Fix and re-enable eliminate_permutations graph transformation (#824)

Summary:
Pull Request resolved: #824

Fix eliminate_permutations graph transformation in AITemplate to retain successive cancelling permutations on strided input Tensor.

Re-enabled the transformation as well.

Differential Revision: D47371853

fbshipit-source-id: a2eb55549ca510482c57f2fa03b81d88642aff04

python/aitemplate/compiler/transform/optimize_graph.py

@@ -51,6 +51,7 @@
 from aitemplate.compiler.transform.transform_odd_alignment import (
     transform_odd_alignment,
 )
+from aitemplate.compiler.transform.transform_permutations import eliminate_permutations
 from aitemplate.compiler.transform.transform_permute_to_reshape import (
     transform_permute_to_reshape,
 )
@@ -125,8 +126,7 @@ def optimize_graph(
         split_large_split_ops,
         transform_permute_to_reshape,
         transform_memory_ops,
-        # FIXME: temporarily disable this due to some accuracy issue
-        # eliminate_permutations,
+        eliminate_permutations,
     ]
 
     if not optimize:

python/aitemplate/compiler/transform/transform_permutations.py

@@ -17,6 +17,7 @@
 import numpy as np
 
 from aitemplate.compiler.base import Operator, Tensor
+from aitemplate.compiler.tensor_accessor import TensorAccessor
 from aitemplate.compiler.transform import transform_utils
 
 
@@ -60,6 +61,32 @@ def remove_second_permutation_from_graph(
     transform_utils.remove_tensor_from_sorted_graph(output_tensor)
 
 
+def _reshaped_or_strided_input_or_output_accessor(op: Operator) -> bool:
+    def _reshaped_or_strided_tensor_accessor(accessor: TensorAccessor) -> bool:
+        if (
+            accessor.actual_shapes is not None
+            and accessor.actual_shapes != accessor.original_shapes
+        ):
+            return True
+
+        # Is it a strided accessor
+        if hasattr(accessor, "stride_dim") and accessor.stride_dim is not None:
+            return True
+
+        return False
+
+    input_accessors = op._attrs.get("input_accessors", None)
+    output_accessors = op._attrs.get("output_accessors", None)
+
+    return (
+        (input_accessors is not None)
+        and _reshaped_or_strided_tensor_accessor(input_accessors[0])
+    ) or (
+        (output_accessors is not None)
+        and _reshaped_or_strided_tensor_accessor(output_accessors[0])
+    )
+
+
 def eliminate_permutations(
     sorted_graph: List[Tensor], workdir: str = None
 ) -> List[Tensor]:
@@ -73,12 +100,7 @@ def eliminate_permutations(
                 continue
             if not cur_op._attrs["op"].startswith("permute"):
                 continue
-            input_accessors = cur_op._attrs.get("input_accessors", None)
-            if (
-                input_accessors is not None
-                and hasattr(input_accessors[0], "strided_dim")
-                and input_accessors[0].strided_dim is not None
-            ):
+            if _reshaped_or_strided_input_or_output_accessor(cur_op):
                 continue
             curr_op_output = cur_op._attrs["outputs"][0]
             dst_ops = curr_op_output._attrs["dst_ops"]
@@ -89,6 +111,8 @@ def eliminate_permutations(
             for next_op in dst_ops:
                 if not next_op._attrs["op"].startswith("permute"):
                     continue
+                if _reshaped_or_strided_input_or_output_accessor(next_op):
+                    continue
                 p1 = get_permutation(cur_op)
                 p2 = get_permutation(next_op)
                 if not np.all(np.array(p1)[p2] == np.arange(0, len(p1))):

tests/unittest/compiler/test_eliminate_permutations.py

@@ -28,7 +28,6 @@
 )
 
 
-@unittest.skip("Skip until we fix the accuracy issue")
 class EliminatePermutationTestCase(unittest.TestCase):
     def test_eliminate_permutation(self):
         dtype = "float"
@@ -190,6 +189,126 @@ def test_eliminate_permutation_all_permutations(self):
         self.assertEqual(len(result_graph), 3)
         self.assertTrue(graph_has_op(result_graph, "permute"))
 
+    def test_do_not_eliminate_permutation_of_strided_input(self):
+        dtype = "float"
+        shape = [3, 2, 4]
+        new_shape = [3, 2 * 2]
+        target = detect_target()
+
+        x = Tensor(shape, name="x", dtype=dtype, is_input=True)
+        s1 = ops.dynamic_slice()(
+            x, start_indices=[0, 0, 2], end_indices=[2147483647, 2147483647, 4]
+        )
+        p1 = ops.permute()(s1, dims=[0, 2, 1])
+        p2 = ops.permute()(p1, dims=[0, 2, 1])
+        z = ops.reshape()(p2, new_shape)
+        z._attrs["is_output"] = True
+        z._attrs["name"] = "z"
+
+        with compile_model(
+            z, target, "./tmp", "test_do_not_eliminate_permutation_of_strided_input"
+        ) as module:
+            # Verify the generated graph.
+            sorted_graph = module.debug_sorted_graph
+            self.assertEqual(len(sorted_graph), 4)
+            self.assertTrue(graph_has_op(sorted_graph, "permute021"))
+
+            x_pt = get_random_torch_tensor(shape, dtype)
+            z_pt = get_torch_empty_tensor(new_shape, dtype)
+
+            module.run_with_tensors({"x": x_pt}, {"z": z_pt})
+
+            self.assertTrue(
+                torch.equal(
+                    torch.reshape(torch.split(x_pt, 2, dim=2)[1], new_shape), z_pt
+                )
+            )
+
+    def test_do_not_eliminate_permutation_of_strided_input2(self):
+        dtype = "float"
+        shape = [3, 4, 2]
+        new_shape = [3, 2 * 2]
+        target = detect_target()
+
+        x = Tensor(shape, name="x", dtype=dtype, is_input=True)
+        p1 = ops.permute()(x, dims=[0, 2, 1])
+        s1 = ops.dynamic_slice()(
+            p1, start_indices=[0, 0, 2], end_indices=[2147483647, 2147483647, 4]
+        )
+        p2 = ops.permute()(s1, dims=[0, 2, 1])
+        z = ops.reshape()(p2, new_shape)
+        z._attrs["is_output"] = True
+        z._attrs["name"] = "z"
+
+        with compile_model(
+            z, target, "./tmp", "test_do_not_eliminate_permutation_of_strided_input2"
+        ) as module:
+            # Verify the generated graph.
+            sorted_graph = module.debug_sorted_graph
+            self.assertEqual(len(sorted_graph), 4)
+            self.assertTrue(graph_has_op(sorted_graph, "permute021"))
+
+            x_pt = get_random_torch_tensor(shape, dtype)
+            z_pt = get_torch_empty_tensor(new_shape, dtype)
+
+            module.run_with_tensors({"x": x_pt}, {"z": z_pt})
+
+            self.assertTrue(
+                torch.equal(
+                    torch.reshape(
+                        torch.permute(
+                            torch.split(torch.permute(x_pt, (0, 2, 1)), 2, dim=2)[1],
+                            (0, 2, 1),
+                        ),
+                        new_shape,
+                    ),
+                    z_pt,
+                )
+            )
+
+    def test_do_not_eliminate_permutation_of_reshaped_input(self):
+        dtype = "float"
+        shape = [3, 2, 4]
+        new_shape = [3, 2, 4]
+        target = detect_target()
+
+        x = Tensor(shape, name="x", dtype=dtype, is_input=True)
+        p1 = ops.permute()(x, dims=[0, 2, 1])
+        r1 = ops.reshape()(p1, new_shape)
+        p2 = ops.permute()(r1, dims=[0, 2, 1])
+        z = ops.dynamic_slice()(
+            p2, start_indices=[0, 0, 1], end_indices=[2147483647, 2147483647, 2]
+        )
+        z._attrs["is_output"] = True
+        z._attrs["name"] = "z"
+
+        with compile_model(
+            z, target, "./tmp", "test_do_not_eliminate_permutation_of_reshaped_input"
+        ) as module:
+            # Verify the generated graph.
+            sorted_graph = module.debug_sorted_graph
+            self.assertEqual(len(sorted_graph), 4)
+            self.assertTrue(graph_has_op(sorted_graph, "permute021"))
+
+            x_pt = get_random_torch_tensor(shape, dtype)
+            z_pt = get_torch_empty_tensor([3, 4, 1], dtype)
+
+            module.run_with_tensors({"x": x_pt}, {"z": z_pt})
+
+            self.assertTrue(
+                torch.equal(
+                    torch.split(
+                        torch.permute(
+                            torch.reshape(torch.permute(x_pt, (0, 2, 1)), new_shape),
+                            (0, 2, 1),
+                        ),
+                        1,
+                        dim=2,
+                    )[1],
+                    z_pt,
+                )
+            )
+
 
 if __name__ == "__main__":
     unittest.main()