merrymercy commented on a change in pull request #7156:
URL: https://github.com/apache/tvm/pull/7156#discussion_r548452026
##########
File path: python/tvm/auto_scheduler/search_task.py
##########
@@ -265,10 +277,6 @@ def apply_best(self, log_file, layout_rewrite_option=None):
"Cannot find any valid schedule for %s in file %s" %
(self.workload_key, log_file)
)
- if layout_rewrite_option is None:
- layout_rewrite_option = LayoutRewriteOption.NO_REWRITE
- if self.target.kind.name == "llvm":
- layout_rewrite_option =
LayoutRewriteOption.INSERT_TRANSFORM_STAGE
Review comment:
This change breaks the current behavior of our tutorial.
In the single op tutorial, we want to be able to control the layout rewrite
optimization with just one line
https://github.com/apache/tvm/blob/0b2ab56705a1c26aef402d1c0d48a887fc76dbd3/tutorials/auto_scheduler/tune_matmul_x86.py#L63
With this attrs, we should do layout rewrite by inserting a new stage.
Without this attr, we should not do layout rewrite.
But after your change, we need to control the layout rewrite in two places,
and the tutorial won't perform layout rewrite.
I propose the following changes:
1. set the default value of `layout_rewrite_option` in `apply_best` as
`None`. If it is `None`, use `task.layout_rewrite_option` in this function.
2. set the default value of `layout_rewrite_option` in `SearchTask` as
'None'. If the target is cpu/mali, use `INSERT_TRANSFORM_STAGE`. If the target
is cuda, use `NO_REWRITE`
----------------------------------------------------------------
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]
