Add support for filt_height==1 for streaming quartus conv2d (fastmachinelearning#886)

Author: jmitrevs

hls4ml/templates/quartus/firmware/nnet_utils/nnet_conv2d_stream.h

@@ -69,7 +69,7 @@ template <class data_T, typename CONFIG_T>
 void shift_line_buffer_2d(
     const data_T &in_elem,
     nnet::shift_reg<typename data_T::value_type, CONFIG_T::pad_left + CONFIG_T::in_width + CONFIG_T::pad_right>
-        line_buffer[CONFIG_T::filt_height - 1][CONFIG_T::n_chan],
+        line_buffer[MAX(CONFIG_T::filt_height - 1, 1)][CONFIG_T::n_chan],
     typename data_T::value_type shift_buffer[CONFIG_T::filt_height][CONFIG_T::n_chan]) {
 // For every channel, insert the incoming pixel at end of the shift buffer
 UpdateBuffer:

test/pytest/test_pointwiseconv.py

@@ -11,10 +11,8 @@
 
 padds_options = ['same', 'valid']
 chans_options = ['channels_last']
-io_type_options = ['io_parallel', 'io_stream']
 strides1d_options = [(1,), (2,)]
 strides2d_options = [(1, 1), (2, 2)]
-strategy_options = ['Latency', 'Resource']
 
 
 @pytest.mark.parametrize('chans', chans_options)
@@ -24,6 +22,7 @@
     'backend, io_type, strategy',
     [
         ('Quartus', 'io_parallel', 'resource'),
+        ('Quartus', 'io_stream', 'resource'),
         ('Vivado', 'io_parallel', 'resource'),
         ('Vitis', 'io_parallel', 'resource'),
         ('Vivado', 'io_parallel', 'latency'),
@@ -54,7 +53,7 @@ def test_pointwiseconv1d(chans, padds, strides, backend, io_type, strategy):
     X_input = np.random.rand(100, *input_shape)
     keras_prediction = model.predict(X_input)
 
-    default_precision = 'ac_fixed<32,16,true>' if backend == 'Quartus' else 'ap_fixed<32,16>'
+    default_precision = 'fixed<32,16>'
     config = hls4ml.utils.config_from_keras_model(model, default_precision=default_precision)
     config['Model']['Strategy'] = strategy
 
@@ -70,7 +69,9 @@ def test_pointwiseconv1d(chans, padds, strides, backend, io_type, strategy):
     hls_model.compile()
     hls_prediction = hls_model.predict(X_input).reshape(keras_prediction.shape)
 
-    assert 'Pointwise' in list(hls_model.graph.values())[1].class_name
+    if not (backend == 'Quartus' and io_type == 'io_stream'):
+        # Quartus io_stream does not currently have a special pointwise implementation
+        assert 'Pointwise' in list(hls_model.graph.values())[1].class_name
     np.testing.assert_allclose(hls_prediction, keras_prediction, rtol=0, atol=0.001)
 
 
@@ -81,6 +82,7 @@ def test_pointwiseconv1d(chans, padds, strides, backend, io_type, strategy):
     'backend, io_type, strategy',
     [
         ('Quartus', 'io_parallel', 'resource'),
+        ('Quartus', 'io_stream', 'resource'),
         ('Vivado', 'io_parallel', 'resource'),
         ('Vivado', 'io_parallel', 'latency'),
         ('Vivado', 'io_stream', 'latency'),
@@ -107,7 +109,7 @@ def test_pointwiseconv2d(chans, padds, strides, backend, io_type, strategy):
     X_input = np.random.rand(100, *input_shape)
     keras_prediction = model.predict(X_input)
 
-    default_precision = 'ac_fixed<32, 9, true>' if backend == 'Quartus' else 'ap_fixed<32, 9>'
+    default_precision = 'fixed<32, 9>'
 
     config = hls4ml.utils.config_from_keras_model(model, default_precision=default_precision)
     config['Model']['Strategy'] = strategy
@@ -125,7 +127,9 @@ def test_pointwiseconv2d(chans, padds, strides, backend, io_type, strategy):
     hls_model.compile()
     hls_prediction = hls_model.predict(X_input).reshape(keras_prediction.shape)
 
-    assert 'Pointwise' in list(hls_model.graph.values())[1].class_name
+    if not (backend == 'Quartus' and io_type == 'io_stream'):
+        # Quartus io_stream does not currently have a special pointwise implementation
+        assert 'Pointwise' in list(hls_model.graph.values())[1].class_name
     np.testing.assert_allclose(hls_prediction, keras_prediction, rtol=0, atol=0.001)