Add selective build support for prim ops

Summary:
This diff implements selective build functionality for primitive operations (prim ops) in ExecutorTorch, allowing users to include only specific prim ops in their builds to reduce binary size and compilation time.

## Key Changes:

1. **Conditional compilation in register_prim_ops.cpp**: Wrapped each of the prim op registrations with conditional compilation macros that check both selective build enablement (`EXECUTORCH_ENABLE_PRIM_OPS_SELECTIVE_BUILD`) and individual op selection (e.g., `INCLUDE_EXECUTORCH_PRIM_ET_VIEW_DEFAULT`).

2. **Code generation tool**: Added `gen_selected_prim_ops.py` that takes comma-separated prim op names and generates a header file (`selected_prim_ops.h`) containing appropriate `#define` statements for selected ops. The tool normalizes op names to macro-safe format (e.g., `executorch_prim::et_view.default` → `INCLUDE_EXECUTORCH_PRIM_ET_VIEW_DEFAULT`).

3. **Build system integration**:
  In order to make et_operator_library also handle prim of selective build we
  make a few changes.
  1. Extract prim ops in et_operator_library
  2. Similar to gen_op_list, we invoke script that geneates selected_prim_ops.h
  file per et_operator_library target. Thus et_operator_library now generates
  selected_operators.yaml and selected_prim_ops.h.
     Note that in order to make these work we have to allow et_operator_libray to
     handle the following cases.
     1. All ops are aten ops
     2. All ops are prim ops
     3. Mix
     To do this we must make sure that the genrule continues to produce the file
     it says it will produce. In the case of 1 we have to produce empty
     selected_prim_opsh. and in case 2 we have to produce emtpy
     selected_operators.yaml
  3. In gen_all_oplist we allow for empty selected_operators.yaml and skip the
  file.
  4. Similar to gen_all_oplist we introduce another binary that combines all
  selected_prim_ops.h.
  5. Then in executorch_generated_lib we query targets from 4 that have
  selected_prim_ops and use those to compile register_prim_ops.cpp.
  In executorch_generate_lib we introduce include_all_prim_ops which by default
  is True. Hence if one wants to enable selective build for prim ops one must
  turn off that flag

## Usage:

Users can now specify prim ops like:
```
et_operator_library(name="my_aten_prim_ops", ops=["aten::mul.out", "executorch_prim::et_view.default", "aten::sym_size.int"])
executorch_generated_lib(name="my_lib", deps=[":my_aten_prim_ops"] +
other_deps, include_all_prim_ops=False)
```

Reviewed By: ivayloen, larryliu0820

Differential Revision: D81648030

Author: kimishpatel

codegen/tools/combine_prim_ops_headers.py

@@ -0,0 +1,145 @@
+#!/usr/bin/env python3
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+Script to combine multiple selected_prim_ops.h header files into a single header.
+This is used by selected_prim_operators_genrule to merge prim ops headers from dependencies.
+"""
+
+import argparse
+import os
+import sys
+from pathlib import Path
+from typing import List, Set
+
+
+def read_header_file(file_path: Path) -> Set[str]:
+    """
+    Read a selected_prim_ops.h file and extract the macros and comments.
+
+    Args:
+        file_path: Path to the header file
+
+    Returns:
+        macros_set where macros_set contains unique macro defines
+    """
+    macros = set()
+
+    try:
+        with open(file_path, 'r') as f:
+            for line in f:
+                line = line.strip()
+
+                # Extract #define statements for prim ops
+                if line.startswith('#define INCLUDE_') and not line.startswith('#define EXECUTORCH_ENABLE'):
+                    macros.add(line)
+    except FileNotFoundError:
+        print(f"Warning: Header file not found: {file_path}", file=sys.stderr)
+    except Exception as e:
+        print(f"Error reading {file_path}: {e}", file=sys.stderr)
+
+    return macros
+
+
+def combine_prim_ops_headers(header_file_paths: List[str], output_path: str) -> None:
+    """
+    Combine multiple selected_prim_ops.h files into a single header.
+
+    Args:
+        header_files: List of paths to header files to combine
+        output_path: Path to output the combined header
+    """
+    all_macros = set()
+    has_selective_build = False
+
+    # Read all header files and collect unique macros
+    for header_file_path in header_file_paths:
+        header_file = Path(header_file_path) / 'selected_prim_ops.h'
+        if os.path.exists(header_file):
+            macros = read_header_file(header_file)
+            all_macros.update(macros)
+            if len(all_macros) > 0:
+                has_selective_build = True
+        else:
+            print(f"Warning: Header file does not exist: {header_file}", file=sys.stderr)
+
+    # Generate combined header
+    header_content = [
+        "// Combined header for selective prim ops build",
+        "// This file is auto-generated by combining multiple selected_prim_ops.h files",
+        "// Do not edit manually.",
+        "",
+        "#pragma once",
+        "",
+    ]
+
+    if all_macros and has_selective_build:
+        header_content.extend([
+            "// Enable selective build for prim ops",
+            "#define EXECUTORCH_ENABLE_PRIM_OPS_SELECTIVE_BUILD",
+            "",
+            "// Combined prim ops macros from all dependencies",
+        ])
+
+        # Sort macros for deterministic output
+        sorted_macros = sorted(all_macros)
+        header_content.extend(sorted_macros)
+    else:
+        header_content.extend([
+            "// No prim ops found in dependencies - all prim ops will be included",
+            "// Selective build is disabled",
+        ])
+
+    header_content.append("")
+
+    # Write the combined header
+    os.makedirs(os.path.dirname(output_path), exist_ok=True)
+    with open(output_path, 'w') as f:
+        f.write('\n'.join(header_content))
+
+def _get_header_file_paths_from_query_output(query_output_file: str) -> List[str]:
+    """
+    Parse the output of a Buck query command to extract header file paths.
+
+    Args:
+        query_output_file: Path to the file containing the query output
+
+    Returns:
+        List of header file paths
+    """
+    header_file_paths = []
+    assert query_output_file[0] == '@', "query_output_file is not a valid file path, or it doesn't start with '@'."
+    query_output_file = query_output_file[1:]
+
+    with open(query_output_file, 'r') as f:
+        for line in f:
+            # Extract the header file path from the query output
+            header_file_paths += line.split()
+    return header_file_paths
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Combine multiple selected_prim_ops.h header files')
+    parser.add_argument('--header_files', required=True, help='Comma-separated list of header file paths')
+    parser.add_argument('--output_dir', required=True, help='Output directory for combined header')
+
+    args = parser.parse_args()
+    import os
+    header_file_paths = _get_header_file_paths_from_query_output(args.header_files)
+
+    if not header_file_paths:
+        print("Error: No header files provided", file=sys.stderr)
+        sys.exit(1)
+
+    # Generate output path
+    output_path = os.path.join(args.output_dir, 'selected_prim_ops.h')
+
+    combine_prim_ops_headers(header_file_paths, output_path)
+
+
+if __name__ == '__main__':
+    main()

codegen/tools/gen_all_oplist.py

@@ -10,7 +10,7 @@
 import sys
 from functools import reduce
 from pathlib import Path
-from typing import Any, List
+from typing import Any, Dict, List
 
 import yaml
 from torchgen.selective_build.selector import (
@@ -71,6 +71,16 @@ def _raise_if_check_prim_ops_fail(options):
             )
             raise Exception(error)
 
+def _selected_ops_model_dict_is_empty(model_dict: Dict[str, Any]) -> bool:
+    return (
+        not model_dict.get("build_features", []) and
+        not model_dict.get("custom_classes", []) and
+        not model_dict.get("et_kernel_metadata", None) and
+        not model_dict.get("include_all_non_op_selectives", False) and
+        not model_dict.get("include_all_operators", False) and
+        not model_dict.get("kernel_metadata", {}) and
+        not model_dict.get("operators", {})
+    )
 
 def main(argv: List[Any]) -> None:
     """This binary generates 3 files:
@@ -171,6 +181,11 @@ def main(argv: List[Any]) -> None:
                 ), f"{model_file_name} is not a valid file path. This is likely a BUCK issue."
                 with open(model_file_name, "rb") as model_file:
                     model_dict = yaml.safe_load(model_file)
+                    # It is possible that we created an empty yaml file.
+                    # This is because et_operator_library may only contain prim ops.
+                    # In that case selected_operators.yaml will be empty.
+                    if _selected_ops_model_dict_is_empty(model_dict):
+                        continue
                     resolved = resolve_model_file_path_to_buck_target(model_file_name)
                     for op in model_dict["operators"]:
                         model_dict["operators"][op]["debug_info"] = [resolved]

codegen/tools/gen_oplist.py

@@ -10,6 +10,7 @@
 import sys
 from enum import IntEnum
 from typing import Any, Dict, List, Optional, Set
+from pathlib import Path
 
 import yaml
 
@@ -219,13 +220,16 @@ def gen_oplist(
     ops_dict: Optional[str] = None,
     include_all_operators: bool = False,
 ):
-    assert (
+    if not (
         model_file_path
         or ops_schema_yaml_path
         or root_ops
         or ops_dict
         or include_all_operators
-    ), "Need to provide either model_file_path or ops_schema_yaml_path or root_ops or ops_dict or include_all_operators."
+    ):
+        # dump empty yaml file
+        _dump_yaml([], output_path)
+        return
 
     assert output_path, "Need to provide output_path for dumped yaml file."
     op_set = set()
@@ -326,9 +330,10 @@ def main(args: List[Any]) -> None:
     )
     options = parser.parse_args(args)
 
+    output_path = Path(options.output_path) / "selected_operators.yaml"
     try:
         gen_oplist(
-            output_path=options.output_path,
+            output_path=output_path,
             model_file_path=options.model_file_path,
             ops_schema_yaml_path=options.ops_schema_yaml_path,
             root_ops=options.root_ops,

codegen/tools/gen_selected_prim_ops.py

@@ -0,0 +1,95 @@
+#!/usr/bin/env fbpython
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-unsafe
+
+import argparse
+import os
+import sys
+from typing import Any, List
+
+from torchgen.code_template import CodeTemplate
+
+
+selected_prim_ops_h_template_str = """#pragma once
+/**
+ * Generated by executorch/codegen/tools/gen_selected_prim_ops.py
+ */
+
+$defines
+"""
+selected_prim_ops_h_template = CodeTemplate(selected_prim_ops_h_template_str)
+
+
+def normalize_op_name(op_name: str) -> str:
+    """
+    Normalize an operator name to a macro-safe format.
+    Convert op names like "executorch_prim::et_view.default" to "EXECUTORCH_PRIM_ET_VIEW_DEFAULT"
+    or "aten::sym_size.int" to "ATEN_SYM_SIZE_INT"
+    """
+    # Remove namespace separator and replace with underscore
+    normalized = op_name.replace("::", "_")
+    # Replace dots with underscores
+    normalized = normalized.replace(".", "_")
+    # Convert to uppercase
+    normalized = normalized.upper()
+    # Add INCLUDE_ prefix
+    normalized = f"INCLUDE_{normalized}"
+    return normalized
+
+
+def write_selected_prim_ops(prim_op_names: List[str], output_dir: str) -> None:
+    """
+    Generate selected_prim_ops.h from a list of prim op names.
+
+    Args:
+        prim_op_names: List of prim op names like ["executorch_prim::et_view.default", "aten::sym_size.int"]
+        output_dir: Directory where to write selected_prim_ops.h
+    """
+    # Generate #define statements for each op
+    defines = []
+    for op_name in prim_op_names:
+        macro_name = normalize_op_name(op_name)
+        defines.append(f"#define {macro_name}")
+
+    # Join all defines with newlines
+    defines_str = "\n".join(defines)
+
+    # Generate header content
+    header_contents = selected_prim_ops_h_template.substitute(defines=defines_str)
+
+    # Write to file
+    selected_prim_ops_path = os.path.join(output_dir, "selected_prim_ops.h")
+    with open(selected_prim_ops_path, "wb") as out_file:
+        out_file.write(header_contents.encode("utf-8"))
+
+
+def main(argv: List[Any]) -> None:
+    parser = argparse.ArgumentParser(description="Generate selected prim ops header")
+    parser.add_argument(
+        "--prim-op-names",
+        "--prim_op_names",
+        help="Comma-separated list of prim op names to include",
+        required=True,
+    )
+    parser.add_argument(
+        "--output-dir",
+        "--output_dir",
+        help="The directory to store the output header file (selected_prim_ops.h)",
+        required=True,
+    )
+
+    options = parser.parse_args(argv)
+
+    # Parse comma-separated prim op names
+    prim_op_names = [name.strip() for name in options.prim_op_names.split(",") if name.strip()]
+
+    write_selected_prim_ops(prim_op_names, options.output_dir)
+
+
+if __name__ == "__main__":
+    main(sys.argv[1:])

codegen/tools/targets.bzl

@@ -103,6 +103,26 @@ def define_common_targets(is_fbcode = False):
         _is_external_target = True,
     )
 
+    runtime.python_library(
+        name = "combine_prim_ops_headers_lib",
+        srcs = ["combine_prim_ops_headers.py"],
+        base_module = "executorch.codegen.tools",
+        visibility = ["//executorch/..."],
+    )
+
+    runtime.python_binary(
+        name = "combine_prim_ops_headers",
+        main_module = "executorch.codegen.tools.combine_prim_ops_headers",
+        package_style = "inplace",
+        visibility = [
+            "PUBLIC",
+        ],
+        deps = [
+            ":combine_prim_ops_headers_lib",
+        ],
+        _is_external_target = True,
+    )
+
     runtime.python_test(
         name = "test_gen_all_oplist",
         srcs = [
@@ -155,6 +175,27 @@ def define_common_targets(is_fbcode = False):
         _is_external_target = True,
     )
 
+    runtime.python_library(
+        name = "gen_selected_prim_ops_lib",
+        srcs = ["gen_selected_prim_ops.py"],
+        base_module = "executorch.codegen.tools",
+        visibility = ["//executorch/..."],
+        external_deps = ["torchgen"],
+    )
+
+    runtime.python_binary(
+        name = "gen_selected_prim_ops",
+        main_module = "executorch.codegen.tools.gen_selected_prim_ops",
+        package_style = "inplace",
+        visibility = [
+            "PUBLIC",
+        ],
+        deps = [
+            ":gen_selected_prim_ops_lib",
+        ],
+        _is_external_target = True,
+    )
+
     if not runtime.is_oss:
         runtime.cxx_python_extension(
             name = "selective_build",