NXP backend: Delegate input/output format transpositions by default.

Author: MartinPavella

backends/nxp/backend/ir/conversion_config.py

@@ -13,8 +13,7 @@ def __init__(self, args: dict | None = None):
 
         :param args: Optional dictionary with conversion arguments. Unknown arguments are ignored.
         """
-        self.use_neutron_for_format_conversion: bool = False
-        self.keep_io_format: bool = False
+        self.use_neutron_for_format_conversion: bool = True
         self.allow_inputs_stripping: bool = True
         self.qdq_aware_conversion: bool = True
         self.symbolic_dimensions_mapping: dict[str, int] | None = None

backends/nxp/nxp_backend.py

@@ -45,7 +45,7 @@ def __init__(self):
         self.output_format = None
         self.operators_not_to_delegate: List[str] = []
         self.neutron_converter_flavor = None
-        self.use_neutron_for_format_conversion = False
+        self.use_neutron_for_format_conversion = True
 
     def _replace_colons(self, operator: str) -> str:
         """
@@ -59,7 +59,7 @@ def neutron_compile_spec(
         neutron_converter_flavor: str,
         extra_flags: Optional[str] = None,
         operators_not_to_delegate: Optional[List[str]] = None,
-        use_neutron_for_format_conversion: bool = False,
+        use_neutron_for_format_conversion: bool = True,
     ):
         """
         Generate compile spec for Neutron NPU
@@ -127,7 +127,7 @@ def generate_neutron_compile_spec(
     system_config: Optional[str] = None,
     extra_flags: Optional[str] = None,
     operators_not_to_delegate: Optional[List[str]] = None,
-    use_neutron_for_format_conversion: bool = False,
+    use_neutron_for_format_conversion: bool = True,
 ) -> List[CompileSpec]:
     return (
         NeutronCompileSpecBuilder()

backends/nxp/tests/executorch_pipeline.py

@@ -105,7 +105,7 @@ def to_quantized_edge_program(
     remove_quant_io_ops=False,
     custom_delegation_options=CustomDelegationOptions(),  # noqa B008
     get_quantizer_fn=None,
-    use_neutron_for_format_conversion=False,
+    use_neutron_for_format_conversion=True,
 ) -> EdgeProgramManager:
     _neutron_target_spec = NeutronTargetSpec(target, neutron_converter_flavor)
     if get_quantizer_fn is None:
@@ -156,8 +156,13 @@ def to_quantized_edge_program(
 def to_quantized_executorch_program(
     model: torch.nn.Module,
     input_spec: tuple[ModelInputSpec, ...] | tuple[int, ...] | list[tuple[int, ...]],
+    use_neutron_for_format_conversion: bool = True,
 ) -> ExecutorchProgramManager:
-    edge_program_manager = to_quantized_edge_program(model, input_spec)
+    edge_program_manager = to_quantized_edge_program(
+        model,
+        input_spec,
+        use_neutron_for_format_conversion=use_neutron_for_format_conversion,
+    )
 
     return edge_program_manager.to_executorch(
         config=ExecutorchBackendConfig(extract_delegate_segments=False)

backends/nxp/tests/ir/converter/node_converter/test_abs_converter.py

@@ -14,9 +14,10 @@
 from executorch.backends.nxp.tests.executors import (
     convert_run_compare,
     graph_contains_any_of_ops,
-    ToNCHWPreprocess,
-    ToNHWCPreprocess,
+    ToChannelFirstPreprocess,
+    ToChannelLastPreprocess,
 )
+
 from executorch.exir.dialects._ops import ops as exir_ops
 from torch.export import ExportedProgram
 
@@ -67,7 +68,9 @@ def test_conv_abs(mocker, input_shape: tuple[int] = (1, 3, 112, 112)):
 
     converter_spy = mocker.spy(EdgeProgramToIRConverter, "convert_program")
 
-    quantized_program = to_quantized_edge_program(model, input_shape).exported_program()
+    quantized_program = to_quantized_edge_program(
+        model, input_shape, use_neutron_for_format_conversion=False
+    ).exported_program()
 
     tflite_flatbuffers_model, io_formats = converter_spy.spy_return
     exported_program: ExportedProgram = converter_spy.call_args.args[1]
@@ -80,8 +83,8 @@ def test_conv_abs(mocker, input_shape: tuple[int] = (1, 3, 112, 112)):
     convert_run_compare(
         exported_program,
         tfl_model=tflite_flatbuffers_model,
-        tflite_input_preprocess=ToNHWCPreprocess(),
-        tflite_output_preprocess=ToNCHWPreprocess(),
+        tflite_input_preprocess=ToChannelLastPreprocess(),
+        tflite_output_preprocess=ToChannelFirstPreprocess(),
         input_data=input_data,
         atol=1.0,
     )

backends/nxp/tests/ir/converter/node_converter/test_adaptive_avg_pool2d_converter.py

@@ -47,7 +47,9 @@ def test_adaptive_avg_pool_2d_delegated_quant_conversion(
     converter_spy = mocker.spy(EdgeProgramToIRConverter, "convert_program")
 
     # Run conversion
-    edge_program = to_quantized_edge_program(model, input_shape).exported_program()
+    edge_program = to_quantized_edge_program(
+        model, input_shape, use_neutron_for_format_conversion=False
+    ).exported_program()
     nodes = [str(node) for node in edge_program.graph.nodes]
 
     # Input size is a multiple of output size, can be converted to AveragePool, node is delegated
@@ -91,7 +93,9 @@ def test_adaptive_avg_pool_2d_non_delegated_quant_conversion(
     converter_spy = mocker.spy(EdgeProgramToIRConverter, "convert_program")
 
     # Run conversion
-    edge_program = to_quantized_edge_program(model, input_shape).exported_program()
+    edge_program = to_quantized_edge_program(
+        model, input_shape, use_neutron_for_format_conversion=False
+    ).exported_program()
     nodes = list(edge_program.graph.nodes)
 
     # Input size is not a multiple of output size, cannot be converted to AveragePool, node is not delegated
@@ -122,7 +126,9 @@ def test_adaptive_avg_pool_2d_mean_dim_quant_conversion(mocker):
     converter_spy = mocker.spy(EdgeProgramToIRConverter, "convert_program")
 
     # Run conversion
-    _ = to_quantized_edge_program(model, input_shape)
+    _ = to_quantized_edge_program(
+        model, input_shape, use_neutron_for_format_conversion=False
+    )
 
     # Capture generated model
     tflite_flatbuffers_model, io_formats = converter_spy.spy_return

backends/nxp/tests/ir/converter/node_converter/test_add_tensor_converter.py

@@ -103,7 +103,9 @@ def test_add_tensor_w_conv_quant_conversion(mocker, input_shape):
     converter_spy = mocker.spy(EdgeProgramToIRConverter, "convert_program")
 
     # Run conversion
-    _ = to_quantized_edge_program(model, input_shape)
+    _ = to_quantized_edge_program(
+        model, input_shape, use_neutron_for_format_conversion=False
+    )
 
     # Capture generated model
     tflite_flatbuffers_model, io_formats = converter_spy.spy_return

backends/nxp/tests/ir/converter/node_converter/test_avg_pool2d_converter.py

@@ -6,10 +6,11 @@
 import numpy as np
 import pytest
 import torch
-
 from executorch.backends.nxp.backend.edge_program_converter import (
     EdgeProgramToIRConverter,
 )
+
+from executorch.backends.nxp.backend.ir.conversion_config import ConversionConfig
 from executorch.backends.nxp.backend.ir.converter.builder.model_builder import (
     ModelBuilder,
 )
@@ -91,6 +92,9 @@ def test_avg_pool_2d_conversion(input_shape, padding, count_include_pad):
         input_data,
         tflite_input_preprocess=ToNHWCPreprocess(),
         tflite_output_preprocess=ToNCHWPreprocess(),
+        conversion_config=ConversionConfig(
+            {"use_neutron_for_format_conversion": False}
+        ),
     )
 
 
@@ -145,7 +149,9 @@ def test_avg_pool_2d_quant_conversion(mocker, input_shape, padding, count_includ
     converter_spy = mocker.spy(EdgeProgramToIRConverter, "convert_program")
 
     # Run conversion
-    _ = to_quantized_edge_program(model, input_shape)
+    _ = to_quantized_edge_program(
+        model, input_shape, use_neutron_for_format_conversion=False
+    )
 
     # Capture generated model
     tflite_flatbuffers_model, io_formats = converter_spy.spy_return
@@ -172,7 +178,9 @@ def test_avg_pool_2d_quant_conversion__padded(mocker):
     ops_spy = mocker.spy(ModelBuilder, "finish")
 
     # Run conversion
-    _ = to_quantized_edge_program(model, input_shape)
+    _ = to_quantized_edge_program(
+        model, input_shape, use_neutron_for_format_conversion=False
+    )
 
     # Capture the converter operators.
     ops = ops_spy.spy_return.sub_graphs[0].operators.vector

backends/nxp/tests/ir/converter/node_converter/test_clone_converter.py

@@ -113,7 +113,7 @@ def test_conv_dropout_quant(self, inplace_dropout: bool, input_shape: tuple[int]
             owner=EdgeProgramToIRConverter,
         ) as converter_spy:
             quantized_program = to_quantized_edge_program(
-                model, input_shape
+                model, input_shape, use_neutron_for_format_conversion=False
             ).exported_program()
 
             tflite_flatbuffers_model, _ = converter_spy.calls[-1].return_value

backends/nxp/tests/ir/converter/node_converter/test_constant_pad_nd_converter.py

@@ -7,6 +7,7 @@
 import pytest
 import torch
 
+from executorch.backends.nxp.backend.ir.conversion_config import ConversionConfig
 from executorch.backends.nxp.tests.executorch_pipeline import (
     to_edge_program,
     to_quantized_edge_program,
@@ -101,6 +102,9 @@ def test_constant_pad_nd_conversion__channels_first(input_shape, paddings):
         input_data,
         tflite_input_preprocess=ToNHWCPreprocess(),
         tflite_output_preprocess=ToNCHWPreprocess(),
+        conversion_config=ConversionConfig(
+            {"use_neutron_for_format_conversion": False}
+        ),
     )
 
 

backends/nxp/tests/ir/converter/node_converter/test_conv_converter.py

@@ -10,6 +10,7 @@
 from executorch.backends.nxp.backend.edge_program_converter import (
     EdgeProgramToIRConverter,
 )
+from executorch.backends.nxp.backend.ir.conversion_config import ConversionConfig
 from executorch.backends.nxp.backend.ir.converter.builder.model_builder import (
     ModelBuilder,
 )
@@ -375,7 +376,9 @@ def test_conv2d_quant_conversion(mocker, model: torch.nn.Module, input_shape):
     converter_spy = mocker.spy(EdgeProgramToIRConverter, "convert_program")
 
     # Run conversion
-    _ = to_quantized_edge_program(model, input_shape)
+    _ = to_quantized_edge_program(
+        model, input_shape, use_neutron_for_format_conversion=False
+    )
 
     # Capture generated model
     tflite_flatbuffers_model, io_formats = converter_spy.spy_return
@@ -451,6 +454,7 @@ def test_conv2d_conversion__depthwise__quantized(
             kernel_size=kernel_shape,
         ),
         tuple(input_shape),
+        use_neutron_for_format_conversion=False,
     ).exported_program()
 
     ops = spy.spy_return.sub_graphs[0].operators.vector
@@ -485,6 +489,9 @@ def test_conv2d_conversion__depthwise__padded(padding, mocker):
         tflite_input_preprocess=ToChannelLastPreprocess(),
         tflite_output_preprocess=ToChannelFirstPreprocess(),
         atol=4e-7,
+        conversion_config=ConversionConfig(
+            {"use_neutron_for_format_conversion": False}
+        ),
     )
     conversion_result = spy.spy_return
     ops = conversion_result.sub_graphs[0].operators.vector
@@ -505,6 +512,7 @@ def test_conv2d_conversion__depthwise__padded__quantized(padding, mocker):
             group=group, in_channels=group, out_channels=group, padding=padding
         ),
         tuple(input_shape),
+        use_neutron_for_format_conversion=False,
     ).exported_program()
 
     ops = spy.spy_return.sub_graphs[0].operators.vector