Details

SE sends errors to the C7x CPU with accompanying zeroed out data. It zeroes out the data by de-asserting the clr_n pin of the interface-driving register for the data via a sticky "error is present to go to CPU" bit. After an SEBRK SE clears that error bit in preparation to send valid data again, but, in the case of SEBRK prematurely ending the stream, it ends up doing so after pushing the first end-of-stream reference to the interface.

Because of this, the first end-of-stream reference will be created with zeroed out data. When FILLVAL is enabled (set to non-zero fill mode), then that non-zero fill data is supposed to be sent after the end of the stream, but this bug corrupts it for the first reference by forcing the data to actual zero for a single cycle.

Workaround

There is no true workaround for this issue, but the scenario can be avoided by doing any of the following in a given stream:

Not using stream breaks that will end the stream prematurely

Not using end-of-stream references for computation

Specifically not using the first end-of-stream reference for computation

Not using the FILLVAL feature

This issue has been discussed with MMALIB/TIDL teams, who are currently the only users of the FILLVAL feature and the requestors to have it added for J7AEP and J7AHP. End-of-stream references are not utilized for computation, making this bug un-hittable, so performing an ECO to fix it is not currently planned.