merrymercy commented on a change in pull request #6073:
URL: https://github.com/apache/incubator-tvm/pull/6073#discussion_r456549272
##########
File path: src/auto_scheduler/compute_dag.cc
##########
@@ -276,10 +276,18 @@ std::pair<te::Schedule, Array<te::Tensor>>
ComputeDAG::ApplySteps(
// return value, so the ApplyToSchedule is not able to be merged to single
interface
if (auto ps = step.as<ReorderStepNode>()) {
ps->ApplyToSchedule(stages, stage_to_axes);
+ } else if (auto ps = step.as<ComputeAtStepNode>()) {
Review comment:
Why do you use a different order from the order in our internal repo?
I propose we can use one of the following two orders.
1. Alphabetical order
2. Logical order from easy to hard.
- primitives working on one stage: reorder, split, fuse, annotation
- primitives working on multiple stages: compute_at, compute_root,
compute_inline
- primitive adding new stages: cache_read, cache_write, ...
And we should keep the same order in all places.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
