attempted fix for streaming normalize_binary_tanh and normalize_ternary_tanh

Author: jmitrevs

hls4ml/backends/fpga/passes/bn_quant.py

@@ -9,6 +9,7 @@
 batchnorm_quantized_tanh_config_template = """struct config{index} : nnet::batchnorm_quantized_tanh_config {{
     static const unsigned n_in = {n_in};
     static const unsigned n_filt = {n_filt};
+    static const unsigned n_scale_bias = (n_filt == -1) ? n_in : n_filt;
     static const unsigned io_type = nnet::{iotype};
     static const unsigned reuse_factor = {reuse};
 }};\n"""

hls4ml/templates/vivado/nnet_utils/nnet_batchnorm_stream.h

@@ -51,16 +51,16 @@ void normalize(hls::stream<data_T> &data, hls::stream<res_T> &res, typename CONF
 //       Merged Batch Normalization and Quantized Tanh
 // ****************************************************
 template <class data_T, typename CONFIG_T>
-void normalize_binary_tanh(hls::stream<data_T> &data, hls::stream<nnet::array<ap_uint<1>, CONFIG_T::n_in>> &res,
-                           typename data_T::value_type threshold[CONFIG_T::n_in]) {
+void normalize_binary_tanh(hls::stream<data_T> &data, hls::stream<nnet::array<ap_uint<1>, CONFIG_T::n_scale_bias>> &res,
+                           typename data_T::value_type threshold[CONFIG_T::n_scale_bias]) {
     #pragma HLS ARRAY_PARTITION variable=threshold complete
 
 BinaryNormLoop:
     for (int i = 0; i < CONFIG_T::n_in / data_T::size; i++) {
         #pragma HLS PIPELINE
 
         data_T in_data = data.read();
-        nnet::array<ap_uint<1>, CONFIG_T::n_in> out_data;
+        nnet::array<ap_uint<1>, CONFIG_T::n_scale_bias> out_data;
         PRAGMA_DATA_PACK(out_data)
 
     BatchNormPack:
@@ -74,9 +74,9 @@ void normalize_binary_tanh(hls::stream<data_T> &data, hls::stream<nnet::array<ap
 }
 
 template <class data_T, typename CONFIG_T>
-void normalize_ternary_tanh(hls::stream<data_T> &data, hls::stream<nnet::array<ap_int<2>, CONFIG_T::n_in>> &res,
-                            typename data_T::value_type threshold_hi[CONFIG_T::n_in],
-                            typename data_T::value_type threshold_lo[CONFIG_T::n_in]) {
+void normalize_ternary_tanh(hls::stream<data_T> &data, hls::stream<nnet::array<ap_int<2>, CONFIG_T::n_scale_bias>> &res,
+                            typename data_T::value_type threshold_hi[CONFIG_T::n_scale_bias],
+                            typename data_T::value_type threshold_lo[CONFIG_T::n_scale_bias]) {
     #pragma HLS ARRAY_PARTITION variable=threshold_hi complete
     #pragma HLS ARRAY_PARTITION variable=threshold_lo complete
 
@@ -85,7 +85,7 @@ void normalize_ternary_tanh(hls::stream<data_T> &data, hls::stream<nnet::array<a
         #pragma HLS PIPELINE
 
         data_T in_data = data.read();
-        nnet::array<ap_int<2>, CONFIG_T::n_in> out_data;
+        nnet::array<ap_int<2>, CONFIG_T::n_scale_bias> out_data;
         PRAGMA_DATA_PACK(out_data)
 
     BatchNormPack:

test/pytest/test_binary_cnn.py

@@ -90,11 +90,4 @@ def test_model2(backend, io_type):
     y = model2.predict(np.zeros((1, 28, 28, 1)))
     y_hls = hls_model.predict(np.zeros((1, 28, 28, 1)))
 
-    print(f"{y_hls=}")
-    print(f"{y=}")
-
     np.testing.assert_allclose(np.squeeze(y_hls), np.squeeze(y), rtol=1e-2, atol=0.01)
-
-
-if __name__ == "__main__":
-    test_model2("Vivado", "io_stream")