cancel
Showing results for 
Search instead for 
Did you mean: 
Highlighted
685 Views
Registered: ‎11-09-2018

Is the FPGA SOC fast enough for a CNN running in a bare metal environment using NEON intrinsics?

I am trying to create an object detector based on the example below. I have converted the keras model trained with over 100 images and converted to C++ using kerasify. The challenge is to convert kerasify to run on the Xilinx Zynq 7000 (Zedboard) by removing STL and using the NEON instrinsics without high level synthesis at this stage. Currently running a modified kerasify using Xilinx libraries, I am getting a maximum of 3 frames per second. Is there any benefits to carry on with the conversion or move straight to HLS?

https://www.pyimagesearch.com/2017/12/18/keras-deep-learning-raspberry-pi/

 

0 Kudos
5 Replies
Highlighted
589 Views
Registered: ‎11-09-2018

Re: Is the FPGA SOC fast enough for a CNN running in a bare metal environment using NEON intrinsics?

I have answered my own question.
0 Kudos
Highlighted
Visitor
Visitor
408 Views
Registered: ‎06-21-2019

Re: Is the FPGA SOC fast enough for a CNN running in a bare metal environment using NEON intrinsics?

Hello,

 

I am also currently facing an issue to run kerasify c++ model on xilinx zynq 7000. Can you please explain how to change the kerasify library to make it work on zynq 7000. 

Thank you.

PS: I have very little knowledge in ARM-NEON programming.

0 Kudos
Highlighted
382 Views
Registered: ‎11-09-2018

Re: Is the FPGA SOC fast enough for a CNN running in a bare metal environment using NEON intrinsics?

Have you tried compiling kerasify using the xilinx compiler? It is standard C++ and libraries which should be supported by the Xilinx compiler which is based on GCC. 

0 Kudos
Highlighted
Visitor
Visitor
355 Views
Registered: ‎06-21-2019

Re: Is the FPGA SOC fast enough for a CNN running in a bare metal environment using NEON intrinsics?

Yeah i tried that. They do compile in the SDK but i am getting runtime error. I think the error is at the LoadModel function. I have the model file on SD card and the standard ifstream is unable to find the model file. May be something needs to be changed at the LoadModel fucntion to find the file on SD card, not sure what exactly would work.
0 Kudos
Highlighted
272 Views
Registered: ‎11-09-2018

Re: Is the FPGA SOC fast enough for a CNN running in a bare metal environment using NEON intrinsics?

From memory you have to replace the function using the xilinx library to load the model from SD card.

0 Kudos