diff --git a/src/finn/transformation/streamline/reorder.py b/src/finn/transformation/streamline/reorder.py
index b46b82c77a3f1b70a3b05d87cd3c48fc1d94fd45..a1bd16f6d0b70193122d5d067ccdee395260c7b1 100644
--- a/src/finn/transformation/streamline/reorder.py
+++ b/src/finn/transformation/streamline/reorder.py
@@ -32,6 +32,7 @@ from onnx import helper as oh
from finn.transformation import Transformation
from finn.transformation.infer_shapes import InferShapes
+from finn.core.datatype import DataType
from finn.core.onnx_exec import execute_node
from finn.util.basic import get_by_name
from finn.custom_op.registry import getCustomOp
@@ -338,6 +339,71 @@ class MoveScalarMulPastConv(Transformation):
return (model, graph_modified)
+class MoveMulPastDWConv(Transformation):
+ """Move channelwise mul operations past depthwise conv operations. We want to have muls
+ next to each other such that they can be collapsed into a single mul."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ node_ind += 1
+ if (
+ n.op_type == "Mul"
+ and not model.is_fork_node(n)
+ and not model.is_join_node(n)
+ ):
+ consumer = model.find_consumer(n.output[0])
+ if (
+ consumer is not None
+ and consumer.op_type == "Conv"
+ and not model.is_join_node(consumer)
+ ):
+ mul_weight_name = n.input[1]
+ A = model.get_initializer(mul_weight_name)
+ if A is None:
+ warnings.warn(
+ """Mul weight tensor is not set. If it is a constant,
+ please use set_initializer to set the tensor."""
+ )
+ continue
+ conv_node = consumer
+ mul_node = n
+ start_name = mul_node.input[0]
+ conv_in_name = conv_node.input[0]
+ conv_in_shape = model.get_tensor_shape(conv_in_name)
+ ifm_ch = conv_in_shape[1]
+ group_attribute = get_by_name(consumer.attribute, "group")
+ if group_attribute is None:
+ continue
+ group_attribute = group_attribute.i
+ conv_out_name = conv_node.output[0]
+ conv_out_shape = model.get_tensor_shape(conv_out_name)
+ if A.shape == (1, ifm_ch, 1, 1) and ifm_ch == group_attribute:
+ # if the mul is channelwise and conv is depthwise,
+ # we can simply swap the order of ops
+ # rewire mul input to be conv input
+ conv_node.input[0] = start_name
+ model.set_tensor_shape(start_name, conv_in_shape)
+ model.set_tensor_datatype(start_name, DataType.FLOAT32)
+ # use old conv input tensor as conv output
+ conv_node.output[0] = conv_in_name
+ model.set_tensor_shape(conv_in_name, conv_out_shape)
+ model.set_tensor_datatype(conv_in_name, DataType.FLOAT32)
+ # use new conv output as new mul node input
+ mul_node.input[0] = conv_in_name
+ # use old conv output as new mul node output
+ mul_node.output[0] = conv_out_name
+ model.set_tensor_datatype(conv_out_name, DataType.FLOAT32)
+ # move mul node past conv node
+ graph.node.remove(mul_node)
+ graph.node.insert(node_ind, mul_node)
+ graph_modified = True
+ model = model.transform(InferShapes())
+ return (model, graph_modified)
+
+
class MoveLinearPastEltwiseAdd(Transformation):
"""Move linear operations (mul, add) past elementwise add operations where possible.
Specifically,matches and transforms the following patterns:
diff --git a/tests/transformation/test_move_mul_past_dw_conv.py b/tests/transformation/test_move_mul_past_dw_conv.py
new file mode 100644
index 0000000000000000000000000000000000000000..1ae8fbfe89986d58d3d71f5f8735a98469d9d1e3
--- /dev/null
+++ b/tests/transformation/test_move_mul_past_dw_conv.py
@@ -0,0 +1,93 @@
+import pytest
+
+from onnx import helper, TensorProto
+from finn.custom_op.im2col import compute_conv_output_dim
+import finn.core.onnx_exec as oxe
+from finn.core.datatype import DataType
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.infer_datatypes import InferDataTypes
+from finn.transformation.infer_shapes import InferShapes
+from finn.util.basic import gen_finn_dt_tensor
+from finn.transformation.streamline.reorder import MoveMulPastDWConv
+
+
+# input dimension
+@pytest.mark.parametrize("ifm_dim", [4, 7])
+# input channels
+@pytest.mark.parametrize("ifm_ch", [2, 3])
+# kernel size
+@pytest.mark.parametrize("k", [2, 3])
+# stride
+@pytest.mark.parametrize("stride", [1, 2])
+# padding
+@pytest.mark.parametrize("pad_amt", [0, 1])
+# depthwise
+@pytest.mark.parametrize("dw", [0, 1])
+def test_move_mul_past_dw_conv(ifm_dim, ifm_ch, k, stride, pad_amt, dw):
+ if dw == 1:
+ ofm_ch = ifm_ch
+ groups = ifm_ch
+ W_shape = [ofm_ch, 1, k, k]
+ else:
+ ofm_ch = ifm_ch + 2
+ groups = 1
+ W_shape = [ofm_ch, ifm_ch, k, k]
+
+ ofm_dim = compute_conv_output_dim(ifm_dim, k, stride, pad_amt)
+
+ # set up onnx model
+ inp = helper.make_tensor_value_info(
+ "inp", TensorProto.FLOAT, [1, ifm_ch, ifm_dim, ifm_dim]
+ )
+ mul = helper.make_tensor_value_info("mul", TensorProto.FLOAT, [1, ifm_ch, 1, 1])
+ W = helper.make_tensor_value_info("W", TensorProto.FLOAT, W_shape)
+ outp = helper.make_tensor_value_info(
+ "outp", TensorProto.FLOAT, [1, ofm_ch, ofm_dim, ofm_dim]
+ )
+
+ Mul_node = helper.make_node("Mul", ["inp", "mul"], ["mul_out"])
+
+ Conv_node = helper.make_node(
+ "Conv",
+ ["mul_out", "W"],
+ ["outp"],
+ group=groups,
+ kernel_shape=[k, k],
+ pads=[pad_amt, pad_amt, pad_amt, pad_amt],
+ strides=[stride, stride],
+ )
+
+ graph = helper.make_graph(
+ nodes=[Mul_node, Conv_node],
+ name="mulpastconv_graph",
+ inputs=[inp],
+ outputs=[outp],
+ value_info=[mul, W],
+ )
+
+ model = helper.make_model(graph, producer_name="mulpastconv-model")
+ model = ModelWrapper(model)
+ inp_values = gen_finn_dt_tensor(DataType.INT2, [1, ifm_ch, ifm_dim, ifm_dim])
+ mul_values = gen_finn_dt_tensor(DataType.INT2, [1, ifm_ch, 1, 1])
+ W_values = gen_finn_dt_tensor(DataType.INT2, W_shape)
+ model.set_initializer("W", W_values)
+ model.set_initializer("mul", mul_values)
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ idict = {"inp": inp_values}
+ odict = oxe.execute_onnx(model, idict, True)
+ out_before = odict["outp"]
+
+ # move channelwise multiplication past depthwise conv
+ model_transformed = model.transform(MoveMulPastDWConv())
+ odict = oxe.execute_onnx(model_transformed, idict, True)
+ out_after = odict["outp"]
+
+ assert (out_before == out_after).all()
+
+ if dw == 0:
+ assert model.graph.node[0].op_type == model_transformed.graph.node[0].op_type
+ assert model.graph.node[1].op_type == model_transformed.graph.node[1].op_type
+ else:
+ assert model.graph.node[0].op_type == model_transformed.graph.node[1].op_type
+ assert model.graph.node[1].op_type == model_transformed.graph.node[0].op_type