refactor 2d inner reduction heuristics #3333
Conversation
|
!build |
|
!test --diff --diff-bench |
should the target of this one be |
|
!test --diff --diff-bench |
Yes, |
| // Max vectorization factor | ||
| int64_t vect_factor = std::min( | ||
| scheduler_utils::lastPow2(max_unroll), (int64_t)max_vectorize_factor); | ||
| int64_t after_vect = total_reduction_numel / vect_factor; |
There was a problem hiding this comment.
This was a ceilDiv earlier.
Is it an intended change?
There was a problem hiding this comment.
Yes, vect_factor was derived from inner most dim, total_reduction_numel must be disivible by vect_factor
Following #3331
refactor 2d inner reduction heuristics, prepare for further heuristics tunning.