Add precision inherition for when generating stream clone

hls4ml/__init__.py

@@ -1,3 +1,5 @@
+import os
+
 from hls4ml import converters, report, utils  # noqa: F401
 
 try:
@@ -22,3 +24,5 @@ def reseed(newseed):
         torch.manual_seed(newseed)
     except ImportError:
         print('\nPyTorch seed not set')
+
+    os.environ['RANDOM_SEED'] = f'{newseed}'

hls4ml/backends/fpga/passes/clone.py

@@ -20,21 +20,19 @@ def initialize(self):
 class CloneFunctionTemplate(FunctionCallTemplate):
     def __init__(self):
         super().__init__(Clone, include_header=clone_include_list)
-        self.template = None  # to be filled once number of clones known
 
     def format(self, node):
         params = self._default_function_params(node)
         for i, _output in enumerate(node.outputs):
             params['output' + str(i + 1)] = node.variables[node.outputs[i]].name
 
-        if self.template is None:
-            self.template = (
-                'nnet::clone_stream<{input_t}, {output_t}, {size}>({input}, '
-                + ', '.join(['{output' + str(i + 1) + '}' for i in range(len(node.outputs))])
-                + ');'
-            )
+        template = (
+            'nnet::clone_stream<{input_t}, {output_t}, {size}>({input}, '
+            + ', '.join(['{output' + str(i + 1) + '}' for i in range(len(node.outputs))])
+            + ');'
+        )
 
-        return self.template.format(**params)
+        return template.format(**params)
 
 
 def register_clone(backend):
@@ -79,13 +77,17 @@ def transform(self, model, node):
                     attrs = {'size': np.prod(out_var.shape)}
                     idx = layer.inputs.index(output)
                     layer.inputs[idx] = output + '_cpy' + str(i)
-                clone_layer = model.make_node(
+
+                clone_layer: Clone = model.make_node(
                     Clone,
                     'clone_' + node.name,
                     attrs,
                     [output],
                     [output + '_cpy' + str(i + 1) for i in range(len(output_map[output]))],
                 )
+                for i in range(len(output_map[output])):
+                    key = output + '_cpy' + str(i + 1)
+                    clone_layer.attributes[key].type = node.attributes['result_t']
                 model.insert_node(clone_layer)
                 transformed = True
 

test/pytest/test_stream_clone.py

@@ -0,0 +1,93 @@
+from pathlib import Path
+
+import numpy as np
+import pytest
+from keras.layers import Add, Dense
+from tensorflow import keras
+
+from hls4ml.converters import convert_from_keras_model
+
+test_root_path = Path(__file__).parent
+
+
+@pytest.fixture(scope='module')
+def model_clone_precision_inherition():
+    inp = keras.Input(shape=(10,), name='inp')
+    x = Dense(10, name='x')(inp)
+    y = Dense(10, name='y')(inp)
+    out = Add(name='out')([x, y])
+    model = keras.Model(inp, out)
+    return model
+
+
+@pytest.fixture(scope='module')
+def model_multi_clone():
+    inp = keras.Input(shape=(10,))
+    x = Dense(10)(inp)
+    y = Dense(10)(inp)
+    z = Dense(10)(inp)
+    xy = Add()([x, y])
+    xy = Add()([xy, y])
+    out = Add()([xy, z])
+    model = keras.Model(inp, out)
+    return model
+
+
+@pytest.fixture(scope='module')
+def data():
+    X = np.random.normal(0, 1, (1000, 10))
+    X = np.clip(X, -16, 15)
+    return X
+
+
+@pytest.mark.parametrize('backend', ['Vivado', 'Quartus', 'Vitis'])
+def test_multi_clone(model_multi_clone, data, backend: str):
+    output_dir = str(test_root_path / f'hls4mlprj_stream_clone_multiclone_{backend}')
+    hls_config = {'Model': {'Precision': 'fixed<32,5>', 'ReuseFactor': 1}}
+    model_hls = convert_from_keras_model(
+        model_multi_clone,
+        backend=backend,
+        output_dir=output_dir,
+        hls_config=hls_config,
+        io_type='io_stream',  # clone only happens with stream io.
+    )
+    model_hls.compile()
+    r_hls = model_hls.predict(data)
+    r_keras = model_multi_clone(data).numpy()
+
+    assert np.allclose(r_hls, r_keras, atol=1e-5, rtol=0)
+
+
+@pytest.mark.parametrize('backend', ['Vivado', 'Quartus', 'Vitis'])
+def test_clone_precision_inherition(model_clone_precision_inherition, data, backend: str):
+    output_dir = str(test_root_path / f'hls4mlprj_stream_clone_precision_{backend}')
+    layer_config = {
+        'inp': {'Precision': 'fixed<32,5>'},
+        'x': {'Precision': 'fixed<32,5>'},
+        'x_linear': {'Precision': 'fixed<32,5>'},
+        'y': {'Precision': 'fixed<32,5>'},
+        'y_linear': {'Precision': 'fixed<32,5>'},
+        'out': {'Precision': 'fixed<32,5>'},
+    }
+    hls_config = {'Model': {'Precision': 'fixed<1,0>', 'ReuseFactor': 1}, 'LayerName': layer_config}
+    model_hls = convert_from_keras_model(
+        model_clone_precision_inherition,
+        backend=backend,
+        output_dir=output_dir,
+        hls_config=hls_config,
+        io_type='io_stream',  # clone only happens with stream io.
+    )
+    assert model_hls.graph['clone_inp'].attributes['inp_cpy1'].type.precision.width == 32
+    assert model_hls.graph['clone_inp'].attributes['inp_cpy1'].type.precision.integer == 5
+    assert model_hls.graph['clone_inp'].attributes['inp_cpy2'].type.precision.width == 32
+    assert model_hls.graph['clone_inp'].attributes['inp_cpy2'].type.precision.integer == 5
+
+    model_hls.compile()
+    r_hls = model_hls.predict(data)
+    r_keras = model_clone_precision_inherition(data).numpy()
+
+    assert np.allclose(r_hls, r_keras, atol=1e-5, rtol=0)
+
+
+if __name__ == '__main__':
+    test_clone_precision_inherition(model_clone_precision_inherition(), data(), 'Vivado')