Use FusedLocation to hold layer metadata

Author: yizhuoz004

tripy/nvtripy/backend/mlir/compiler.py

@@ -58,6 +58,7 @@ def _get_compilation_task(self, trt_builder_opt_level):
         if config.enable_mlir_debug or config.enable_tensorrt_debug:
             opts.append("--debug=true")
             if config.enable_mlir_debug:
+                opts.append(f"--mlir-elide-elementsattrs-if-larger=32")
                 opts.append(f"--debug-only={config.mlir_debug_types}")
                 opts.append(f"--mlir-print-ir-after-all")
                 opts.append(f"--mlir-print-ir-tree-dir={config.mlir_debug_tree_path}")

tripy/nvtripy/backend/mlir/utils.py

@@ -124,8 +124,9 @@ def make_mlir_tensor(
 OUTPUT_SEPARATOR = ";;<out>;;"
 
 
-def make_tensor_location(input_names: List[str], output_names: List[str]) -> ir.Location:
-    return ir.Location.name(f"{','.join(input_names)}{OUTPUT_SEPARATOR}{','.join(output_names)}")
+def make_tensor_location(input_names: List[str], output_names: List[str], metadata: str) -> ir.Location:
+    loc = ir.Location.name(f"{','.join(input_names)}{OUTPUT_SEPARATOR}{','.join(output_names)}")
+    return ir.Location.fused([loc], ir.StringAttr.get(metadata))
 
 
 # The way locations are printed by MLIR-TRT differs from how they are printed by TRT, hence all the `?`s.

tripy/nvtripy/trace/trace.py

@@ -159,8 +159,29 @@ def to_mlir_impl():
                             layer_input_ops = [mlir_ops[inp.name] for inp in op.inputs]
                             output_types = [out.to_mlir() for out in op.outputs]
 
+                            metadata = (
+                                f"<TraceOp: {op}, Stack Info: "
+                                + ", ".join(
+                                    (
+                                        str_from_stack_info(
+                                            # include_code_index points to the first "useful" frame, i.e. usually the API
+                                            # that the user calls, or a few frames below that:
+                                            out.stack_info[out.stack_info.include_code_index or 0 :],
+                                            enable_color=False,
+                                            fetch_source_code=False,
+                                        )
+                                        if out.stack_info
+                                        else ""
+                                    ).replace("\n", " ")
+                                    for out in op.outputs
+                                )
+                                + " >, "
+                            )
+
                             with make_tensor_location(
-                                [inp.name for inp in op.inputs], [out.name for out in op.outputs]
+                                [inp.name for inp in op.inputs],
+                                [out.name for out in op.outputs],
+                                metadata,
                             ):
                                 mlir_output_ops = op.to_mlir(layer_input_ops, output_types)
 
@@ -180,25 +201,6 @@ def num_known_dims(ranked_tensor_type):
                                 if num_known_dims(output_type) >= num_known_dims(mlir_output_op.type):
                                     mlir_output_op.set_type(output_type)
 
-                                mlir_output_op.owner.attributes["metadata"] = ir.StringAttr.get(
-                                    f"<TraceOp: {op}, Stack Info: "
-                                    + ", ".join(
-                                        (
-                                            str_from_stack_info(
-                                                # include_code_index points to the first "useful" frame, i.e. usually the API
-                                                # that the user calls, or a few frames below that:
-                                                out.stack_info[out.stack_info.include_code_index or 0 :],
-                                                enable_color=False,
-                                                fetch_source_code=False,
-                                            )
-                                            if out.stack_info
-                                            else ""
-                                        ).replace("\n", " ")
-                                        for out in op.outputs
-                                    )
-                                    + " >, "
-                                )
-
                             mlir_ops.update(zip([out.name for out in op.outputs], mlir_output_ops))
 
                         func_dialect.ReturnOp([mlir_ops[o.name] for o in self.outputs])