| |
Node Splitting
- Loop parallelization is impossible when the statements are involved in a dependence cycle
- Sometimes dependence cycles can be broken resulting in total/partial parallelization of loop
- Flow dependence cycles are very hard to break
- Anti-and output-dependence cycles can be broken by renaming variables
- Following loop can not be parallelized
for I = 1,n
A[i] = B[i] + C[i]
D[i] = A[i-1] + A[i+1]
endfor |
Node Splitting …
for I = 1,n
A[i] = B[i] + C[i]
temp[i] = A[i+1]
D[i] = A[i-1] + temp[i]
endfor |
for I = 1,n
temp[i] = A[i+1]
A[i] = B[i] + C[i]
D[i] = A[i-1] + temp[i]
endfor |
temp[1..n] = A[2..n+1]
A[1..n] = B[1..n] + C[1..n]
D[1..n] = A[0..n-1] + temp[1..n] |
Cycle Shrinking
- In many loops dependence cycles are impossible to break
- Such dependence cycles usually involve only few dependences
- Cycle shrinking is used to extract any parallelism that may be present in the loop
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
|
|
