/write

same variable

in direction of sequential ordering

RepresentationIteration space

Dependence vectors• Distance vectors

• Direction vectors

Parallelization Testing