Allow removing permute pairs in addition to transpose pairs (#10501)

Summary:

As titled. Gets us 27% better cycles on Activity Classification (at opt level 3).

Can be improved further, task is T222295719

Reviewed By: zonglinpeng

Differential Revision: D73619452

Author: mcremon-meta

backends/cadence/aot/fuse_ops.py

@@ -14,7 +14,7 @@
 import operator
 from collections import deque
 from numbers import Number
-from typing import cast, Sequence
+from typing import cast
 
 # Import these for the cadence function signatures.
 import executorch.backends.cadence.aot.ops_registrations  # noqa: F401
@@ -881,9 +881,9 @@ def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
 
 
 @register_cadence_pass(CadencePassAttribute(opt_level=1))
-class FuseTransposeOpPairsPass(FuseOpPairsAcrossBranchesPass):
+class FuseTransposeOrPermuteOpPairsPass(FuseOpPairsAcrossBranchesPass):
     """
-    Fuse transpose op pairs to a single view op.
+    Fuse transpose or permute op pairs to a single view op.
     """
 
     # A list of ops that can be bypassed when looking for a
@@ -907,42 +907,28 @@ def can_fuse_for_chain(
         if not super().can_fuse_for_chain(producer, consumer, consumer_op_packets):
             return False
 
-        def get_dims(node: torch.fx.Node) -> tuple[int, int]:
-            def canonicalize(dim: int) -> int:
-                if dim < 0:
-                    dim += len(node.meta["val"].shape)
-                return dim
-
-            return tuple(canonicalize(cast(int, d)) for d in node.args[1:3])
-
-        def is_equivalent(
-            shape: Sequence[int],
-            transpose0: tuple[int, int],
-            transpose1: tuple[int, int],
-        ) -> bool:
-            def permute_order(
-                order: Sequence[int], dims: tuple[int, int]
-            ) -> Sequence[int]:
-                new_order = list(order)
-                new_order[dims[0]], new_order[dims[1]] = (
-                    new_order[dims[1]],
-                    new_order[dims[0]],
-                )
-                return new_order
+        input_shape = list(cast(torch.fx.Node, producer.args[0]).meta["val"].shape)
 
-            order = permute_order(range(len(shape)), transpose0)
-            order = permute_order(order, transpose1)
+        intermediate_shape = (
+            get_transposed_dims(producer, input_shape)
+            if producer.target == exir_ops.edge.aten.transpose_copy.int
+            else get_permuted_dims(producer, input_shape)
+        )
 
-            non_unit_dims = [dim for dim in range(len(shape)) if shape[dim] != 1]
-            non_unit_dims_permuted = [dim for dim in order if shape[dim] != 1]
+        final_shape = (
+            get_transposed_dims(consumer, intermediate_shape)
+            if consumer.target == exir_ops.edge.aten.transpose_copy.int
+            else get_permuted_dims(consumer, intermediate_shape)
+        )
 
-            return non_unit_dims == non_unit_dims_permuted
+        non_unit_dims = [
+            input_shape[dim] for dim in range(len(input_shape)) if input_shape[dim] != 1
+        ]
+        non_unit_dims_permuted = [
+            final_shape[dim] for dim in range(len(final_shape)) if final_shape[dim] != 1
+        ]
 
-        return is_equivalent(
-            cast(torch.fx.Node, producer.args[0]).meta["val"].shape,
-            get_dims(producer),
-            get_dims(consumer),
-        )
+        return non_unit_dims == non_unit_dims_permuted
 
     def get_fused_node(
         self,
@@ -960,13 +946,20 @@ def get_fused_node(
         return view
 
     def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
-        # Remove any dequantize op that has only quantize ops as its users.
+        # Remove any transpose op pair that cancel each other.
         self.find_and_fuse(
             graph_module,
             producer_op_packets={exir_ops.edge.aten.transpose_copy},
             consumer_op_packets={exir_ops.edge.aten.transpose_copy},
             bypass_ops=self.bypass_ops,
         )
+        # Remove any permute op pair that cancel each other.
+        self.find_and_fuse(
+            graph_module,
+            producer_op_packets={exir_ops.edge.aten.permute_copy},
+            consumer_op_packets={exir_ops.edge.aten.permute_copy},
+            bypass_ops=self.bypass_ops,
+        )
         result = super().call(graph_module)
         return result
 
@@ -1028,5 +1021,5 @@ class CadenceFuseOpsInGraph:
         FuseQuantDequantToRequantizePass,
         FuseMulIntoDequantPass,
         FuseFullThenReshapePass,
-        FuseTransposeOpPairsPass,
+        FuseTransposeOrPermuteOpPairsPass,
     ]

backends/cadence/aot/passes.py

@@ -14,7 +14,7 @@
 from executorch.backends.cadence.aot.fuse_ops import (
     CadenceFuseOpsInGraph,
     FuseFullThenReshapePass,
-    FuseTransposeOpPairsPass,
+    FuseTransposeOrPermuteOpPairsPass,
 )
 from executorch.backends.cadence.aot.pass_utils import (
     CadencePassAttribute,
@@ -83,7 +83,7 @@ def get_passes_in_default_order() -> List[ExportPass]:
         CadenceSimplifyOpsInGraph.passes,
         FinalizePipeline,
         FuseFullThenReshapePass,
-        FuseTransposeOpPairsPass,
+        FuseTransposeOrPermuteOpPairsPass,
         RemoveNopSliceOrViewOpPass,
     ]
     return pytree.tree_flatten(passes)[0]

backends/cadence/aot/tests/test_fusion_ops_passes.py

@@ -20,7 +20,7 @@
     FuseFullThenReshapePass,
     FuseMulIntoDequantPass,
     FuseQuantDequantToRequantizePass,
-    FuseTransposeOpPairsPass,
+    FuseTransposeOrPermuteOpPairsPass,
 )
 from executorch.backends.cadence.aot.graph_builder import GraphBuilder
 from executorch.backends.cadence.aot.pass_utils import count_node, op_counts_match
@@ -510,7 +510,7 @@ def test_fuse_then_transpose_pass(self):
         )
 
 
-class TestFuseTransposeOpPairsPass(TestFusionPassesBase):
+class TestFuseTransposeOrPermuteOpPairsPass(TestFusionPassesBase):
     def _create_operator(
         self, builder: GraphBuilder, op: torch._ops.OpOverload, x: ProxyValue
     ) -> ProxyValue:
@@ -536,17 +536,17 @@ def _create_operator(
     def test_fuse_transpose_pairs(self, op: torch._ops.OpOverload):
         # Create a graph with transpose -> quant -> transpose.
         builder = GraphBuilder()
-        x = builder.placeholder("x", torch.randn(2, 3))
-        transpose_node = builder.call_operator(
+        x = builder.placeholder("x", torch.randn(2, 3, 4))
+        transpose_node0 = builder.call_operator(
             op=exir_ops.edge.aten.transpose_copy.int,
             args=(x, 0, 1),
         )
-        quant_node = self._create_operator(builder, op, transpose_node)
-        transpose_node = builder.call_operator(
+        quant_node = self._create_operator(builder, op, transpose_node0)
+        transpose_node1 = builder.call_operator(
             op=exir_ops.edge.aten.transpose_copy.int,
-            args=(quant_node, 0, 1),
+            args=(quant_node, 1, 2),
         )
-        builder.output([transpose_node])
+        builder.output([transpose_node1])
         gm = builder.get_graph_module()
         self.check_op_counts(
             gm,
@@ -557,7 +557,7 @@ def test_fuse_transpose_pairs(self, op: torch._ops.OpOverload):
         )
 
         # Check that the pass fuses the two transpose ops.
-        fusion_pass_result = FuseTransposeOpPairsPass()(gm)
+        fusion_pass_result = FuseTransposeOrPermuteOpPairsPass()(gm)
         self.assertIsNotNone(fusion_pass_result)
         gm_after_pass = fusion_pass_result.graph_module
         self.check_op_counts(
@@ -568,6 +568,47 @@ def test_fuse_transpose_pairs(self, op: torch._ops.OpOverload):
             },
         )
 
+    @parameterized.expand(
+        [
+            exir_ops.edge.quantized_decomposed.quantize_per_tensor.default,
+            exir_ops.edge.cadence.quantized_relu.per_tensor,
+        ],
+    )
+    def test_fuse_permute_pairs(self, op: torch._ops.OpOverload):
+        # Create a graph with permute -> quant -> permute.
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(8, 2, 3, 4))
+        permute_node0 = builder.call_operator(
+            op=exir_ops.edge.aten.permute_copy.default,
+            args=(x, [0, 3, 1, 2]),
+        )
+        quant_node = self._create_operator(builder, op, permute_node0)
+        permute_node1 = builder.call_operator(
+            op=exir_ops.edge.aten.permute_copy.default,
+            args=(quant_node, [0, 2, 3, 1]),
+        )
+        builder.output([permute_node1])
+        gm = builder.get_graph_module()
+        self.check_op_counts(
+            gm,
+            expected_op_counts={
+                exir_ops.edge.aten.permute_copy.default: 2,
+                op: 1,
+            },
+        )
+
+        # Check that the pass fuses the two transpose ops.
+        fusion_pass_result = FuseTransposeOrPermuteOpPairsPass()(gm)
+        self.assertIsNotNone(fusion_pass_result)
+        gm_after_pass = fusion_pass_result.graph_module
+        self.check_op_counts(
+            gm_after_pass,
+            expected_op_counts={
+                exir_ops.edge.aten.permute_copy.default: 0,
+                op: 1,
+            },
+        )
+
     def test_no_fusion_for_transpose_pairs(self):
         # Create a graph with transpose -> quant -> transpose.
         builder = GraphBuilder()
@@ -595,7 +636,7 @@ def test_no_fusion_for_transpose_pairs(self):
         )
 
         # No fusion.
-        gm_after_pass = FuseTransposeOpPairsPass()(gm).graph_module
+        gm_after_pass = FuseTransposeOrPermuteOpPairsPass()(gm).graph_module
         self.check_op_counts(
             gm_after_pass,
             expected_op_counts={
@@ -604,6 +645,42 @@ def test_no_fusion_for_transpose_pairs(self):
             },
         )
 
+    def test_no_fusion_for_permute_pairs(self):
+        # Create a graph with permute -> quant -> permute.
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(2, 3, 4))
+        permute_node = builder.call_operator(
+            op=exir_ops.edge.aten.permute_copy.default,
+            args=(x, [2, 0, 1]),
+        )
+        quant_node = builder.call_operator(
+            op=exir_ops.edge.quantized_decomposed.quantize_per_tensor.default,
+            args=(permute_node, 1.2, 3, 0, 127, torch.int8),
+        )
+        permute_node = builder.call_operator(
+            op=exir_ops.edge.aten.permute_copy.default,
+            args=(quant_node, [2, 0, 1]),
+        )
+        builder.output(permute_node)
+        gm = builder.get_graph_module()
+        self.check_op_counts(
+            gm,
+            expected_op_counts={
+                exir_ops.edge.aten.permute_copy.default: 2,
+                exir_ops.edge.quantized_decomposed.quantize_per_tensor.default: 1,
+            },
+        )
+
+        # No fusion.
+        gm_after_pass = FuseTransposeOrPermuteOpPairsPass()(gm).graph_module
+        self.check_op_counts(
+            gm_after_pass,
+            expected_op_counts={
+                exir_ops.edge.aten.permute_copy.default: 2,
+                exir_ops.edge.quantized_decomposed.quantize_per_tensor.default: 1,
+            },
+        )
+
     def test_fusion_for_forked_transposes(self):
         # Create a graph with transpose -> quant -> transpose.
         builder = GraphBuilder()
@@ -636,7 +713,7 @@ def test_fusion_for_forked_transposes(self):
         )
 
         # Fuse the all the transpose ops.
-        gm_after_pass = FuseTransposeOpPairsPass()(gm).graph_module
+        gm_after_pass = FuseTransposeOrPermuteOpPairsPass()(gm).graph_module
         self.check_op_counts(
             gm_after_pass,
             expected_op_counts={