NXP backend: Enable new Neutron C flow support for Clamp operator

Author: StrycekSimon

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

@@ -3,15 +3,29 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+from copy import copy
+
+import numpy as np
+import torch
 from executorch.backends.nxp.backend.edge_helper import try_get_arg
 from executorch.backends.nxp.backend.ir.converter.node_converter import (
+    _is_dequant_node,
+    _is_quant_node,
     CustomDelegationOptions,
     is_not_qdq_node,
     NodeConverter,
 )
+from executorch.backends.nxp.backend.ir.converter.quantization_utils import (
+    propagate_quantization,
+)
 from executorch.backends.nxp.backend.ir.lib.tflite.BuiltinOperator import (
     BuiltinOperator,
 )
+from executorch.backends.nxp.backend.ir.tflite_generator import tflite_model
+from executorch.backends.nxp.backend.ir.tflite_generator.builtin_options import (
+    maximum_options,
+    minimum_options,
+)
 from executorch.backends.nxp.backend.neutron_operator_support import (
     activation_supported_on_target,
 )
@@ -21,15 +35,25 @@
 from torch.nn import Parameter
 
 
+def _is_convertible_to_relu(node):
+    bounds = ClampConverter._get_clamp_bounds(node)
+
+    # Only some specific bounds are supported on the target hardware.
+    if bounds not in ClampConverter.RELU_COMPATIBLE_BOUNDS.values():
+        return False
+
+    return True
+
+
 class ClampConverter(NodeConverter):
-    SUPPORTED_BOUNDS = {
+    RELU_COMPATIBLE_BOUNDS = {
         "ReluN1To1": (-1, 1),
         "Relu0To1": (0, 1),
         "Relu6": (0, 6),
         "Relu": (0, None),
     }
 
-    BOUNDS_TO_NEUTRON_IR_OP = {
+    BOUNDS_TO_RELU_NEUTRON_IR_OP = {
         (-1, 1): BuiltinOperator.RELU_N1_TO_1,
         (0, 1): BuiltinOperator.RELU_0_TO_1,
         (0, 6): BuiltinOperator.RELU6,
@@ -53,27 +77,52 @@ def _is_supported_in_IR(
         # No NeutronIR-specific restrictions.
         return True
 
+    @staticmethod
+    def _io_quant_is_same(node: Node):
+        quant = next(iter(node.users.keys()))
+        dequant = node.args[0]
+
+        if not _is_dequant_node(dequant):
+            return False
+
+        if not _is_quant_node(quant):
+            return False
+
+        q_params = quant.args[1:]
+        dq_params = dequant.args[1:]
+        return all([q == dq for q, dq in zip(q_params, dq_params)])
+
     @staticmethod
     def _is_supported_on_target(
         node: Node,
         neutron_target_spec: NeutronTargetSpec,
         parameters_mapping: dict[str, Parameter],
         custom_delegation_options: CustomDelegationOptions,
     ) -> bool:
-        bounds = ClampConverter._get_clamp_bounds(node)
+        relu_compatible = _is_convertible_to_relu(node)
+
+        if neutron_target_spec.use_new_flow_neutron_c:
+            io_quant_consistent = ClampConverter._io_quant_is_same(node)
+            quant_supported = NodeConverter.uses_quantization_type_for_io(
+                node,
+                supported_types=[torch.int8, torch.uint8],
+                input_indices=[0],
+                output_indices=[0],
+            )
 
-        # Only some specific bounds are supported on the target hardware.
-        if bounds not in ClampConverter.SUPPORTED_BOUNDS.values():
-            return False
+            # We either convert to ReLU -> SingleInputQuantization pattern
+            # or we convert to Min/Max, which requires same quantization on
+            # both input and output.
+            return (relu_compatible | io_quant_consistent) and quant_supported
 
-        return True
+        return relu_compatible
 
     @classmethod
     def supports_partitioning_result(
         cls,
         node: Node,
         partition_list: list[Partition],
-        custom_delegation_options: CustomDelegationOptions,
+        _: CustomDelegationOptions,
         neutron_target_spec: NeutronTargetSpec,
         parameters_mapping: dict[str, Parameter],
     ) -> bool:
@@ -82,7 +131,10 @@ def supports_partitioning_result(
         # Neutron cannot delegate a partition where ReLU or ReLU6 is the only operator
         # and at the same time the node does not satisfy delegation requirements.
         # In contrast, ReLUN1To1 and ReLU0To1 are supported and delegated successfuly.
-        if bounds in [cls.SUPPORTED_BOUNDS["Relu"], cls.SUPPORTED_BOUNDS["Relu6"]]:
+        if bounds in [
+            cls.RELU_COMPATIBLE_BOUNDS["Relu"],
+            cls.RELU_COMPATIBLE_BOUNDS["Relu6"],
+        ]:
             is_alone_in_partition = cls.is_node_alone_in_partition(
                 node, partition_list, filter_fn=is_not_qdq_node
             )
@@ -91,8 +143,21 @@ def supports_partitioning_result(
 
         return True
 
+    @staticmethod
+    def _quantize_value(
+        value: int,
+        zp: int,
+        scale: float,
+        quant_min: int,
+        quant_max: int,
+        dtype: type = np.int8,
+    ) -> np.integer:
+        rescaled_value = round(value / scale) + zp
+        return dtype(np.clip(rescaled_value, quant_min, quant_max))
+
     def convert(self, node: Node):
-        """Convert the `aten.clamp.default` operator to Neutron IR `Relu*` operators.
+        """Convert the `aten.clamp.default` operator to either
+        Neutron IR `Relu*` operator or combination of `Min` and `Max`.
         The schema is:
             aten::clamp(
                 Tensor self,
@@ -101,13 +166,69 @@ def convert(self, node: Node):
             ) -> Tensor
         """
         self.assert_convertible(node)
+        to_relu = _is_convertible_to_relu(node)
 
         bounds = self._get_clamp_bounds(node)
-
         t_op = self._create_tflite_op_with_io_tensors(node)
 
-        # noinspection PyTypeChecker,PyUnboundLocalVariable
-        t_op.opcode_index = self.builder.op_code_index_for_op_type(
-            self.BOUNDS_TO_NEUTRON_IR_OP[bounds]
-        )
-        self.builder.append_operators([t_op])
+        if not self.neutron_target_spec.use_new_flow_neutron_c or to_relu:
+            # noinspection PyTypeChecker,PyUnboundLocalVariable
+            t_op.opcode_index = self.builder.op_code_index_for_op_type(
+                self.BOUNDS_TO_RELU_NEUTRON_IR_OP[bounds]
+            )
+            self.builder.append_operators([t_op])
+            return
+
+        q_node = node.args[0]
+        assert _is_dequant_node(q_node)
+        _, scale, zp, quant_min, quant_max, _ = q_node.args
+
+        x = t_op.tmp_inputs[0]
+        y = t_op.tmp_outputs[0]
+
+        if x.quantization is not None and y.quantization is None:
+            propagate_quantization(x, y)
+
+        if x.quantization != y.quantization:
+            raise AssertionError(
+                "Input and output quantization should be same in order to convert to max/min."
+            )
+
+        min_value, max_value = bounds
+
+        if min_value is not None:
+            min_value = self._quantize_value(min_value, zp, scale, quant_min, quant_max)
+            min_tensor = self.builder.create_tensor_for_data(
+                np.array([min_value], np.int8), "min"
+            )
+            propagate_quantization(x, min_tensor)
+
+        if max_value is not None:
+            max_value = self._quantize_value(max_value, zp, scale, quant_min, quant_max)
+            max_tensor = self.builder.create_tensor_for_data(
+                np.array([max_value], np.int8), "max"
+            )
+            propagate_quantization(x, max_tensor)
+
+        if None not in bounds:
+            tmp_y = self.builder.duplicate_tensor(x)
+            tmp_x = tmp_y
+            propagate_quantization(x, tmp_y)
+        else:
+            tmp_y = y
+            tmp_x = x
+
+        ops_to_add = []
+        if max_value is not None:
+            min_op = tflite_model.Operator(builtin_options=minimum_options.Minimum())
+            min_op.tmp_inputs = [x, max_tensor]
+            min_op.tmp_outputs = [tmp_y]
+            ops_to_add.append(min_op)
+
+        if min_value is not None:
+            max_op = tflite_model.Operator(builtin_options=maximum_options.Maximum())
+            max_op.tmp_inputs = [tmp_x, min_tensor]
+            max_op.tmp_outputs = [y]
+            ops_to_add.append(max_op)
+
+        self.builder.append_operators(ops_to_add)

backends/nxp/quantizer/patterns.py

@@ -10,6 +10,9 @@
 from functools import partial
 
 import torch
+from executorch.backends.nxp.backend.ir.converter.node_converters.ops_converters.clamp_converter import (
+    _is_convertible_to_relu,
+)
 from executorch.backends.nxp.quantizer.utils import (
     get_bias_qparams,
     get_bias_qparams_transp_conv,
@@ -115,8 +118,9 @@ class SharedSpecPattern(QuantizationPattern):
     def partition_types(self) -> list[torch.nn.Module]:
         pass
 
-    def get_anchors(
-        self, gm: fx.GraphModule, fused_partition: list[fx.GraphModule]
+    @staticmethod
+    def get_shared_spec_anchors(
+        gm: fx.GraphModule, fused_partition: list[fx.GraphModule]
     ) -> PartitionAnchors | None:
         node = fused_partition[0].nodes[-1]
         assert len(fused_partition[0].input_nodes) == 1
@@ -137,15 +141,21 @@ def get_anchors(
             ],
         )
 
+    def get_anchors(
+        self, gm: fx.GraphModule, fused_partition: list[fx.GraphModule]
+    ) -> PartitionAnchors | None:
+        return self.get_shared_spec_anchors(gm, fused_partition)
+
 
 class SingleInputBasicPattern(QuantizationPattern):
     @abstractmethod
     def partition_types(self) -> list[OpOverload]:
         pass
 
-    def get_anchors(
-        self, gm: fx.GraphModule, fused_partition: list[fx.GraphModule]
-    ) -> PartitionAnchors | None:
+    @staticmethod
+    def get_single_input_anchors(
+        gm: fx.GraphModule, fused_partition: list[fx.GraphModule]
+    ):
         node = fused_partition[0].nodes[-1]
 
         return PartitionAnchors(
@@ -155,6 +165,11 @@ def get_anchors(
             output=[(node,)],
         )
 
+    def get_anchors(
+        self, gm: fx.GraphModule, fused_partition: list[fx.GraphModule]
+    ) -> PartitionAnchors | None:
+        return self.get_single_input_anchors(gm, fused_partition)
+
 
 class BatchNormPattern(QuantizationPattern):
     def partition_types(self) -> list[OpOverload]:
@@ -408,12 +423,25 @@ def get_anchors(
         )
 
 
-class ClampPattern(SingleInputBasicPattern):
+class ClampPattern(QuantizationPattern):
     """Quantizer for the `aten.clamp.default` operator."""
 
     def partition_types(self):
         return [torch.ops.aten.clamp.default]
 
+    def get_anchors(
+        self, gm: fx.GraphModule, fused_partition: list[fx.GraphModule]
+    ) -> PartitionAnchors | None:
+        node = fused_partition[0].nodes[-1]
+
+        if (
+            self.neutron_quantizer.neutron_target_spec.use_new_flow_neutron_c
+            and not _is_convertible_to_relu(node)
+        ):
+            return SharedSpecPattern.get_shared_spec_anchors(gm, fused_partition)
+        else:
+            return SingleInputBasicPattern.get_single_input_anchors(gm, fused_partition)
+
 
 def _is_batch_norm(node_: Node) -> bool:
     return node_.op == "call_function" and node_.target in [