Introduce apply_torch_ops_passes to test mul fusion e2e (#11741)

Summary:
Pull Request resolved: #11741

This diff
1) extends export_to_edge() with apply_torch_ops_passes() call
We need this to grab the actual mul argument value and fusing in later passes. Otherwise, the constant gets lifted by _lower_ep_to_edge()
2) implements e2e mul fusion test.
We test that both ReplaceMulTensorWithMulAndFullOpsPass() and FuseMulTensorIntoQuantPass() passes get applied correctly and which results into mul.Tensor being removed completely.

Reviewed By: hsharma35

Differential Revision: D76469613

Author: eigen-k

backends/cadence/aot/compiler.py

@@ -41,7 +41,8 @@
 from torch.export.exported_program import ExportedProgram
 from torchao.quantization.pt2e.quantize_pt2e import convert_pt2e, prepare_pt2e
 
-from .passes import get_cadence_passes
+from .passes import get_edge_passes
+from .passes import apply_torch_ops_passes
 
 from .utils import print_ops_info
 
@@ -265,6 +266,9 @@ def export_to_edge(
     # Export the model into an ExportedProgram.
     expo_program = trace(model, inputs)
 
+    # Apply passes which transform the ExportedProgram before it gets lowered to edge.
+    apply_torch_ops_passes(expo_program)
+
     # Lower the model to edge IR.
     edge_prog_manager = _lower_ep_to_edge(expo_program, dump_graphs, constant_methods)
 
@@ -306,7 +310,7 @@ def _lower_ep_to_cadence(
     Lower an existing ExportedProgram to edge IR and apply frontend optimization passes.
     """
     edge_prog_manager = _lower_ep_to_edge(program, dump_graphs=dump_graphs)
-    cadence_passes = get_cadence_passes(opt_level)
+    cadence_passes = get_edge_passes(opt_level)
 
     # Run a couple required passes for quant/dequant ops
     cadence_prog_manager = edge_prog_manager.transform(
@@ -324,7 +328,7 @@ def export_to_cadence(
     opt_level: int = 1,
 ) -> EdgeProgramManager:
     edge_prog_manager = export_to_edge(model, inputs, dump_graphs=dump_graphs)
-    cadence_passes = get_cadence_passes(opt_level)
+    cadence_passes = get_edge_passes(opt_level)
 
     # Run a couple required passes for quant/dequant ops
     cadence_prog_manager = edge_prog_manager.transform(
@@ -368,7 +372,7 @@ def export_to_executorch_gen_etrecord(
     memory_config: Optional[MemoryConfig] = None,
     dump_graphs: bool = False,
 ) -> ExecutorchProgramManager:
-    cadence_passes = get_cadence_passes(opt_level)
+    cadence_passes = get_edge_passes(opt_level)
     edge_prog_manager = export_to_edge(model, inputs, dump_graphs)
 
     # Run a couple required passes for quant/dequant ops

backends/cadence/aot/passes.py

@@ -6,7 +6,7 @@
 
 # pyre-strict
 
-from typing import Any, List, Optional
+from typing import Any, Callable, List, Optional
 
 import torch
 import torch.fx
@@ -28,13 +28,17 @@
     RemoveRedundantOps,
 )
 from executorch.backends.cadence.aot.reorder_ops import CadenceReorderOpsInGraph
-from executorch.backends.cadence.aot.replace_ops import CadenceReplaceOpsInGraph
+from executorch.backends.cadence.aot.replace_ops import (
+    CadenceReplaceOpsInGraph,
+    ReplaceMulTensorWithMulAndFullOpsPass,
+)
 from executorch.backends.cadence.aot.simplify_ops import CadenceSimplifyOpsInGraph
 from executorch.exir.pass_base import ExportPass, PassResult
 from executorch.exir.pass_manager import PassManager, PassType
 from executorch.exir.passes import dead_code_elimination_pass
 from executorch.exir.passes.scalar_to_tensor_pass import ScalarToTensorPass
 from executorch.exir.passes.spec_prop_pass import SpecPropPass
+from torch.export.exported_program import ExportedProgram
 
 
 @register_cadence_pass(CadencePassAttribute(opt_level=0))
@@ -89,7 +93,7 @@ def get_passes_in_default_order() -> List[ExportPass]:
     return pytree.tree_flatten(passes)[0]
 
 
-def get_cadence_passes(
+def get_edge_passes(
     opt_level: int,
 ) -> List[Optional[PassResult]]:
     passes = get_passes_in_default_order()
@@ -100,3 +104,14 @@ def get_cadence_passes(
         for filtered_pass in list(filter(pass_filter, passes))
     ]
     return filtered_passes
+
+def apply_torch_ops_passes(expo_program: ExportedProgram) -> None:
+    """
+    Applies compiler passes on torch.ops IR, including torch.ops.aten, torch.ops.cadence, etc.
+    expo_program is expected to be the output of the torch.export.export().
+    """
+
+    aten_passes: List[Callable[[torch.fx.GraphModule], Optional[PassResult]]] = [
+        ReplaceMulTensorWithMulAndFullOpsPass()
+    ]
+    PassManager(aten_passes)(expo_program.graph_module)