search

Xilinx / BNN-PYNQ

Quantized Neural Networks (QNNs) on PYNQ
https://xilinx.github.io/finn/
BSD 3-Clause "New" or "Revised" License
671 stars 306 forks source link

Source code for low level driver #29

Closed eerobert closed 6 years ago

eerobert commented 6 years ago

Hi,

I am porting the LFC to a smaller device and found that it cannot fits in. So I modified the whole thing to SFC and it's already working on HLS CSIM.

Now I have a problem on loading the .so and would like to seek for your help. It would be very nice if you can at least release the API definition of those function in .so

Or if possible source code will definitely good for further customization.

giuliogamba commented 6 years ago

Hi,

the source code for the generation of the .so is open-sourced as well here . You can execute the command make-sw.sh script on the PYNQ board and compile it as explained here. The important thing is that, at compilation time, the main_python.cpp is linking the config.h to get information about layer dimensions. Hope this help.

eerobert commented 6 years ago

Hi,

OO.. thanks. I miss that point.

Thanks so much.

eerobert commented 6 years ago

Hi,

I tried to compiled the LFC PYTHON SW version, but the SW inference fail. It seems like the recompiled python_sw-lfc-pynq.so is different from the bundled so. Would you mind provide me more instruction on this issue?

giuliogamba commented 6 years ago

The .so might differ since it depends on the Vivado HLS library your compiling against. I've just re-compiled locally linking library from HLS 2017.4 and it works on my side, so I am not able to reproduce. Which version of the HLS library are you using?

eerobert commented 6 years ago

Thanks for your prompt reply.

I am now able to make python_sw working. it seems like the on board g++ is not the same arm-linux-gnueabihf-g++ 2017.4.

But I still having problem in running python_hw.

I wish to know why the "thePlatform" under foldedmv-offload is using the class donutdriver instead of xlnkdriver. The program throw an exception in the line foldedmv-offload.cpp:398 as the donutdriver return 0 on allocAccelBuffer.

giuliogamba commented 6 years ago

Hi,

are you running the executable as root?

eerobert commented 6 years ago

Hi,

I am using Xilinx SDK to step debug the executable.

After reviewing the code, I wonder why the platform-xlnk.cpp:63 cannot override the donutdriver virtual allocAccelBuffer with xlnkdriver allocAccelBuffer.

eerobert commented 6 years ago

Hi,

After rebuild everything, it works. Thanks