Trying to port People Counting application on ZCU104 platform, where we want to Off load ML Part to FPGA and other Pre/Post processing modules wanted to use ARM CPU Cores. When we run the application we see that Pre/Post processing modules were taking lot of time. So we wanted to implement using Neon Intrinsics .
Here we see issue, when we compiled float and neon code with -O3 flag we see same latency numbers .
Can you please suggests any tips or how to analyse it further on this?
Thanks and Regards,
You probably need to share a bit more on what you're getting / how you're compiling etc. Also what cpu cores are in the platform.
As a starting point try: https://community.arm.com/developer/tools-software/oss-platforms/b/android-blog/posts/arm-neon-optimization which mentions Neon latency issues.
The clang compiler is very mature now for Neon, so a naïve intrinsics implementation is often worse than what the compiler manages. It is important to lay out the data and loops in a way that parallelises easily to maximise the SIMD utilization - this can improve the non-intrinsics performance, as well as allowing you better intrinsics optimization.
Thanks a lot for your time and reply on this .
I am looking into the optimisation link you have shared here and update you further on this.
how you're compiling?
g++ -O3 decode.cpp -o decode
Decode is a c++ floating point implementation where I have decode the bounding box.
View all questions in Machine Learning forum