Revert "ArBackend: Enable Pybindings for tosa_serialization lib (pytorch#15356)"

This reverts commit fdfeaa4.

Author: GregoryComer

.mypy.ini

@@ -83,12 +83,6 @@ ignore_missing_imports = True
 [mypy-tosa_tools.*]
 ignore_missing_imports = True
 
-[mypy-tosa_serializer]
-ignore_missing_imports = True
-
-[mypy-tosa_serializer.*]
-ignore_missing_imports = True
-
 [mypy-setuptools.*]
 ignore_missing_imports = True
 

backends/arm/common/debug.py

@@ -7,9 +7,8 @@
 import os
 from typing import Optional
 
+import serializer.tosa_serializer as ts
 import torch
-
-import tosa_serializer as ts
 from executorch.exir.print_program import inspect_node
 
 logger = logging.getLogger(__name__)
@@ -51,20 +50,29 @@ def get_node_debug_info(
     return output
 
 
-# Output TOSA flatbuffer for debugging
-def debug_tosa_dump(tosa_graph: bytes, path: str, suffix: str = ""):
+# Output TOSA flatbuffer and test harness file
+def debug_tosa_dump(tosa_graph: ts.TosaSerializer, path: str, suffix: str = ""):
     filename = f"output{suffix}.tosa"
 
     logger.info(f"Emitting debug output to: {path=}, {suffix=}")
 
     os.makedirs(path, exist_ok=True)
 
+    fb = tosa_graph.serialize()
+    js = tosa_graph.writeJson(filename)
+
     filepath_tosa_fb = os.path.join(path, filename)
     with open(filepath_tosa_fb, "wb") as f:
-        f.write(tosa_graph)
+        f.write(fb)
     if not os.path.exists(filepath_tosa_fb):
         raise IOError("Failed to write TOSA flatbuffer")
 
+    filepath_desc_json = os.path.join(path, f"desc{suffix}.json")
+    with open(filepath_desc_json, "w") as f:
+        f.write(js)
+    if not os.path.exists(filepath_desc_json):
+        raise IOError("Failed to write TOSA JSON")
+
 
 def debug_fail(
     node,
@@ -73,7 +81,7 @@ def debug_fail(
     path: Optional[str] = None,
 ):
     logger.warning("Internal error due to poorly handled node:")
-    if tosa_graph is not None and path:
-        debug_tosa_dump(tosa_graph.serialize(), path)
+    if tosa_graph is not None and path is not None:
+        debug_tosa_dump(tosa_graph, path)
         logger.warning(f"Debug output captured in '{path}'.")
     debug_node(node, graph_module)

backends/arm/debug/schema.py

@@ -10,8 +10,8 @@
 from dataclasses import asdict, dataclass
 from typing import Any, Optional
 
+import serializer.tosa_serializer as ts
 import torch
-import tosa_serializer as ts
 
 from executorch.backends.arm.common.arm_compile_spec import ArmCompileSpec
 
@@ -114,18 +114,23 @@ def to_dict(self) -> dict[str, Any]:
 class DebugHook:
     def __init__(self, debug_mode: ArmCompileSpec.DebugMode) -> None:
         self._debug_events: list[DebugSchema] = []
+        self.__op_id_to_name = {}
         self.mode = debug_mode
 
-    def add(self, node: torch.fx.Node, tosa_op: Any, tosa_op_id: ts.Op) -> DebugSchema:
+        # Build up a mapping from TOSA 1.0 operator IDs to their names
+        for name, val in vars(ts.Op).items():
+            self.__op_id_to_name[val] = name
+
+    def add(self, node: torch.fx.Node, tosa_op: Any, tosa_op_id: int) -> DebugSchema:
         tosa_debug_info = None
 
         # If the debug data is being embedded into the TOSA flatbuffer
         # do not collect TOSADebugSchema data, it's redundent
         if self.mode != ArmCompileSpec.DebugMode.TOSA:
             tosa_debug_info = TosaDebugSchema(
                 node_name=str(tosa_op),
-                operator_name=str(tosa_op_id),
-                operator_id=int(tosa_op_id),
+                operator_name=self.__op_id_to_name[tosa_op_id],
+                operator_id=tosa_op_id,
             )
 
         aten_debug_info = ATenDebugSchema.from_node(node)

backends/arm/ethosu/backend.py

@@ -51,15 +51,6 @@ def _compile_tosa_flatbuffer(
                 "compile_flags are required in the CompileSpec list for EthosUBackend"
             )
 
-        # Vela tooling only supports flatbuffers up to 2 GiB.
-        max_flatbuffer_size = 2 * 1024 * 1024 * 1024
-        flatbuffer_size = len(tosa_flatbuffer)
-        if flatbuffer_size > max_flatbuffer_size:
-            raise RuntimeError(
-                "TOSA flatbuffer is too large for Vela "
-                f"({flatbuffer_size} bytes > {max_flatbuffer_size} bytes limit)."
-            )
-
         # Pass on the TOSA flatbuffer to the vela compiler.
         binary = vela_compile(
             tosa_flatbuffer,

backends/arm/operators/node_visitor.py

@@ -9,7 +9,6 @@
 from typing import Any, Dict, List, Optional
 
 import torch
-import tosa_serializer as ts
 
 from executorch.backends.arm.common.arm_compile_spec import ArmCompileSpec
 from executorch.backends.arm.debug.schema import DebugHook
@@ -47,12 +46,12 @@ def _serialize_operator(
         self,
         node: torch.fx.Node,
         tosa_graph: Any,
-        tosa_op: ts.Op,
+        tosa_op: Any,
         inputs: List[str],
         outputs: List[str],
         attributes: Optional[Any] = None,
     ) -> None:
-        op_location = ts.TosaOpLocation()
+        op_location = ""
         if self.debug_hook:
             debug_info = self.debug_hook.add(
                 node,
@@ -61,7 +60,7 @@ def _serialize_operator(
             )
 
             if self.debug_hook.mode == ArmCompileSpec.DebugMode.TOSA:
-                op_location.text = json.dumps(debug_info.to_dict())
+                op_location = json.dumps(debug_info.to_dict())
 
         tosa_graph.addOperator(
             tosa_op,

backends/arm/operators/op_abs.py

@@ -6,7 +6,7 @@
 # pyre-unsafe
 from typing import Any, List
 
-import tosa_serializer as ts
+import serializer.tosa_serializer as ts
 
 from executorch.backends.arm.operators.node_visitor import (
     NodeVisitor,
@@ -48,13 +48,11 @@ def define_node(
             output.tosa_spec,
         )
 
-        attr = ts.TosaSerializerAttribute()
-        attr.AbsAttribute()
-        self._serialize_operator(
-            node,
-            tosa_graph,
-            ts.Op.ABS,
-            [inputs[0].name],
+        tosa_graph.addOperator(
+            ts.TosaOp.Op().ABS,
+            [
+                inputs[0].name,
+            ],
             [output.name],
-            attr,
+            None,
         )

backends/arm/operators/op_add.py

@@ -9,7 +9,7 @@
 
 import executorch.backends.arm.tosa.quant_utils as tqutils
 import executorch.backends.arm.tosa.utils as tutils
-import tosa_serializer as ts
+import serializer.tosa_serializer as ts
 
 from executorch.backends.arm.operators.node_visitor import (
     NodeVisitor,
@@ -81,16 +81,15 @@ def define_node(
             add_output = output
 
         input1, input2 = rescaled_inputs
-        attr = ts.TosaSerializerAttribute()
-        attr.AddAttribute()
+
         # Do the INT32 Add
         self._serialize_operator(
             node,
             tosa_graph,
-            ts.Op.ADD,
+            ts.TosaOp.Op().ADD,
             [input1.name, input2.name],
             [add_output.name],
-            attr,
+            None,
         )
 
         if output.dtype == ts.DType.INT8:
@@ -144,14 +143,13 @@ def define_node(
             )
 
             input1, input2 = inputs
-            attr = ts.TosaSerializerAttribute()
-            attr.AddAttribute()
+
             # FP lowering
             self._serialize_operator(
                 node,
                 tosa_graph,
-                ts.Op.ADD,
+                ts.TosaOp.Op().ADD,
                 [input1.name, input2.name],
                 [output.name],
-                attr,
+                None,
             )

backends/arm/operators/op_amax.py

@@ -4,7 +4,7 @@
 # LICENSE file in the root directory of this source tree.
 from typing import Any, List
 
-import tosa_serializer as ts
+import serializer.tosa_serializer as ts
 
 from executorch.backends.arm._passes.arm_pass_utils import get_first_fake_tensor
 from executorch.backends.arm.operators.node_visitor import (
@@ -60,12 +60,11 @@ def define_node(
             )
 
         attr = ts.TosaSerializerAttribute()
-        nan_mode = ts.NanPropagationMode.PROPAGATE
-        attr.ReduceMaxAttribute(axis=input.dim_order.index(dim), nan_mode=nan_mode)
+        attr.ReduceMaxAttribute(axis=input.dim_order.index(dim), nan_mode=1)
         self._serialize_operator(
             node,
             tosa_graph,
-            ts.Op.REDUCE_MAX,
+            ts.TosaOp.Op().REDUCE_MAX,
             [input.name],
             [output.name],
             attr,

backends/arm/operators/op_amin.py

@@ -4,7 +4,7 @@
 # LICENSE file in the root directory of this source tree.
 from typing import Any, List
 
-import tosa_serializer as ts
+import serializer.tosa_serializer as ts
 
 from executorch.backends.arm._passes.arm_pass_utils import get_first_fake_tensor
 from executorch.backends.arm.operators.node_visitor import (
@@ -60,13 +60,11 @@ def define_node(
             )
 
         attr = ts.TosaSerializerAttribute()
-        attr.ReduceMinAttribute(
-            axis=input.dim_order.index(dim), nan_mode=ts.NanPropagationMode.PROPAGATE
-        )
+        attr.ReduceMinAttribute(axis=input.dim_order.index(dim), nan_mode=1)
         self._serialize_operator(
             node,
             tosa_graph,
-            ts.Op.REDUCE_MIN,
+            ts.TosaOp.Op().REDUCE_MIN,
             [input.name],
             [output.name],
             attr,

backends/arm/operators/op_any.py

@@ -6,7 +6,7 @@
 # pyre-unsafe
 from typing import Any, cast, List
 
-import tosa_serializer as ts
+import serializer.tosa_serializer as ts
 
 from executorch.backends.arm.operators.node_visitor import (  # type: ignore
     NodeVisitor,
@@ -55,7 +55,7 @@ def define_node(
         self._serialize_operator(
             node,
             tosa_graph,
-            ts.Op.REDUCE_ANY,
+            ts.TosaOp.Op().REDUCE_ANY,
             [inputs[0].name],
             [output.name],
             attr,