xref: /aosp_15_r20/external/pytorch/tools/autograd/gen_annotated_fn_args.py (revision da0073e96a02ea20f0ac840b70461e3646d07c45)
1"""
2For procedural tests needed for __torch_function__, we use this function
3to export method names and signatures as needed by the tests in
4test/test_overrides.py.
5
6python -m tools.autograd.gen_annotated_fn_args \
7       aten/src/ATen/native/native_functions.yaml \
8       aten/src/ATen/native/tags.yaml \
9       $OUTPUT_DIR \
10       tools/autograd
11
12Where $OUTPUT_DIR is where you would like the files to be
13generated.  In the full build system, OUTPUT_DIR is
14torch/testing/_internal/generated
15"""
16
17from __future__ import annotations
18
19import argparse
20import os
21import textwrap
22from collections import defaultdict
23from typing import Any, Sequence, TYPE_CHECKING
24
25import torchgen.api.python as python
26from torchgen.context import with_native_function
27from torchgen.gen import parse_native_yaml
28from torchgen.utils import FileManager
29
30from .gen_python_functions import (
31    is_py_fft_function,
32    is_py_linalg_function,
33    is_py_nn_function,
34    is_py_special_function,
35    is_py_torch_function,
36    is_py_variable_method,
37    should_generate_py_binding,
38)
39
40
41if TYPE_CHECKING:
42    from torchgen.model import Argument, BaseOperatorName, NativeFunction
43
44
45def gen_annotated(
46    native_yaml_path: str, tags_yaml_path: str, out: str, autograd_dir: str
47) -> None:
48    native_functions = parse_native_yaml(
49        native_yaml_path, tags_yaml_path
50    ).native_functions
51    mappings = (
52        (is_py_torch_function, "torch._C._VariableFunctions"),
53        (is_py_nn_function, "torch._C._nn"),
54        (is_py_linalg_function, "torch._C._linalg"),
55        (is_py_special_function, "torch._C._special"),
56        (is_py_fft_function, "torch._C._fft"),
57        (is_py_variable_method, "torch.Tensor"),
58    )
59    annotated_args: list[str] = []
60    for pred, namespace in mappings:
61        groups: dict[BaseOperatorName, list[NativeFunction]] = defaultdict(list)
62        for f in native_functions:
63            if not should_generate_py_binding(f) or not pred(f):
64                continue
65            groups[f.func.name.name].append(f)
66        for group in groups.values():
67            for f in group:
68                annotated_args.append(f"{namespace}.{gen_annotated_args(f)}")
69
70    template_path = os.path.join(autograd_dir, "templates")
71    fm = FileManager(install_dir=out, template_dir=template_path, dry_run=False)
72    fm.write_with_template(
73        "annotated_fn_args.py",
74        "annotated_fn_args.py.in",
75        lambda: {
76            "annotated_args": textwrap.indent("\n".join(annotated_args), "    "),
77        },
78    )
79
80
81@with_native_function
82def gen_annotated_args(f: NativeFunction) -> str:
83    def _get_kwargs_func_exclusion_list() -> list[str]:
84        # functions that currently don't work with kwargs in test_overrides.py
85        return [
86            "diagonal",
87            "round_",
88            "round",
89            "scatter_",
90        ]
91
92    def _add_out_arg(
93        out_args: list[dict[str, Any]], args: Sequence[Argument], *, is_kwarg_only: bool
94    ) -> None:
95        for arg in args:
96            if arg.default is not None:
97                continue
98            out_arg: dict[str, Any] = {}
99            out_arg["is_kwarg_only"] = str(is_kwarg_only)
100            out_arg["name"] = arg.name
101            out_arg["simple_type"] = python.argument_type_str(
102                arg.type, simple_type=True
103            )
104            size_t = python.argument_type_size(arg.type)
105            if size_t:
106                out_arg["size"] = size_t
107            out_args.append(out_arg)
108
109    out_args: list[dict[str, Any]] = []
110    _add_out_arg(out_args, f.func.arguments.flat_positional, is_kwarg_only=False)
111    if f"{f.func.name.name}" not in _get_kwargs_func_exclusion_list():
112        _add_out_arg(out_args, f.func.arguments.flat_kwarg_only, is_kwarg_only=True)
113
114    return f"{f.func.name.name}: {repr(out_args)},"
115
116
117def main() -> None:
118    parser = argparse.ArgumentParser(description="Generate annotated_fn_args script")
119    parser.add_argument(
120        "native_functions", metavar="NATIVE", help="path to native_functions.yaml"
121    )
122    parser.add_argument("tags", metavar="TAGS", help="path to tags.yaml")
123    parser.add_argument("out", metavar="OUT", help="path to output directory")
124    parser.add_argument(
125        "autograd", metavar="AUTOGRAD", help="path to template directory"
126    )
127    args = parser.parse_args()
128    gen_annotated(args.native_functions, args.tags, args.out, args.autograd)
129
130
131if __name__ == "__main__":
132    main()
133