[BYOC] Configurable optimize pass for PartitionGraph #6777
Conversation
| @@ -62,7 +72,7 @@ def partition_for_arm_compute_lib(mod, params=None): | |||
| transform.InferType(), | |||
| transform.MergeComposite(arm_compute_lib_pattern_table()), | |||
| transform.AnnotateTarget("arm_compute_lib"), | |||
| transform.PartitionGraph(), | |||
| transform.PartitionGraph(optimize), | |||
There was a problem hiding this comment.
Why cant we just run the optimize pass before partition graph here ? Is there a reason why it has to be inside PartitionGraph() ?
There was a problem hiding this comment.
This pass applies to each partitioned function, so it has to be called after the partitioned function has been created. See #6068
There was a problem hiding this comment.
Hmm, then why not after PartitionGraph ? (using kCompiler as a filter) ?
There was a problem hiding this comment.
As being said in that PR, we can definitely do that. In fact we are planning an RFC right now, so depending on one the RFC, we probably don't need this PR anymore.
There was a problem hiding this comment.
OK, sounds good :). I think PartitionGraph should just perform partitioning unless its really unavoidable to perform the needed post-processing after the pass.
|
@comaniac consider close this PR as it seems to be stale? |
Ah sorry we don't have bandwidth to work on this recently. I'll close the PR for now. |
Currently,
PartitionGraphimplicitly invokesrelay.ext.*.optimizepass for each partitioned Relay function, but this is hard to maintain and trace. This PR adds an optional argument toPartitionGraphpass to make it explicit.cc @zhiics @masahi @lhutton1