Skip to content

MLBEDSW-10573: Fix buffered weights and unbuffered scales

Johan Gunnarsson requested to merge MLBEDSW-10573 into main

When two convolution ops share the same weight tensor, but have different bias tensor or output scaling, we can end up with a convolution op that has buffered weights and unbuffered scales. In this case we didn't properly handle depth slicing of the scales tensor. The SCALE_BASE register was always set to the first slice of encoded scales.

Also change high level command stream log output to print addresses as hex values.

Signed-off-by: Johan Gunnarsson johan.gunnarsson@arm.com Change-Id: I59c067e36e9c67c84b885bfa72380354c3707591

Merge request reports