MLBEDSW-4338 Randomized int16 PAD output diff

The issue was that the AveragePool in these test cases were
translated to DepthwiseConv2DBias and int16 convolutions
always runs with reduced scale. Fixed so that reduced scale
is not used in this case.

Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>
Change-Id: Ice956eabbb37c8aa1991464870006971c6ecec43

2 files changed, 16 insertions, 5 deletions

diff --git a/ethosu/vela/tflite_graph_optimiser.py b/ethosu/vela/tflite_graph_optimiser.py
index 0630ef41..6b454e3d 100644
--- a/ethosu/vela/tflite_graph_optimiser.py
+++ b/ethosu/vela/tflite_graph_optimiser.py
@@ -385,7 +385,7 @@ def convert_resizenn_ac_to_depthwise_conv(op, upscale_factor):
     # need to append the bias tensor as resize ops only have 2 inputs
     assert len(op.inputs) == 2
     op.inputs.append(None)
-    fixup_bias_tensors(op, None, None)
+    fixup_bias_tensors(op, None, None, DataType.int32)
 
     # finally update the shape incase we've change the tensor shapes or connections
     op.set_ifm_ofm_shapes()
@@ -1324,7 +1324,7 @@ def replace_pad_by_hw_pad(op: Operation, arch, nng):
             op.add_input_tensor(weight_tens)
             # Add bias tensor, all biases set to 0
             op.inputs.append(None)
-            fixup_bias_tensors(op, arch, nng)
+            fixup_bias_tensors(op, arch, nng, DataType.int32)
             # Add other inputs
             op.inputs.extend(other_inputs)
             op.rounding_mode = NpuRoundingMode.NATURAL
@@ -1407,12 +1407,22 @@ def convert_pad(op: Operation, arch, nng):
     return avgpool_op
 
 
-def fixup_bias_tensors(op, arch, nng):
+def fixup_bias_tensors(op, arch, nng, dtype=None):
     if op.type.needs_bias() and op.bias is None:
         # Op has no bias, add bias tensor filled with zeros
         nr_biases = op.inputs[1].shape[-1]
         bias_values = [0] * nr_biases
-        bias_tensor = create_const_tensor(op.name + "_bias", [nr_biases], DataType.int32, bias_values)
+        # The DataType of the bias tensor can be explicitly provided or deduced from the ifm
+        # DataType. Default is int32 bias for 8-bit ifms and int64 for int16 ifms.
+        # For int16 the selected bias DataType will have an impact on the scaling
+        # used when encoding the scales and biases later. The default mode will match the
+        # refence with reduced scaling for int64 bias.
+        # This means that in cases (in the graph optimiser) where DepthwiseConv2DBias
+        # is used to emulate average pool int32 bias should be selected for full precision
+        # int16 scaling.
+        if dtype is None:
+            dtype = DataType.int64 if op.ifm.dtype == DataType.int16 else DataType.int32
+        bias_tensor = create_const_tensor(op.name + "_bias", [nr_biases], dtype, bias_values)
         op.set_input_tensor(bias_tensor, op.type.info.indices.biases[0])
 
     return op
diff --git a/ethosu/vela/weight_compressor.py b/ethosu/vela/weight_compressor.py
index 78c43511..db225fb6 100644
--- a/ethosu/vela/weight_compressor.py
+++ b/ethosu/vela/weight_compressor.py
@@ -275,7 +275,8 @@ def _prepare_scale_and_bias(arch, tens, rescale_for_faf, explicit_scaling):
         quantised_scales = [(int(m), int(s)) for s, m in zip(explicit_scaling.shift, explicit_scaling.multiplier)]
     else:
         # quantise all of the weight scales into (scale_factor, shift)
-        if ifm_dtype == DataType.int16:
+        if ifm_dtype == DataType.int16 and bias_tens.dtype == DataType.int64:
+            # Reference uses reduced scaling for int16 with int64 bias
             quantised_scales = [reduced_quantise_scale(scale) for scale in scales]
         else:
             quantised_scales = [quantise_scale(scale) for scale in scales]