ekalda commented on a change in pull request #10763:
URL: https://github.com/apache/tvm/pull/10763#discussion_r836288601
##########
File path: python/tvm/relay/backend/contrib/ethosu/te/convolution.py
##########
@@ -187,7 +187,7 @@ def conv2d_compute(
[1, 0, 0, 0, 0, 0],
[0, 1, 0, 0, 0, 0],
[0, 0, 0, 1, 0, 0],
- [0, 0, 16, 0, 1, -16],
+ [0, 0, 0, 0, 0, ofm_channels],
Review comment:
Initially I didn't make that change since the old transformation matrix
wasn't "harmful" in the case of these operators (it was only a problem for
operators that had weights or some kind of a kernel), but I think it makes
sense to unify the transform for all the NPU ops since then we can indeed
easily define the layout transform matrices in one place and easily reuse them
across all the TEs and also in the tests.
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
