NXP backend: added support for aten.conv_transpose1 and refactored convolution_converter

backends/nxp/aten_passes/neutron_aten_pass_manager.py

@@ -7,6 +7,9 @@
 
 import torch
 
+from executorch.backends.nxp.aten_passes.convert_1d_conv_to_2d import (
+    ConvertConv1dToConv2dPass,
+)
 from executorch.backends.nxp.aten_passes.convert_div_to_mul import ConvertDivToMulPass
 from executorch.backends.nxp.aten_passes.decompose_split_to_slices_pass import (
     DecomposeSplitToSlicesPass,
@@ -49,6 +52,7 @@ def _get_default_passes(neutron_target_spec, qat_mode: bool = False) -> list[Pas
         FuseLinearAndAddPass(),
         MoveActivationBeforeConcat(neutron_target_spec),
         ConvertDivToMulPass(),
+        ConvertConv1dToConv2dPass(),
     ]
 
     if not qat_mode:

backends/nxp/backend/ir/converter/node_converters/ops_converters/convolution_converter.py

@@ -15,7 +15,6 @@
 from executorch.backends.nxp.backend.ir.converter.conversion import (
     aten_translator,
     common,
-    translator,
 )
 from executorch.backends.nxp.backend.ir.converter.conversion.common import try_get_input
 from executorch.backends.nxp.backend.ir.converter.conversion.translator import (
@@ -42,7 +41,6 @@
 from executorch.backends.nxp.backend.ir.tflite_generator.builtin_options import (
     conv_2d_options,
     depthwise_conv_2d_options,
-    reshape_options,
     transpose_conv_options,
 )
 from executorch.backends.nxp.backend.neutron_target_spec import NeutronTargetSpec
@@ -70,8 +68,9 @@ def _is_supported_on_target(
             return False
 
         if conv_params.transposed:
-            # TransposeConv1d is not supported on Neutron
-            if len(conv_params.dilation) == 1:
+            # TransposeConv2d with groups > 1 is not supported
+            # TODO: split into multiple convs with groups = 1
+            if conv_params.groups > 1:
                 return False
             if not node_is_effectively_static_tensor(weights, parameters_mapping):
                 # Only supported if the weights are static, because TFLite `TransposeConv` uses permuted
@@ -187,99 +186,6 @@ def _get_convolution_arguments(
             groups,
         )
 
-    def _convert_1d_conv(
-        self, t_op: tflite_model.Operator, conv_params: ConvParameters
-    ) -> list[tflite_model.Operator]:
-        """Convert the 'Conv' operator with a 1D kernel to TFLite 'Conv2D'.
-        TFLite doesn't support 1D convolution, but this behaviour can be represented using
-               Reshape -> Conv2D -> Reshape.
-        The first reshape introduces a 4th dimension with size 1. The second Reshape removes the temporary dimension.
-        """
-        # -- Calculate the shapes for equivalent 2D convolution --
-        conv_2d_input_shape = translator.nhc_dimensions_to_nhwc(
-            t_op.tmp_inputs[0].shape.vector
-        )
-        conv_2d_weight_shape = translator.nhc_dimensions_to_nhwc(
-            t_op.tmp_inputs[1].shape.vector
-        )
-        conv_2d_output_shape = translator.nhc_dimensions_to_nhwc(
-            t_op.tmp_outputs[0].shape.vector
-        )
-
-        # -- Generate tensors taking part in the conversion --
-        reshape1_input = t_op.tmp_inputs[0]
-
-        reshape1_output = self.builder.duplicate_tensor(
-            reshape1_input, name_suffix="_4D_"
-        )
-        reshape1_output.shape = tflite_model.Shape(conv_2d_input_shape)
-
-        reshape2_input = self.builder.duplicate_tensor(
-            t_op.tmp_outputs[0], name_suffix="_4D_"
-        )
-        reshape2_input.shape = tflite_model.Shape(conv_2d_output_shape)
-
-        reshape2_output = t_op.tmp_outputs[0]
-
-        pre_reshapes = []
-
-        # Extend the weights tensor to 4D
-        weights_tensor = t_op.tmp_inputs[1]
-        if tensor_has_data(weights_tensor):
-            # Do it statically
-            weights_tensor.shape = tflite_model.Shape(conv_2d_weight_shape)
-            weights_tensor.tmp_buffer.data = weights_tensor.tmp_buffer.data.reshape(
-                conv_2d_weight_shape
-            )
-
-        else:
-            # Add a Reshape before the weights tensor
-            new_weights_tensor = self.builder.duplicate_tensor(
-                weights_tensor, name_suffix="_4D_"
-            )
-            new_weights_tensor.shape = tflite_model.Shape(conv_2d_weight_shape)
-
-            weight_reshape = tflite_model.Operator(
-                builtin_options=reshape_options.Reshape(conv_2d_weight_shape)
-            )
-            weight_reshape.tmp_inputs = [weights_tensor]
-            weight_reshape.tmp_outputs = [new_weights_tensor]
-
-            pre_reshapes.append(weight_reshape)
-
-            # Save the new weights tensor, to assign it later.
-            weights_tensor = new_weights_tensor
-
-        # -- Create the new operators --
-        reshape1 = tflite_model.Operator(
-            builtin_options=reshape_options.Reshape(conv_2d_input_shape)
-        )
-        reshape1.tmp_inputs = [reshape1_input]
-        reshape1.tmp_outputs = [reshape1_output]
-        pre_reshapes.append(reshape1)
-
-        reshape2 = tflite_model.Operator(
-            builtin_options=reshape_options.Reshape(reshape2_output.shape.vector)
-        )
-        reshape2.tmp_inputs = [reshape2_input]
-        reshape2.tmp_outputs = [reshape2_output]
-
-        # Assign the new input and output of the Conv2D
-        t_op.tmp_inputs = [reshape1_output, weights_tensor] + t_op.tmp_inputs[
-            2:
-        ]  # Add bias as well, if present
-        t_op.tmp_outputs = [reshape2_input]
-
-        # Extend all Conv attributes to 2D
-        common.extend_1d_stride_to_2d(conv_params.stride)
-        common.extend_1d_dilation_to_2d(conv_params.dilation)
-        common.extend_1d_padding_to_2d(conv_params.padding)
-
-        # Convert the now 2D Conv
-        converted_conv_ops = self._convert_2d_conv(t_op, conv_params)
-
-        return pre_reshapes + converted_conv_ops + [reshape2]
-
     # noinspection PyPep8Naming
     def _convert_unpadded_2D(
         self, t_op: tflite_model.Operator, conv_params: ConvParameters
@@ -523,9 +429,7 @@ def convert(self, node: Node):
         )
 
         rank = t_op.tmp_inputs[1].shape.len()
-        if rank == 3:  # Conv1D
-            ops_to_add = self._convert_1d_conv(t_op, conv_params)
-        elif rank == 4:  # Conv2D
+        if rank == 4:  # Conv2D
             ops_to_add = self._convert_2d_conv(t_op, conv_params)
         else:
             raise NotImplementedError(

backends/nxp/quantizer/neutron_quantizer.py

@@ -23,7 +23,6 @@
     BMMPattern,
     CatPattern,
     ClampPattern,
-    Conv1dPattern,
     Conv2dPattern,
     ConvTranspose2dPattern,
     DropoutPattern,
@@ -266,9 +265,10 @@ def __init__(self, neutron_target_spec: NeutronTargetSpec, is_qat: bool = False)
                 OpQuantizer(BMMPattern(is_qat=is_qat), static_qconfig),
                 OpQuantizer(CatPattern(is_qat=is_qat), static_qconfig),
                 OpQuantizer(ClampPattern(is_qat=is_qat), static_qconfig),
-                OpQuantizer(Conv1dPattern(is_qat=is_qat), static_qconfig),
                 OpQuantizer(Conv2dPattern(self, is_qat=is_qat), static_qconfig),
-                OpQuantizer(ConvTranspose2dPattern(is_qat=is_qat), static_qconfig),
+                OpQuantizer(
+                    ConvTranspose2dPattern(self, is_qat=is_qat), static_qconfig
+                ),
                 OpQuantizer(DropoutPattern(is_qat=is_qat), static_qconfig),
                 OpQuantizer(FlattenPattern(is_qat=is_qat), static_qconfig),
                 OpQuantizer(HardTanhPattern(is_qat=is_qat), static_qconfig),

backends/nxp/quantizer/patterns.py

@@ -7,10 +7,14 @@
 
 from abc import ABC, abstractmethod
 from dataclasses import dataclass, field
+from functools import partial
 
 import torch
 
-from executorch.backends.nxp.quantizer.utils import get_bias_qparams
+from executorch.backends.nxp.quantizer.utils import (
+    get_bias_qparams,
+    get_padded_bias_qparams,
+)
 from torch import fx
 from torch._ops import OpOverload
 from torch.fx import Node
@@ -482,16 +486,6 @@ def get_anchors(
         )
 
 
-class Conv1dPattern(ConvPattern):
-    def partition_types(self) -> list[OpOverload]:
-        return [torch.ops.aten.conv1d.default]
-
-
-class ConvTranspose1dPattern(ConvPattern):
-    def partition_types(self) -> list[OpOverload]:
-        return [torch.ops.aten.conv_transpose1d.default]
-
-
 class Conv2dPattern(ConvPattern):
     def __init__(self, neutron_quantizer, is_qat: bool = False):
         super().__init__(is_qat=is_qat)
@@ -572,6 +566,14 @@ def get_anchors(
 
 
 class ConvTranspose2dPattern(QuantizationPattern):
+    def __init__(self, neutron_quantizer, is_qat: bool = False):
+        super().__init__(is_qat=is_qat)
+
+        self.neutron_quantizer = neutron_quantizer
+        self.neutron_target_info = (
+            self.neutron_quantizer.neutron_target_spec.neutron_target_info
+        )
+
     def partition_types(self) -> list[OpOverload]:
         return [torch.ops.aten.conv_transpose2d.input]
 
@@ -580,48 +582,81 @@ def get_anchors(
     ) -> PartitionAnchors:
         conv_node = fused_partition[0].nodes[-1]
 
+        # When `groups` > 1, the per-channel weight qparams have shape (`out_channels` / `groups`),
+        # but bias qparams have shape (`out_channels`) - not divided by `groups`.
+        # So the weight qparams must be expanded to match the shape correctly.
+        groups = 1 if len(conv_node.args) < 7 else conv_node.args[6]
+        if groups > 1:
+            out_channels = conv_node.meta["val"].shape[1]
+            derive_qparams_fn = partial(
+                get_padded_bias_qparams, out_channels=out_channels
+            )
+
+        else:
+            derive_qparams_fn = get_bias_qparams
+
         bias_quantization_qspec = DerivedQuantizationSpec(
             derived_from=[
                 (conv_node.args[0], conv_node),
                 (conv_node.args[1], conv_node),
             ],
-            derive_qparams_fn=get_bias_qparams,
+            derive_qparams_fn=derive_qparams_fn,
             dtype=torch.int32,
             quant_min=-(2**31) + 1,
             quant_max=2**31 - 1,
             qscheme=torch.per_channel_symmetric,
             ch_axis=0,
         )
 
-        weight_observer_or_fake_quant_ctr = PerChannelMinMaxObserver
+        w_ch_axis = 1
+        weight_observer_or_fake_quant_ctr = (
+            FakeQuantize.with_args(
+                observer=MovingAveragePerChannelMinMaxObserver, ch_axis=w_ch_axis
+            )
+            if self.is_qat
+            else PerChannelMinMaxObserver.with_args(ch_axis=w_ch_axis)
+        )
         weight_quantization_spec = QuantizationSpec(
             dtype=torch.int8,
             observer_or_fake_quant_ctr=weight_observer_or_fake_quant_ctr,
             quant_min=-127,
             quant_max=127,
             qscheme=torch.per_channel_symmetric,
-            ch_axis=1,
+            ch_axis=w_ch_axis,
         )
 
         # Keep bias empty if not supplied
         bias = []
         if len(conv_node.args) > 2 and conv_node.args[2] is not None:
             bias = [(conv_node, NodeArgsIdx(2), bias_quantization_qspec)]
 
-        output_specs = [(conv_node,)]
+        # If the following node is a fusable activation, quantize together with activation
+        output = [(conv_node,)]
+        if len(
+            conv_node.users
+        ) == 1 and self.neutron_target_info.is_supported_fused_activation__aten(
+            activation := next(iter(conv_node.users))
+        ):
+            activation_quantizer = self.neutron_quantizer.op_to_quantizer[
+                activation.target
+            ]
+            activation_quantizer.annotate(gm)
+            output = []
+            activation.meta["quantization_annotation"].input_qspec_map = {}
+
         # In order for QAT to be numerically correct, there should be no quantization between
         # convolution node and batch norm node.
         if self.is_qat:
             conv_users = conv_node.users
             possibly_bn = list(conv_users.keys())[0] if len(conv_users) == 1 else None
             if possibly_bn and _is_batch_norm(possibly_bn):
-                output_specs = []
+                output = []
 
         return PartitionAnchors(
             inputs=[(conv_node, NodeArgsIdx(0))],
             weights=[(conv_node, NodeArgsIdx(1), weight_quantization_spec)],
             biases=bias,
-            output=output_specs,
+            output=output,
         )
 
 

backends/nxp/quantizer/utils.py

@@ -73,6 +73,32 @@ def get_bias_qparams(
     return bias_scale, bias_zero_point
 
 
+def get_padded_bias_qparams(
+    obs_or_fqs: List[ObserverOrFakeQuantize],
+    out_channels: int | None = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    act_scale, _ = obs_or_fqs[0].calculate_qparams()
+    weight_scale, _ = obs_or_fqs[1].calculate_qparams()
+
+    # It may happen that `torch.ao` incorrectly sets the weight qparams, not matching bias qparams.
+    # If `out_channels` is given, ensure bias qparams are per-output-channel:
+    # So for example  w = [w1, w2, w3]  ->  [w1, w2, w3, w1, w2, w3, ...]
+    if out_channels is not None:
+        weight_scale = weight_scale.flatten()
+        if weight_scale.numel() != out_channels:
+            if out_channels % weight_scale.numel() != 0:
+                raise RuntimeError(
+                    "Weight qparams cannot be repeated if not divisible by `out_channels`."
+                )
+            weight_scale = weight_scale.repeat(out_channels // weight_scale.numel())
+
+        act_scale = act_scale.flatten()[0]
+
+    bias_scale = act_scale * weight_scale
+    bias_zero_point = torch.zeros_like(bias_scale, dtype=torch.int64)
+    return bias_scale, bias_zero_point
+
+
 def get_aten_node_target_partitions(
     graph: torch.fx.Graph,
     wanted_original_aten_op: List[OpOverload],