diff options
| -rw-r--r-- | ethosu/vela/tflite_graph_optimiser.py | 18 | ||||
| -rw-r--r-- | ethosu/vela/weight_compressor.py | 3 |
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] |