ACL Execution Provider Performance

[JAEWOOKe]

Dear all,

I'm running ONNX Runtime 1.6.0 version with ARM Compute Library 19.05 on board with zynqmp (with ARM-A53).

I have profiled ResNet-50 and show CPUExecutionProvider is much faster than ACLExecutionProvider.

I thought ACL would accelerate ARM core so I'm a little confused.

I have build ACL with the following command.

scons -j 4 extra_cxx_flags="-fPIC" Werror=0 debug=0 asserts=0 neon=1 opencl=0 os=linux arch=arm64-v8a examples=1

These are my profiling data for CPU and ACL on ResNet-50.

CPU
https://drive.google.com/file/d/1cFKMR9KiK4zMSWWPkKZLoNJmVjzaxQs8/view?usp=sharing
ACL
https://drive.google.com/file/d/13ZVTjSKVb2m9HO1Sd_0ZTUaSZIywUgSv/view?usp=sharing

Can anyone give me any advice?

Thanks,
Jaewook