Test calculation output shape for Broadcast op, relax restrictions for partially dynamic input data #1247
Conversation
mbencer
requested review from
tsocha,
postrational,
arogowie-intel,
ilya-lavrenov and
jane-intel
There was a problem hiding this comment.
I haven't looked through all tests. However after review of spreadsheet with matrix of cases we came to conclusion we should clarify the resulting shapes.
My understanding is that, we should leverage as much as possible information available from target_shape input in case of numpy and explicit broadcasting rules, which I'd like to emphasize are uni-directional modes.
| const bool is_output_rank_static = | ||
| output_shape.rank().is_static() && output_shape[0].is_static(); |
There was a problem hiding this comment.
The variable name is misleading. The conditions are more restrictive than the variable name. In fact they boil down to output_shape.is_static() - since you've already enforced 1D target shape and is_static function checks whether partial shape rank is static as well as all of its dimensions.
There was a problem hiding this comment.
right, I changed it to is_target_shape_known
You are right. I've changed rules for |
mbencer
requested review from
a team and
lazarevevgeny
and removed request for
ilya-lavrenov
Co-authored-by: Adam Osewski <[email protected]>
…r partially dynamic input data (openvinotoolkit#1247)
…r partially dynamic input data (openvinotoolkit#1247)
Tested combinations from CVS-31915