Skip to content

Enforce function pipelining when using io_parallel with Resource strategy

Javier Duarte requested to merge github/fork/vloncar/pipeline_functions into main

Created by: vloncar

Description

Resource strategy implies using DATAFLOW at the top-level function, which doesn't cause loop unrolling and function pipelining. In this case, functions that don't explicitly call #pragma HLS PIPELINE II=... won't be pipelined and will suffer from long latency. We pipeline most functions, but there were a few missing. This PR adds this directive to the functions missing it.

While I was at it, I also cleaned up the pooling implementation and fixed the concatenate1d for io_stream.

The transpose_3d function (corresponding to the Permute Keras layer) is left out, because this function needs to be replaced with something synthesizable. This will be addressed in a follow up PR.

Type of change

  • Bug fix (non-breaking change that fixes an issue)

Tests

The included pytest only tests the validity of the merge functions. To test the synthesis difference, same functions can be used, just use Resource strategy and call build to see it before and after the change.

For example:

import hls4ml
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Add, Average, Maximum, Minimum, Multiply, Subtract

def test_merge(merge_layer):
    input_shape = (10, 10, 3)

    in1 = Input(shape=input_shape)
    in2 = Input(shape=input_shape)
    out = merge_layer()([in1, in2])
    
    model = Model(inputs=[in1, in2], outputs=out)
    model.compile(optimizer='adam', loss='mse')

    config = hls4ml.utils.config_from_keras_model(model, default_precision='ap_fixed<32,16>')
    config['Model']['Strategy'] = 'Resource'
    output_dir = 'hls4mlprj_merge_{}'.format(merge_layer.__name__.lower())
    hls_model = hls4ml.converters.convert_from_keras_model(model, hls_config=config, output_dir=output_dir, io_type='io_parallel')
    report = hls_model.build(synth=True)
    print(report['CSynthesisReport'])

for cls in [Add, Average, Maximum, Minimum, Multiply, Subtract]:
    test_merge(cls)

Checklist

  • I have read the guidelines for contributing.
  • I have commented my code, particularly in hard-to-understand areas.
  • My changes generate no new warnings.
  • I have added tests that prove my fix is effective or that my feature works.B

Merge request reports