The ANN model in pynq perform bad compared with HLS simulation

I think it may be the reason of interface. However pay some attention to the pynq framework may aslo work because sometimes the reason for error could be very simple.

I make a artificial neural network(ANN) model by HLS IDE and integrated it in in Vivado and then transfr the .bit in pynq. The model perform bad in pynq board.

even it perform well in HLS IDE simulation:

I define interface in axi interface for teatures vector input, and then get the interface result with axilite interface. Please ignore the out interface, I use them for debug.


the address is here:
the block design in vivado in here. I choose HP0 because I want a higher transportation time and i choose 32 bits.

Someone may fine I use astype(int) in jupyter notebook for pynq board. I have tried float before, the result is same as int type with Xlnk’s dtype is float.

What exactly is your problem?

Do you mean the performance of the design is slow?
You get the wrong results?
The performance does not match what you expect from the HLS reports?


yeap,the model perform well in hls simulation, so i generated IP with this model in HLS ,but it perform bad in real board T_T

I find the reason, because when the software take the result from HLS io, the result hasn’t been calculated! that is the reason!!! I tryied many times with different design with serval days!!!