MLBEDSW-8497: [MLCE] Avoid modifying FC with dynamic weights

 - If a npu op is followed by a convolution op with dynamic
weights the optimized file ends up containing a duplicated
tensor called _cpu.
 - Another problem is also that an empty bias tensor is added
in the reader.
 - The fix is to ignore these cpu ops both in the reader
and the writer.

Change-Id: I476b4f6062e26cca4ba589df694a99ef79b0f6d4
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

2 files changed, 15 insertions, 10 deletions

diff --git a/ethosu/vela/tflite_reader.py b/ethosu/vela/tflite_reader.py
index 85acb6b..e732f19 100644
--- a/ethosu/vela/tflite_reader.py
+++ b/ethosu/vela/tflite_reader.py
@@ -153,18 +153,21 @@ class TFLiteSubgraph:
             self.virtual_outputs.append(tens)
 
         if op.type.is_depthwise_conv2d_op() or op.type.is_conv2d_op() or op.type == Op.FullyConnected:
+            # Reshape and add bias for ops with constant weights
+            # Do not modify ops with dynamic data since they will run on CPU
             if inputs[1].values is not None:
                 if op.type == Op.FullyConnected:
                     inputs[1] = clone_and_reshape_tensor(inputs[1], (1, 0), False)
                 else:
                     inputs[1] = clone_and_reshape_tensor(inputs[1], (1, 2, 3, 0), False)
-            if op.type.needs_bias() and len(inputs) <= op_type.info.indices.biases[0]:
-                # No Bias tensor
-                inputs.append(None)
-            if inputs[-1] and inputs[-1].values is not None:
-                # Since bias tensor is used for both bias and scale,
-                # a clone with a unique equivalence_id is needed.
-                inputs[-1] = clone_and_reshape_tensor(inputs[-1], None, True)
+
+                if op.type.needs_bias() and len(inputs) <= op_type.info.indices.biases[0]:
+                    # No Bias tensor
+                    inputs.append(None)
+                if inputs[-1] and inputs[-1].values is not None:
+                    # Since bias tensor is used for both bias and scale,
+                    # a clone with a unique equivalence_id is needed.
+                    inputs[-1] = clone_and_reshape_tensor(inputs[-1], None, True)
 
         if opt_serializer is not None:
             op.attrs = opt_serializer.deserialize(op_data)
diff --git a/ethosu/vela/tflite_writer.py b/ethosu/vela/tflite_writer.py
index 44ce711..d4e24a2 100644
--- a/ethosu/vela/tflite_writer.py
+++ b/ethosu/vela/tflite_writer.py
@@ -105,9 +105,11 @@ class TFLiteSerialiser:
                     if op.type.is_conv2d_op() or op.type.is_depthwise_conv2d_op() or op.type == Op.FullyConnected:
                         # Op is run on CPU, make sure the original weight and bias tensors are written back
                         # instead of the cloned/reshaped (see tflite_reader)
-                        for idx, inp in enumerate(op.inputs):
-                            if inp != op.ifm and inp is not None and inp.src_tensor is not None:
-                                op.inputs[idx] = inp.src_tensor
+                        # Do nothing when values are None (dynamic weights)
+                        if op.inputs[1].values is not None:
+                            for idx, inp in enumerate(op.inputs):
+                                if inp != op.ifm and inp is not None and inp.src_tensor is not None:
+                                    op.inputs[idx] = inp.src_tensor
 
         # list of tuple(Op, string, op.version); the custom code is only used for 3rd party custom operators
         self.operator_codes = sorted(set((op.type, op.attrs.get("custom_code", ""), op.version) for op in all_ops))