Summary: MobileNetV2 Fully Lowered to CMSIS-NN

Cortex-M: Enable full MobileNetV2 lowering to CMSIS-NN backend

This PR enables end-to-end export of MobileNetV2 to the CMSIS-NN backend for Cortex-M
targets. All quantized operations (conv2d, depthwise conv2d, linear/addmm, activations)
are now properly lowered to cortex_m::quantized_* operators, enabling efficient inference
on resource-constrained microcontrollers

Test Plan:
python3 -m examples.arm.aot_arm_compiler -m mv2 --target=cortex-m --quantize --intermediates=./mv2_intermediates --output=./mv2_cortex_m.pte

cat ./mv2_intermediates/delegation_info.txt
Delegation info:
Total delegated subgraphs: 0
Number of delegated nodes: 0
Number of non-delegated nodes: 72

Delegation table:
╒════╤═════════════════════════════════════════════╤═══════════════════════════════════╤═══════════════════════════════════════╕
│    │ op_type                                     │   occurrences_in_delegated_graphs │   occurrences_in_non_delegated_graphs │
╞════╪═════════════════════════════════════════════╪═══════════════════════════════════╪═══════════════════════════════════════╡
│  0 │ aten_as_strided_copy_default                │                                 0 │                                     1 │
├────┼─────────────────────────────────────────────┼───────────────────────────────────┼───────────────────────────────────────┤
│  1 │ aten_mean_dim                               │                                 0 │                                     1 │
├────┼─────────────────────────────────────────────┼───────────────────────────────────┼───────────────────────────────────────┤
│  2 │ aten_view_copy_default                      │                                 0 │                                     1 │
├────┼─────────────────────────────────────────────┼───────────────────────────────────┼───────────────────────────────────────┤
│  3 │ cortex_m_dequantize_per_tensor_default      │                                 0 │                                     2 │
├────┼─────────────────────────────────────────────┼───────────────────────────────────┼───────────────────────────────────────┤
│  4 │ cortex_m_quantize_per_tensor_default        │                                 0 │                                     2 │
├────┼─────────────────────────────────────────────┼───────────────────────────────────┼───────────────────────────────────────┤
│  5 │ cortex_m_quantized_add_default              │                                 0 │                                    10 │
├────┼─────────────────────────────────────────────┼───────────────────────────────────┼───────────────────────────────────────┤
│  6 │ cortex_m_quantized_conv2d_default           │                                 0 │                                    35 │
├────┼─────────────────────────────────────────────┼───────────────────────────────────┼───────────────────────────────────────┤
│  7 │ cortex_m_quantized_depthwise_conv2d_default │                                 0 │                                    17 │
├────┼─────────────────────────────────────────────┼───────────────────────────────────┼───────────────────────────────────────┤
│  8 │ cortex_m_quantized_linear_default           │                                 0 │                                     1 │
├────┼─────────────────────────────────────────────┼───────────────────────────────────┼───────────────────────────────────────┤
│ 9  │ dim_order_ops__clone_dim_order_default      │                                 0 │                                     1 │
├────┼─────────────────────────────────────────────┼───────────────────────────────────┼───────────────────────────────────────┤
│ 10 │ Total                                       │                                 0 │                                    71 │
╘════╧═════════════════════════════════════════════╧═══════════════════════════════════╧═══════════════════════════════════════╛

Note E2E Inference tested on Alif E8 Board

Reviewers:

Subscribers:

Tasks:

Tags:

Author: Github Executorch

backends/arm/_passes/fold_qdq_with_annotated_qparams_pass.py

@@ -33,6 +33,14 @@
 from torch.fx import GraphModule, Node
 
 
+# Passthrough ops that preserve quantization parameters from input to output.
+# These ops should be foldable even without explicit annotation metadata.
+PASSTHROUGH_OPS = {
+    exir_ops.edge.aten.hardtanh.default,
+    exir_ops.edge.aten.relu.default,
+    exir_ops.edge.aten.clamp.default,
+}
+
 def _get_special_dtype(qspec: QuantArgs) -> TosaSpecialDtype | None:
     if qspec.dtype == torch.int8:
         if qspec.qmax == 7 and qspec.qmin == -7:
@@ -248,6 +256,26 @@ def _handle_control_flow_node(self, node: Node, graph_module: GraphModule):
                 submodule.graph.erase_node(node_to_remove)
         return
 
+    @staticmethod
+    def _has_dq_input_and_q_output(node: Node) -> bool:
+        """
+        Check if a node has dequantize input(s) and quantize output(s).
+        This indicates the node is part of a quantized computation path.
+        """
+        # Check if any input is from a dequantize op
+        has_dq_input = any(
+            isinstance(arg, Node) and arg.target in DQ_OPS
+            for arg in node.args
+            if isinstance(arg, Node)
+        )
+
+        # Check if any output goes to a quantize op
+        has_q_output = any(
+            user.target in Q_OPS
+            for user in node.users
+        )
+        return has_dq_input and has_q_output
+
     @staticmethod
     def is_foldable(node: Node) -> bool:
         if node.op != "call_function":
@@ -263,6 +291,13 @@ def is_foldable(node: Node) -> bool:
         ):
             return True
 
+        # Passthrough ops (hardtanh, relu, clamp) that have dq inputs and q outputs
+        # should be foldable even without explicit annotation. These ops preserve
+        # quantization parameters and are common in quantized models like MobileNetV2.
+        if node.target in PASSTHROUGH_OPS:
+            if FoldAndAnnotateQParamsPass._has_dq_input_and_q_output(node):
+                return True
+
         # We should not fold q-dq nodes into non-quantized nodes.
         if not (
             ArmAnnotationInfo.CUSTOM_META_KEY in node.meta.get("custom", {})
@@ -335,6 +370,35 @@ def call(self, graph_module: GraphModule) -> PassResult:  # noqa: C901
             ):
                 self._handle_control_flow_node(n, graph_module)
 
+        # Second pass: Propagate qparams through passthrough ops.
+        # For ops like hardtanh that share qparams with their input, we need to:
+        # 1. Copy output_qparams from the passthrough op to its input node
+        # 2. Set input_qparams on the passthrough op
+        for n in graph_module.graph.nodes:
+            n = cast(Node, n)
+            if n.target not in PASSTHROUGH_OPS:
+                continue
+
+            # Check if this passthrough op has output_qparams but missing input_qparams
+            has_output = "output_qparams" in n.meta and len(n.meta.get("output_qparams", {})) > 0
+            has_input = "input_qparams" in n.meta and len(n.meta.get("input_qparams", {})) > 0
+
+            if not has_output or has_input:
+                continue
+
+            # Get the input node
+            if len(n.args) == 0 or not isinstance(n.args[0], Node):
+                continue
+
+            input_node = n.args[0]
+
+            # Propagate: For passthrough ops, output qparams equal input qparams
+            if "output_qparams" not in input_node.meta:
+                input_node.meta["output_qparams"] = n.meta["output_qparams"]
+
+            # Set input_qparams from output_qparams (same for passthrough ops)
+            n.meta["input_qparams"] = {0: n.meta["output_qparams"][0]}
+
         # retrace the graph to update the fake tensor types
         graph_module = super().call(graph_module).graph_module
 

backends/cortex_m/passes/convert_to_cortex_m_pass.py

@@ -69,9 +69,117 @@ def _get_batch_size_from_conv(self, conv_node: torch.fx.Node):
             pass
         return None
 
+    def _get_addmm_replacement(self, node):
+        """
+        Handle aten.addmm (decomposed linear):
+        addmm(bias, input, weight.T) = input @ weight.T + bias
+
+        input_qparams indices for addmm:
+        [0] = bias (int32)
+        [1] = input activation (int8)
+        [2] = weight (int8)
+
+        The weight qparams at index [2] are guaranteed to be present because
+        CortexMQuantizer marks weight nodes as annotated, allowing
+        FoldAndAnnotateQParamsPass to properly fold Q/DQ nodes and populate qparams.
+        """
+        # Validate addmm node structure: addmm(bias, input, weight)
+        if len(node.args) < 3:
+            return None
+
+        bias_node = node.args[0]
+        input_node = node.args[1]
+        weights_node = node.args[2]
+
+        # Validate qparams are present with helpful error messages
+        input_qparams = node.meta.get("input_qparams", {})
+        if 1 not in input_qparams:
+            raise RuntimeError(
+                f"Missing input activation qparams at index 1 for addmm node '{node.name}'. "
+                f"Available input_qparams keys: {list(input_qparams.keys())}. "
+                "Ensure the model is properly quantized and FoldAndAnnotateQParamsPass ran."
+            )
+        if 2 not in input_qparams:
+            raise RuntimeError(
+                f"Missing weight qparams at index 2 for addmm node '{node.name}'. "
+                f"Available input_qparams keys: {list(input_qparams.keys())}. "
+                "Ensure CortexMQuantizer marked weight nodes and PropagateQParamsPass "
+                "propagated qparams through any transpose/permute ops."
+            )
+
+        # Get input activation qparams (index 1, not 0 which is bias!)
+        input_scale = input_qparams[1].scale
+        input_zp = input_qparams[1].zp
+
+        # Get weight qparams (index 2)
+        weight_scale = input_qparams[2].scale
+        weight_zp = input_qparams[2].zp
+
+        # Get output qparams
+        output_scale = node.meta["output_qparams"][0].scale
+        output_zp = node.meta["output_qparams"][0].zp
+        output_min = node.meta["output_qparams"][0].qmin
+        output_max = node.meta["output_qparams"][0].qmax
+
+        # Calculate quantization multiplier and shift
+        quantized_multiplier, quantized_shift = quantize_multiplier_aot(
+            (input_scale * weight_scale) / output_scale
+        )
+
+        # Get the original weight tensor
+        # Trace back through transpose/permute to find the placeholder
+        if weights_node.op == "call_function" and len(weights_node.args) > 0:
+            original_weight_node = weights_node.args[0]
+        else:
+            original_weight_node = weights_node
+
+        weights_tensor = get_param_tensor(self.exported_program, original_weight_node)
+        final_weights = weights_tensor.contiguous()
+
+        # Compute kernel_sum WITHOUT bias (pass None)
+        # Bias is passed separately to the C++ operator
+        kernel_sum_tensor = self._compute_kernel_sum(
+            final_weights, None, -input_zp, -weight_zp
+        )
+
+        # Create placeholders for weights and kernel_sum
+        with node.graph.inserting_after(original_weight_node):
+            weights_placeholder = create_constant_placeholder(
+                self.exported_program,
+                node.graph,
+                node.name + "_weights",
+                InputKind.PARAMETER,
+                final_weights,
+            )
+
+            kernel_sum = create_constant_placeholder(
+                self.exported_program,
+                node.graph,
+                node.name + "_kernel_sum",
+                InputKind.PARAMETER,
+                kernel_sum_tensor,
+            )
+
+        # Build args for cortex_m.quantized_linear
+        args = (
+            input_node,
+            weights_placeholder,
+            bias_node,
+            kernel_sum,
+            -input_zp,
+            -weight_zp,
+            output_zp,
+            [quantized_multiplier],
+            [quantized_shift],
+            output_max,
+            output_min,
+        )
+
+        return exir_ops.edge.cortex_m.quantized_linear.default, args
+
     def _get_linear_replacement(self, node):
         """
-         Let
+        Let
         - yi be the output activations (y1, ... yn)
         - xj be the input activations (x1, ... xm)
         - wij be the weights (w11, ... wnm)
@@ -386,6 +494,11 @@ def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
             match node.target:
                 case exir_ops.edge.aten.linear.default:
                     op, args = self._get_linear_replacement(node)
+                case exir_ops.edge.aten.addmm.default:
+                    result = self._get_addmm_replacement(node)
+                    if result is None:
+                        continue
+                    op, args = result
                 case exir_ops.edge.aten.convolution.default:
                     # Check if it's transposed convolution (arg index 6)
                     transposed = node.args[6] if len(node.args) > 6 else False

backends/cortex_m/passes/cortex_m_pass_manager.py

@@ -14,6 +14,12 @@
 from executorch.backends.transforms.replace_scalar_with_tensor import (
     ReplaceScalarWithTensorArgPass,
 )
+from executorch.backends.arm._passes.decompose_adaptive_avg_pool2d_pass import (
+    DecomposeAdaptiveAvgPool2dPass,
+)
+from executorch.backends.cortex_m.passes.propagate_qparams_pass import (
+    PropagateQParamsPass,
+)
 from executorch.exir.pass_base import ExportPass
 from executorch.exir.pass_manager import PassManager
 from executorch.exir.program._program import _transform
@@ -34,9 +40,11 @@ class CortexMPassManager(PassManager):
         # Run before folding so qparams attach to max_pool2d values, not tuple + getitem.
         RemoveGetItemPass,
         FoldAndAnnotateQParamsPass,
+        PropagateQParamsPass,
         ReplaceScalarWithTensorArgPass,
         ReplaceQuantNodesPass,
         ActivationFusionPass,
+        DecomposeAdaptiveAvgPool2dPass,
         DecomposeHardswishPass,
         QuantizedOpFusionPass,
         ConvertToCortexMPass,
@@ -49,12 +57,22 @@ class CortexMPassManager(PassManager):
         DecomposeMeanPass,
     ]
 
-    def __init__(self, exported_program, passes=None):
+    def __init__(self, exported_program, passes=None, skip_passes=None):
+        """
+        Initialize CortexMPassManager.
+
+        Args:
+            exported_program: The ExportedProgram to transform.
+            passes: Optional custom pass list. Uses default pass_list if None.
+            skip_passes: Optional list of pass classes to skip.
+        """
         self.exported_program = exported_program
         if passes is not None:
             self.passes = passes
         else:
-            self.passes = self.pass_list
+            self.passes = list(self.pass_list)
+        if skip_passes:
+            self.passes = [p for p in self.passes if p not in skip_passes]
 
     def transform_for_annotation(self, model):
         passes = self.pass_list_transform_for_annotation