-
Notifications
You must be signed in to change notification settings - Fork 218
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
NHWC miopenOpTensor Add incorrect results #2860
Comments
Additionally adding the log of running the reproducer
|
As far as I know, miopen tensor operations do not support non-singular stride for the last dimension. It means that any layout change using the strides is not possible, only explicit transpose operation can help, where you physically move C to the last dimension. I'm wondering why the library does not report the unsupported layout. |
This is a bug, I suspect. Can you please set bug correctness and urgency_high on this ticket. |
While trying to do a NHWC conv bias I encountered incorrect results some investigation showed that the miopenOpTensor Add operation produced incorrect results for NHWC format(Works correctly for NCHW).
The used MIOpen library is part of the rocm-5.6.1 package.
I could not find specification regarding the supported formats for miopenOpTensor.
Is this a known/expected limitation?
Additionally looking for similar issues this #2001 seemed related and pointed NHWC not working correctly and a reorder of the tensor data would be the workaround.
#2828 shows reorders not to be working.
What would the solution be? Is there a workaround to this?
I have attached a reproducer showing this behavior:
The text was updated successfully, but these errors were encountered: