Canonicalize on permutes (#18822)

Summary:

First step toward canonicalizing on permute in the graph compiler passes.

Reviewed By: ethansfng

Differential Revision: D100379751

Author: mcremon-meta

backends/cadence/aot/compiler_utils.py

@@ -87,6 +87,21 @@ def broadcastable(shape_1: Sequence[int], shape_2: Sequence[int]) -> bool:
     )
 
 
+def transpose_dims_to_permute_order(ndim: int, dim0: int, dim1: int) -> List[int]:
+    """
+    Convert transpose(dim0, dim1) to an equivalent permute order list.
+    E.g., transpose(0, 1) on a 3D tensor gives [1, 0, 2].
+    """
+    # Normalize negative dims
+    if dim0 < 0:
+        dim0 += ndim
+    if dim1 < 0:
+        dim1 += ndim
+    order = list(range(ndim))
+    order[dim0], order[dim1] = order[dim1], order[dim0]
+    return order
+
+
 def get_transposed_dims(
     node: torch.fx.Node, dims: Optional[List[int]] = None
 ) -> List[int]:

backends/cadence/aot/quantizer/fusion_pass.py

@@ -678,10 +678,13 @@ def call(self, graph_module: fx.GraphModule) -> PassResult:  # noqa: C901
                             quant_node,
                         )
                     elif isinstance(pattern, AddmmPattern):
-                        # Transpose the weight tensor
+                        # Transpose the weight tensor using permute
+                        weight_ndim = len(weights_inputs[0].meta["val"].shape)
+                        perm = list(range(weight_ndim))
+                        perm[0], perm[1] = perm[1], perm[0]
                         transposed_weights = graph_module.graph.call_function(
-                            torch.ops.aten.transpose.int,
-                            (weights_inputs[0], 0, 1),
+                            torch.ops.aten.permute.default,
+                            (weights_inputs[0], perm),
                         )
                         assert (
                             "val" in weights_inputs[0].meta
@@ -692,7 +695,7 @@ def call(self, graph_module: fx.GraphModule) -> PassResult:  # noqa: C901
                         ), "fake_mode is None on weight node"
                         with original_val.fake_mode:
                             transposed_weights.meta["val"] = (
-                                torch.ops.aten.transpose.int(original_val, 0, 1)
+                                torch.ops.aten.permute.default(original_val, perm)
                             )
                         copy_node_metadata(transposed_weights, weights_inputs[0])
 

backends/cadence/aot/replace_ops.py

@@ -19,7 +19,10 @@
 
 import torch
 import torch.fx
-from executorch.backends.cadence.aot.compiler_utils import quantize_tensor_multiplier
+from executorch.backends.cadence.aot.compiler_utils import (
+    quantize_tensor_multiplier,
+    transpose_dims_to_permute_order,
+)
 from executorch.backends.cadence.aot.fuse_ops import FuseCascadedTransposeOrPermuteOps
 from executorch.backends.cadence.aot.pass_utils import (
     CadencePassAttribute,
@@ -355,9 +358,9 @@ def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
 
         # Handle transpose: if mat2 is a transpose op, extract the original tensor
         transposed_mat2 = False
-        if (
-            mat2.op == "call_function"
-            and mat2.target == exir_ops.edge.aten.transpose_copy.int
+        if mat2.op == "call_function" and (
+            mat2.target == exir_ops.edge.aten.transpose_copy.int
+            or mat2.target == exir_ops.edge.aten.permute_copy.default
         ):
             # mat2 is already transposed, so we use the input to the transpose
             mat2 = cast(torch.fx.Node, mat2.args[0])
@@ -405,9 +408,11 @@ def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
         # Transpose mat2 if it wasn't already transposed
         if not transposed_mat2:
             with graph.inserting_before(node):
+                ndim = len(mat2.meta["val"].shape)
+                perm = transpose_dims_to_permute_order(ndim, -1, -2)
                 mat2 = graph.call_function(
-                    exir_ops.edge.aten.transpose_copy.int,
-                    args=(mat2, -1, -2),
+                    exir_ops.edge.aten.permute_copy.default,
+                    args=(mat2, perm),
                 )
 
                 # Metadata copy important
@@ -430,6 +435,35 @@ def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
         return True
 
 
+@register_cadence_pass(CadencePassAttribute(opt_level=1))
+class ReplaceTransposeWithPermutePass(RemoveOrReplacePassInterface):
+    """
+    Replace transpose_copy.int ops with equivalent permute_copy.default ops
+    to canonicalize on permute as the single layout-change op.
+    """
+
+    @property
+    def targets(self) -> list[EdgeOpOverload]:
+        return [exir_ops.edge.aten.transpose_copy.int]
+
+    def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
+        in_tensor = node.args[0]
+        assert isinstance(in_tensor, torch.fx.Node)
+        ndim = len(in_tensor.meta["val"].shape)
+        dim0 = cast(int, node.args[1])
+        dim1 = cast(int, node.args[2])
+        perm = transpose_dims_to_permute_order(ndim, dim0, dim1)
+
+        with node.graph.inserting_before(node):
+            new_node = node.graph.call_function(
+                exir_ops.edge.aten.permute_copy.default,
+                args=(in_tensor, perm),
+            )
+            new_node.meta = node.meta
+        node.replace_all_uses_with(new_node)
+        return True
+
+
 @register_cadence_pass(CadencePassAttribute(opt_level=1))
 class ReplacePermuteWithTransposePass(RemoveOrReplacePassInterface):
     """
@@ -471,7 +505,6 @@ def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
 
         return False
 
-
 @register_cadence_pass(CadencePassAttribute(opt_level=0))
 class ReplaceConvolutionOptionalArgsWithConcreteArgsPass(RemoveOrReplacePassInterface):
     """
@@ -798,9 +831,11 @@ def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
                 # gather stencil. Also, the first two dimensions of weight must be
                 # transposed/interchanged.
                 assert isinstance(weight, torch.fx.Node)
+                weight_ndim = len(weight.meta["val"].shape)
+                perm = transpose_dims_to_permute_order(weight_ndim, 0, 1)
                 transposed_weight = node.graph.call_function(
-                    exir_ops.edge.aten.transpose_copy.int,
-                    args=(weight, 0, 1),
+                    exir_ops.edge.aten.permute_copy.default,
+                    args=(weight, perm),
                 )
                 transposed_weight.meta = weight.meta
 
@@ -1036,18 +1071,19 @@ def targets(self) -> list[EdgeOpOverload]:
     def _transpose_dims(
         self, graph: torch.fx.Graph, node: torch.fx.Node, dim0: int, dim1: int
     ) -> torch.fx.Node:
-        """Helper function to transpose dims of a node."""
+        """Helper function to transpose dims of a node using permute."""
         shape = node.meta["val"].shape
         dim0, dim1 = (
             canonicalize_transposed_dim(dim0, shape),
             canonicalize_transposed_dim(dim1, shape),
         )
         dim0, dim1 = min(dim0, dim1), max(dim0, dim1)
-        transpose_node = graph.call_function(
-            exir_ops.edge.aten.transpose_copy.int, (node, dim0, dim1), {}
+        perm = transpose_dims_to_permute_order(len(shape), dim0, dim1)
+        permute_node = graph.call_function(
+            exir_ops.edge.aten.permute_copy.default, (node, perm), {}
         )
-        transpose_node.meta = node.meta
-        return transpose_node
+        permute_node.meta = node.meta
+        return permute_node
 
     def _change_nchw_to_nhwc(
         self, graph: torch.fx.Graph, node: torch.fx.Node
@@ -1263,18 +1299,19 @@ def targets(self) -> list[EdgeOpOverload]:
     def _transpose_dims(
         self, graph: torch.fx.Graph, node: torch.fx.Node, dim0: int, dim1: int
     ) -> torch.fx.Node:
-        """Helper function to transpose dims of a node."""
+        """Helper function to transpose dims of a node using permute."""
         shape = node.meta["val"].shape
         dim0, dim1 = (
             canonicalize_transposed_dim(dim0, shape),
             canonicalize_transposed_dim(dim1, shape),
         )
         dim0, dim1 = min(dim0, dim1), max(dim0, dim1)
-        transpose_node = graph.call_function(
-            exir_ops.edge.aten.transpose_copy.int, (node, dim0, dim1), {}
+        perm = transpose_dims_to_permute_order(len(shape), dim0, dim1)
+        permute_node = graph.call_function(
+            exir_ops.edge.aten.permute_copy.default, (node, perm), {}
         )
-        transpose_node.meta = node.meta
-        return transpose_node
+        permute_node.meta = node.meta
+        return permute_node
 
     def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
         # Get the dimension argument
@@ -1526,8 +1563,8 @@ def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
         if not channel_last:
             with graph.inserting_before(node):
                 linear_res = graph.call_function(
-                    exir_ops.edge.aten.transpose_copy.int,
-                    args=(linear_res, 1, 2),
+                    exir_ops.edge.aten.permute_copy.default,
+                    args=(linear_res, [0, 2, 1]),
                 )
                 linear_res.meta = node.meta
 
@@ -1717,8 +1754,8 @@ def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
         if not channel_last:
             with graph.inserting_before(node):
                 linear_res = graph.call_function(
-                    exir_ops.edge.aten.transpose_copy.int,
-                    args=(linear_res, 1, 2),
+                    exir_ops.edge.aten.permute_copy.default,
+                    args=(linear_res, [0, 2, 1]),
                 )
                 linear_res.meta = node.meta
 
@@ -2375,9 +2412,11 @@ def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
 
         # Transpose Y_arg
         with graph.inserting_before(node):
+            Y_ndim = len(Y_tensor_val.shape)
+            perm = transpose_dims_to_permute_order(Y_ndim, -1, -2)
             Y_arg_t = graph.call_function(
-                exir_ops.edge.aten.transpose_copy.int,
-                args=(Y_arg, -1, -2),
+                exir_ops.edge.aten.permute_copy.default,
+                args=(Y_arg, perm),
             )
             Y_arg_t.meta = node.meta
 
@@ -2410,13 +2449,10 @@ def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
         result = super().call(graph_module)
         modified = modified or result.modified
         if modified:
-            # Fuse any inserted transpose node with transpose/permute nodes
+            # Fuse any inserted permute node with transpose/permute nodes
             # surrounding it.
             result = FuseCascadedTransposeOrPermuteOps().call(result.graph_module)
             modified = modified or result.modified
-            # Replace permute with transpose.
-            result = ReplacePermuteWithTransposePass().call(result.graph_module)
-            modified = modified or result.modified
 
         return PassResult(result.graph_module, modified)
 
@@ -2640,10 +2676,10 @@ def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
 # graph with another.
 class CadenceReplaceOpsInGraph:
     passes = CommonReplacePasses.passes + [
+        ReplaceTransposeWithPermutePass,
         ReplaceAtenLinalgSvdWithCadenceLinalgSvdPass,
         ReplaceEmptyTensorsWithFullPass,
         ReplaceFunctionallyEquivalentOpTargets,
-        ReplacePermuteWithTransposePass,
         ReplaceConvolutionOptionalArgsWithConcreteArgsPass,
         ReplaceAddMMWithLinearPass,
         ReplacePadWithCatPass,
@@ -2657,6 +2693,8 @@ class CadenceReplaceOpsInGraph:
         ReplaceIm2RowWithViewPass,
         MakeSliceAndCatDimOutermostPass,
         ReplaceMatmulWithTransposedMatmulPass,
+        # Convert permutes back to transposes after all passes that create them.
+        ReplacePermuteWithTransposePass,
         ReplaceNopTransposeOrPermuteWithViewPass,
         ReplaceLinearWithFullyConnectedOpPass,
         ReplaceScalarTensorWithFullPass,