PYNQ-ZU support with VivadoAccelerator backend

[jacopoabramo]

Greetings,

as requested in issue #623 I tested the integration of the PYNQ-ZU board. I managed to create a bitstream of the model presented in the deployment tutorial, and it works fine on the PYNQ-ZU.

A couple of points related to the port itself:

• the part number is incorrect, it should be xczu5eg-sfvc784-1-e
• The final accuracy of the model inferred on the FPGA is around 16% less than the original model (74% QKeras model, 58% on FPGA) with these settings:

hls4ml.model.optimizer.passes.qkeras.OutputRoundingSaturationMode.layers = ['Activation']
hls4ml.model.optimizer.passes.qkeras.OutputRoundingSaturationMode.rounding_mode = 'AP_RND'
hls4ml.model.optimizer.passes.qkeras.OutputRoundingSaturationMode.saturation_mode = 'AP_SAT'

config = hls4ml.utils.config_from_keras_model(model, granularity='name')
config['LayerName']['softmax']['exp_table_t'] = 'ap_fixed<18,8>'
config['LayerName']['softmax']['inv_table_t'] = 'ap_fixed<18,4>'
for layer in ['fc1', 'fc2', 'fc3', 'output']:
    config['LayerName'][layer]['ReuseFactor'] = 64

I assume though that changing some conversion strategy to increase the accuracy the drop should reduce significantly.

[jacopoabramo]

Some issues I encountered unrelated to hls4ml but related to Vivado:

• due to the revision overflow bug affected by versions up to 2021.2, it would be nice to notify users to adapt their Vivado versions with the patch they provide somewhere in the package documentation;
• it happened to me that for some reason when launching Vivado in batch mode the repository paths are not set, hence the pynq-zu board files were not found. I fixed this by adding a Vivado_init.tcl script which would set the repository paths with the command set_param board.repoPaths {paths to board files}. This script goes in <installation_folder>/Xilinx/Vivado/<vivado_version>/scripts/