Output diff when enabling CMSIS_NN_USE_REQUANTIZE_INLINE_ASSEMBLY
https://github.com/ARM-software/CMSIS-NN/pull/171 introduce optional flag CMSIS_NN_USE_REQUANTIZE_INLINE_ASSEMBLY. When this is enabled there are uplifts and as well as downlifts depending on the target and model. For one particular combination that is Arm Cortex-M7 using armclang there was an output diff observed for SSD Mobilenet v3.
This is hard to fix if you don't have access to the model and the compiler. We may be able to extract the "issue" from the model into a unit test to ease debugging.
The MDK Community Edition license will allow full access to the Cortex-M FVP models and the Arm Compiler. Since this is a non-commercial use-case, this is a perfect scenario to use it. Install and activate it either using an IDE of your choice: https://www.keil.arm.com/mdk-community/ Or in a GH CI environment as shown here: https://github.com/Arm-Examples/AVH_CI_Template/blob/main/.github/workflows/basic.yml Or on the command line https://learn.arm.com/learning-paths/embedded-and-microcontrollers/vcpkg-tool-installation/licenseactivation/ This should equip anyone with the tools required to reproduce this.