Previous | Next --- Slide 34 of 48
Back to Lecture Thumbnails
qqkk

I'm wondering what's the implementation of the builtin reduceAdd. Is it O(N), or O(N/P+P), or O(logN) to add all N numbers?

fleventyfive

I think here, reduceAdd has a constant time implementation, since it is used for every single addition. In my mind, all it does is acquire a lock to update diff, add the absolute value of the value by which A[i,j] has changed, and release the lock.

I think what you have in mind is a reduceAdd that synchronizes at the end of all threads, and adds their individual private copies into a single data value. For that, I believe the complexity will be O(NumberOfThreads).

@qqkk could you please elaborate on the other bounds you came up with, like O(logN)? I'm curious.

qqkk

Sorry I messed up with my question. reduceAdd() should be O(NoOfThread). It's just the implementation here to calculate N numbers, complexity should be O(N/P+P).

ZhuansunXt

Assignment here should be performed statically by the compiler, assigning each iteration to program's instances (workers). For example consider it as ISPC foreach semantic.