Pass to replace Adaptive Avg. Pool with Aten Avg. Pool

Differential Revision: D74559775

Pull Request resolved: #10818

Author: skrtskrtfb

backends/cadence/aot/replace_ops.py

@@ -16,6 +16,7 @@
 
 # pyre-unsafe
 
+import logging
 import math
 import operator
 from operator import neg
@@ -2346,6 +2347,66 @@ def resolve_full_arg(self, x_arg, const_arg):
         return const_arg
 
 
+@register_cadence_pass(CadencePassAttribute(opt_level=0))
+class ReplaceAdaptiveAvgPoolWithAtenAvgPoolPass(ExportPass):
+    """
+    Replace the aten adaptive avg_pool op with the aten avg_pool2d op.
+    """
+
+    def call_operator(self, op, args, kwargs, meta):
+        # Only continue for avg_pool op
+        if op not in {exir_ops.edge.aten._adaptive_avg_pool2d.default}:
+            return super().call_operator(op, args, kwargs, meta)
+
+        # Get the input tensor
+        in_tensor = args[0].to_tensor() if isinstance(args[0], ProxyValue) else args[0]
+        # Permute NCHW to NHWC for computation
+        in_tensor_permuted = in_tensor.permute(0, 2, 3, 1)
+        in_tensor_shape = in_tensor_permuted.shape
+
+        output_size = args[1]
+        num_dims = len(output_size)
+
+        # TODO: If in_tensor_shape is not a multiple of output size,
+        # this pass will not work. T224984800
+        dim_multiples = [
+            (in_tensor_shape[i + 1] % output_size[i]) == 0 for i in range(num_dims)
+        ]
+        if not all(dim_multiples):
+            logging.info(
+                f"Unable to replace adaptive average pool with average pool. Input tensor shape of {in_tensor_shape} is not a multiple of output size: {output_size}"
+            )
+            return super().call_operator(op, args, kwargs, meta)
+
+        # Compute stride and kernel_size, then set default values for other arguments
+        stride = [(in_tensor_shape[i + 1] // output_size[i]) for i in range(num_dims)]
+        kernel_size = [
+            in_tensor_shape[i + 1] - (output_size[i] - 1) * stride[i]
+            for i in range(num_dims)
+        ]
+        padding = [0] * num_dims
+        ceil_mode = False
+        count_include_pad = True
+        divisor_override = None
+
+        # Create a new avg_pool node with the updated args
+        new_args = (
+            args[0],
+            kernel_size,
+            stride,
+            padding,
+            ceil_mode,
+            count_include_pad,
+            divisor_override,
+        )
+        return super().call_operator(
+            exir_ops.edge.aten.avg_pool2d.default,
+            new_args,
+            kwargs,
+            meta,
+        )
+
+
 # This class encapsulates all the functions that replace/switch one op in the
 # graph with another.
 class CadenceReplaceOpsInGraph:
@@ -2382,6 +2443,7 @@ class CadenceReplaceOpsInGraph:
         ReplacePT2QuantWithCadenceQuantPass,
         ReplacePT2DequantWithCadenceDequantPass,
         ReplaceSingleElementTensorArgumentsFromFullOpWithScalarPass,
+        ReplaceAdaptiveAvgPoolWithAtenAvgPoolPass,
         ReplaceAtenAvgPoolWithJarvisAvgPoolPass,
         ReplaceWhereWithFullArgsWithWhereScalar,
         ReplaceAtenApproxGeluWithApproxGeluPass,

backends/cadence/aot/tests/test_replace_ops_passes.py

@@ -19,6 +19,7 @@
 from executorch.backends.cadence.aot.replace_ops import (
     ForceChannelLastForConvPass,
     MakeSliceAndCatDimOutermostPass,
+    ReplaceAdaptiveAvgPoolWithAtenAvgPoolPass,
     ReplaceAddMMWithLinearPass,
     ReplaceAtenApproxGeluWithApproxGeluPass,
     ReplaceAtenConvolutionWithJarvisConvolutionPass,
@@ -1936,3 +1937,102 @@ def test_extract_mul_argument_to_full(
                 },
             )
         )
+
+
+class TestReplaceAdaptiveAvgPoolWithAtenAvgPoolPass(unittest.TestCase):
+    def _get_adaptive_avg_pool_gm(
+        self, input_shape: Tuple[int, int, int, int], output_shape: Tuple[int, int]
+    ) -> torch.fx.GraphModule:
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(*input_shape))
+        adaptive_avg_pool2d = builder.call_operator(
+            exir_ops.edge.aten._adaptive_avg_pool2d.default, (x, output_shape)
+        )
+        builder.output([adaptive_avg_pool2d])
+        return builder.get_graph_module()
+
+    def test_replace_adaptive_avg_pool_with_aten_avg_pool(self) -> None:
+        gm = self._get_adaptive_avg_pool_gm((1, 64, 128, 128), (8, 8))
+        self.assertEqual(
+            len(
+                gm.graph.find_nodes(
+                    op="call_function",
+                    target=exir_ops.edge.aten._adaptive_avg_pool2d.default,
+                )
+            ),
+            1,
+        )
+        self.assertEqual(
+            len(
+                gm.graph.find_nodes(
+                    op="call_function",
+                    target=exir_ops.edge.aten.avg_pool2d.default,
+                )
+            ),
+            0,
+        )
+        p = ReplaceAdaptiveAvgPoolWithAtenAvgPoolPass()
+        updated_gm = p.call(gm).graph_module
+        self.assertEqual(
+            len(
+                updated_gm.graph.find_nodes(
+                    op="call_function",
+                    target=exir_ops.edge.aten._adaptive_avg_pool2d.default,
+                )
+            ),
+            0,
+        )
+        avg_pool2d_nodes = updated_gm.graph.find_nodes(
+            op="call_function", target=exir_ops.edge.aten.avg_pool2d.default
+        )
+        self.assertEqual(
+            len(avg_pool2d_nodes),
+            1,
+        )
+        avg_pool2d_node = avg_pool2d_nodes[0]
+
+        self.assertEqual(avg_pool2d_node.args[1], [16, 16])  # kernel_size is 16x16
+        self.assertEqual(avg_pool2d_node.args[2], [16, 16])  # stride is 16, 16
+        self.assertEqual(avg_pool2d_node.args[3], [0, 0])  # padding is 0, 0
+        self.assertEqual(avg_pool2d_node.args[4], False)  # ceil_mode is False
+        self.assertEqual(avg_pool2d_node.args[5], True)  # count_include_pad is True
+        self.assertEqual(avg_pool2d_node.args[6], None)  # divisor_override is None
+
+    def test_replace_adaptive_avg_pool_with_aten_avg_pool_irregular(self) -> None:
+        gm = self._get_adaptive_avg_pool_gm((1, 64, 128, 128), (9, 9))
+        self.assertEqual(
+            len(
+                gm.graph.find_nodes(
+                    op="call_function",
+                    target=exir_ops.edge.aten._adaptive_avg_pool2d.default,
+                )
+            ),
+            1,
+        )
+        self.assertEqual(
+            len(
+                gm.graph.find_nodes(
+                    op="call_function", target=exir_ops.edge.aten.avg_pool2d.default
+                )
+            ),
+            0,
+        )
+        # Shapes are not multiples of each other, so pass will not trigger
+        p = ReplaceAdaptiveAvgPoolWithAtenAvgPoolPass()
+        updated_gm = p.call(gm).graph_module
+        self.assertEqual(
+            len(
+                updated_gm.graph.find_nodes(
+                    op="call_function",
+                    target=exir_ops.edge.aten._adaptive_avg_pool2d.default,
+                )
+            ),
+            1,
+        )
+        avg_pool2d_nodes = updated_gm.graph.find_nodes(
+            op="call_function", target=exir_ops.edge.aten.avg_pool2d.default
+        )
+        self.assertEqual(
+            len(avg_pool2d_nodes),
+            0,
+        )