Hello ArmNN/IMX experts,
I have been able to build from sources (outside eIQ context) ArmNN Static librairies for an iMx6 target.
I have built a static version of the libs because I have been disapointed by the performances of the Shared versions that I also built, and was hoping an improvement as some internet litterature mention ...
Using the SHARED ArmNN libs (*.so) I'm able to run an app which executes inference of a TFLite model on CpuRef (CPU-only) and CpuAcc (using NEON).
Using the STATIC ArmNN libs (*.a), the CPURef is working but not the CpuAcc which generates a segmentation fault at runtime.
I have tried, in the gcc command line the "whole-archive" option, in order to include all the libs, but this is not sufficient and provide a huge binary size.
Since the ArmNN framework does not generates any specific warnings at runtime (even if logging is activated) I am stucked.
## Does someone have an idea ?
For information, I have measured an improvement of ~5% by using the CpuRef with STATIC ArmNN built compared with SHARED (default build option in tutorials).
Kind regards,
Nicolas
