Quartus streaming Dense layer

Author: bo3z

hls4ml/backends/quartus/passes/core_templates.py

@@ -36,7 +36,7 @@
 
 dense_function_template = 'nnet::dense_{strategy}<{input_t}, {output_t}, {config}>({input}, {output}, {w}, {b});'
 
-dense_include_list = ['nnet_utils/nnet_dense.h', 'nnet_utils/nnet_dense_compressed.h']
+dense_include_list = ['nnet_utils/nnet_dense.h', 'nnet_utils/nnet_dense_compressed.h', 'nnet_utils/nnet_dense_stream.h']
 
 class DenseConfigTemplate(LayerConfigTemplate):
     def __init__(self):

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

@@ -0,0 +1,50 @@
+#ifndef NNET_DENSE_STREAM_H_
+#define NNET_DENSE_STREAM_H_
+
+#include "nnet_common.h"
+#include "nnet_types.h"
+#include "nnet_dense.h"
+
+namespace nnet {
+
+template<class data_T, class res_T, typename CONFIG_T>
+void dense_resource(
+    stream<data_T> &data_stream, 
+    stream<res_T> &res_stream,
+    const typename CONFIG_T::weight_t weights[CONFIG_T::n_in*CONFIG_T::n_out],
+    const typename CONFIG_T::bias_t   biases[CONFIG_T::n_out])
+{
+    hls_register typename data_T::value_type data[CONFIG_T::n_in];
+    hls_register typename res_T::value_type res[CONFIG_T::n_out];
+    
+    DataPrepare: 
+    #pragma ii 1
+    for(int i_in = 0; i_in < CONFIG_T::n_in / data_T::size; i_in++) {
+        data_T data_pack = data_stream.read();
+        DataPack: 
+        #pragma unroll
+        for (int i_pack = 0; i_pack < data_T::size; i_pack++) {
+            data[i_in * data_T::size + i_pack] = data_pack[i_pack];
+        }
+    }
+
+    dense_resource<typename data_T::value_type, typename res_T::value_type, CONFIG_T>(data, res, weights, biases);
+
+    ResWrite: 
+    #pragma ii 1
+    for(unsigned i_out = 0; i_out < CONFIG_T::n_out / res_T::size; i_out++) {
+        res_T res_pack;
+        ResPack:
+        #pragma unroll 
+        for (int i_pack = 0; i_pack < res_T::size; i_pack++) {
+            res_pack[i_pack] = res[i_out * res_T::size + i_pack];
+        }
+
+        res_stream.write(res_pack);
+    }
+}
+
+
+}
+
+#endif

test/pytest/test_keras_api.py

@@ -16,7 +16,8 @@
 test_root_path = Path(__file__).parent
 
 @pytest.mark.parametrize('backend', ['Vivado', 'Quartus'])
-def test_dense(backend):
+@pytest.mark.parametrize('io_type', ['io_parallel', 'io_stream'])
+def test_dense(backend, io_type):
     model = tf.keras.models.Sequential()
     model.add(Dense(2,
               input_shape=(1,),
@@ -37,9 +38,9 @@ def test_dense(backend):
     keras_prediction = model.predict(X_input)
 
     config = hls4ml.utils.config_from_keras_model(model)
-    output_dir = str(test_root_path / f'hls4mlprj_keras_api_dense_{backend}')
+    output_dir = str(test_root_path / f'hls4mlprj_keras_api_dense_{backend}_{io_type}')
 
-    hls_model = hls4ml.converters.convert_from_keras_model(model, hls_config=config, output_dir=output_dir, backend=backend)
+    hls_model = hls4ml.converters.convert_from_keras_model(model, hls_config=config, output_dir=output_dir, backend=backend, io_type=io_type)
 
     hls_model.compile()
 
@@ -66,7 +67,8 @@ def test_dense(backend):
                                                  Activation(activation='sigmoid', name='Activation')])
                                                  #ThresholdedReLU(theta=1.0)])
 @pytest.mark.parametrize('backend', ['Vivado', 'Quartus'])
-def test_activations(activation_function, backend):
+@pytest.mark.parametrize('io_type', ['io_parallel', 'io_stream'])
+def test_activations(activation_function, backend, io_type):
     model = tf.keras.models.Sequential()
     model.add(Dense(64,
               input_shape=(1,),
@@ -79,8 +81,8 @@ def test_activations(activation_function, backend):
     X_input = np.random.rand(100,1)
     keras_prediction = model.predict(X_input)
     config = hls4ml.utils.config_from_keras_model(model)
-    output_dir = str(test_root_path / 'hls4mlprj_keras_api_activations_{}_{}'.format(activation_function.__class__.__name__, backend))
-    hls_model = hls4ml.converters.convert_from_keras_model(model, hls_config=config, output_dir=output_dir, backend=backend)
+    output_dir = str(test_root_path / 'hls4mlprj_keras_api_activations_{}_{}_{}'.format(activation_function.__class__.__name__, backend, io_type))
+    hls_model = hls4ml.converters.convert_from_keras_model(model, hls_config=config, output_dir=output_dir, backend=backend, io_type=io_type)
     hls_model.compile()
     hls_prediction = hls_model.predict(X_input)