Merge branch 'dev' into feature/weight_streamers

run-docker.sh

@@ -86,7 +86,7 @@ VIVADO_IP_CACHE=$BUILD_LOCAL/vivado_ip_cache
 git clone --branch feature/finn_onnx_export $BREVITAS_REPO $BREVITAS_LOCAL ||  git -C "$BREVITAS_LOCAL" pull
 git clone $EXAMPLES_REPO $EXAMPLES_LOCAL ||  git -C "$EXAMPLES_LOCAL" pull
 git clone $CNPY_REPO $CNPY_LOCAL ||  git -C "$CNPY_LOCAL" pull
-git clone $FINN_HLS_REPO $FINN_HLS_LOCAL; git -C "$FINN_HLS_LOCAL" checkout b5dc957a16017b8356a7010144b0a4e2f8cfd124 ||  git -C "$FINN_HLS_LOCAL" checkout b5dc957a16017b8356a7010144b0a4e2f8cfd124
+git clone $FINN_HLS_REPO $FINN_HLS_LOCAL ||  git -C "$FINN_HLS_LOCAL" checkout master; git -C "$FINN_HLS_LOCAL" pull
 git clone $PYVERILATOR_REPO $PYVERILATOR_LOCAL ||  git -C "$PYVERILATOR_LOCAL" pull
 git clone $PYNQSHELL_REPO $PYNQSHELL_LOCAL ||  git -C "$PYNQSHELL_LOCAL" pull
 

src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py

@@ -280,13 +280,13 @@ class StreamingFCLayer_Batch(HLSCustomOp):
         out_hls_str = self.get_output_datatype().get_hls_datatype_str()
         wt_hls_str = self.get_weight_datatype().get_hls_datatype_str()
         inp_is_binary = self.get_input_datatype() == DataType.BINARY
-        out_is_binary = self.get_output_datatype() == DataType.BINARY
+        # out_is_binary = self.get_output_datatype() == DataType.BINARY
         wt_is_binary = self.get_weight_datatype() == DataType.BINARY
         bin_xnor_mode = self.get_nodeattr("binaryXnorMode") == 1
         if (inp_is_binary or wt_is_binary) and (not bin_xnor_mode):
             raise Exception("True binary (non-bipolar) inputs not yet supported")
         inp_is_bipolar = self.get_input_datatype() == DataType.BIPOLAR
-        out_is_bipolar = self.get_output_datatype() == DataType.BIPOLAR
+        # out_is_bipolar = self.get_output_datatype() == DataType.BIPOLAR
         wt_is_bipolar = self.get_weight_datatype() == DataType.BIPOLAR
         # reinterpret inp/wt as bipolar if bin_xnor_mode is iset
         inp_is_bipolar = inp_is_bipolar or (inp_is_binary and bin_xnor_mode)
@@ -324,10 +324,7 @@ class StreamingFCLayer_Batch(HLSCustomOp):
                 ret["TWeightI"] = "Slice<%s>" % wt_hls_str
 
         # fill in TDstI
-        if out_is_bipolar or out_is_binary:
-            ret["TDstI"] = "Identity"
-        else:
-            ret["TDstI"] = "Slice<%s>" % out_hls_str
+        ret["TDstI"] = "Slice<%s>" % out_hls_str
         return ret
 
     def get_hls_compatible_weight_tensor(self, orig_weight_matrix):

src/finn/transformation/batchnorm_to_affine.py

@@ -67,6 +67,16 @@ class BatchNormToAffine(Transformation):
                     if consumer.op_type == "Squeeze":
                         bn_output = consumer.output[0]
                 data_shape = model.get_tensor_shape(bn_input)
+                assert A.ndim == B.ndim, "Unexpected mul/add dims in BatchNormToAffine"
+                assert (
+                    len(data_shape) >= A.ndim
+                ), "Unexpected number of dims found in BatchNormToAffine"
+                # reshape the mul/add constants to match the data shape/dims
+                # by adding (1,) dimensions to the right
+                n_spatial_dims = len(data_shape) - 2
+                target_shape = (1, -1) + tuple(1 for i in range(n_spatial_dims))
+                A = A.reshape(target_shape)
+                B = B.reshape(target_shape)
                 # create value_info and initializers for Mul and Add constants
                 mul_const = oh.make_tensor_value_info(
                     model.make_new_valueinfo_name(), TensorProto.FLOAT, A.shape

tests/transformation/test_batchnorm_to_affine.py

@@ -28,10 +28,12 @@
 
 import os
 from pkgutil import get_data
+import pkg_resources as pk
 
 import brevitas.onnx as bo
 import onnx
 import onnx.numpy_helper as nph
+import numpy as np
 
 import finn.core.onnx_exec as oxe
 from finn.core.modelwrapper import ModelWrapper
@@ -58,18 +60,23 @@ def test_batchnorm_to_affine_lfc_w1a1():
     os.remove(export_onnx_path)
 
 
-# cnv batchnorm to affine not yet supported
-
-# def test_batchnorm_to_affine_cnv_w1a1():
-#    lfc = get_test_model_trained("CNV", 1, 1)
-#    bo.export_finn_onnx(lfc, (1, 3, 32, 32), export_onnx_path)
-#    model = ModelWrapper(export_onnx_path)
-#    model = model.transform(InferShapes())
-#    model = model.transform(FoldConstants())
-#    # TODO shape inference failing on transformed model below -- needs debug
-#    new_model = model.transform(BatchNormToAffine())
-#    # check that there are no BN nodes left
-#    # TODO replace this with execution test
-#    op_types = list(map(lambda x: x.op_type, new_model.graph.node))
-#    assert "BatchNormalization" not in op_types
-#    os.remove(export_onnx_path)
+def test_batchnorm_to_affine_cnv_w1a1():
+    lfc = get_test_model_trained("CNV", 1, 1)
+    bo.export_finn_onnx(lfc, (1, 3, 32, 32), export_onnx_path)
+    model = ModelWrapper(export_onnx_path)
+    model = model.transform(InferShapes())
+    model = model.transform(FoldConstants())
+    fn = pk.resource_filename("finn", "data/cifar10/cifar10-test-data-class3.npz")
+    input_tensor = np.load(fn)["arr_0"].astype(np.float32)
+    assert input_tensor.shape == (1, 3, 32, 32)
+    input_dict = {"0": input_tensor}
+    output_dict = oxe.execute_onnx(model, input_dict)
+    expected = output_dict[list(output_dict.keys())[0]]
+    new_model = model.transform(BatchNormToAffine())
+    # check that there are no BN nodes left
+    op_types = list(map(lambda x: x.op_type, new_model.graph.node))
+    assert "BatchNormalization" not in op_types
+    output_dict_p = oxe.execute_onnx(new_model, input_dict)
+    produced = output_dict_p[list(output_dict_p.keys())[0]]
+    assert np.isclose(expected, produced).all()
+    os.remove(export_onnx_path)