Quartus example

[makoeppel]

Hi all,

is there any example code for version 0.6.0 of hls4ml which uses Quartus as a backend? We are using only Intel FPGAs and for me it's not fully clear from the code how to compile the model for Quartus. I do see on the main branch of the repository that the Quartus backend should be somehow supported. I checked out some examples and changed the backend in the configuration but it never worked out:

import hls4ml

#Fetch a keras model from our example repository
#This will download our example model to your working directory and return an example configuration file
config = hls4ml.utils.fetch_example_model('KERAS_3layer.json')

config["Backend"] = "Quartus" # <--- my "simple" change of the example
print(config) #You can print the configuration to see some default parameters

#Convert it to a hls project
hls_model = hls4ml.converters.keras_to_hls(config)

# Print full list of example models if you want to explore more
hls4ml.utils.fetch_example_list()

Cheers,
Marius

[jmitrevs]

The Quartus support is only in the main branch, so you need to check that out from git. (Ignore the version number of the main branch; it says 0.5.2 I believe.) When you use convert_from_keras_model(...), pass it backend='Quartus' and part=Arria10 or whatever.

[bo3z]

Adding to the previous answer - here is an example of a function to convert a model using the Quartus backend and synthesise it:

def convert_to_hls4ml(keras_model, input_data=None, output_data=None, precision='ac_fixed<16, 6, true>', reuse_factor=1):    
    output_dir = r'hls4ml-intel-example'
    hls_config = hls4ml.utils.config_from_keras_model(keras_model, granularity='name', default_precision=precision, default_reuse_factor=reuse_factor)     
    hls_config['Model']['Strategy'] = 'Resource'

    # Device can be Arria10, Stratix10, Agilex, CyloneV, Cyclone10GX etc. - depends on version of Quartus running locally
    hls_model = hls4ml.converters.convert_from_keras_model(
                            keras_model, 
                            hls_config=hls_config,
                            output_dir=output_dir, 
                            backend='Quartus',
                            part='Stratix10')
    hls_model.compile()

    if input_data is not `None:`
        print('Predicting Keras')
        keras_prediction = keras_model.predict(input_data)
        
        print('Predicting HLS with GCC')
        hls_prediction = hls_model.predict(np.ascontiguousarray(input_data))

        # This only works for classification problems - tweak accordingly
        print('Keras Accuracy: {}'.format(accuracy_score(np.argmax(output_data, axis=1), np.argmax(keras_prediction, axis=1))))
        print('HLS Accuracy (GCC): {}'.format(accuracy_score(np.argmax(output_data, axis=1), np.argmax(hls_prediction, axis=1))))
 
    # Depending on the size of the model and available resources, synthesis can take long
    # Set fpgasynth=True if you want to perform Quartus compilation (Synthesis, Fitter, Timing Analysis)
    # For more information during synthesis, set log level to 2
    hls_model.build(synth=True, fpgasynth=False, log_level=1)

Note that all of the tips covered in hls4ml tutorial are applicable to both Vivado and Quartus - especially when it comes to tweaking the default precision, reuse factor and I/O type. For more

[makoeppel]

Thank you very much! I manged to compile the main branch and was able too compile the example from @bo3z as well.