File size: 14,040 Bytes
c61ccee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
# mypy: ignore-errors

import collections
from enum import Enum
from typing import Any, Callable, Dict, List

from .. import variables
from ..current_scope_id import current_scope_id
from ..exc import unimplemented
from ..source import AttrSource, Source
from ..utils import identity, istype


class MutableLocalSource(Enum):
    """

    If the VariableTracker.mutable_local represents a Variable that:

    - already existed that Dynamo began tracking while introspection (Existing)

    - is a new variable that is created during Dynamo introspection (Local)

    """

    Existing = 0
    Local = 1


class ParentsTracker:
    """

    This is a perf optimization to limit the number of objects we need to visit in tx.replace_all.

    This must be a seperate object so that it is not cloned in apply.

    """

    def __init__(self):
        # logically this is a set, but we use a dict to ensure deterministic ordering
        self.parents: Dict[ParentsTracker, bool] = dict()

    def add(self, parent):
        self.parents[parent] = True

    def recursive_parents(self):
        rv = dict(self.parents)
        worklist = list(self.parents)
        while worklist:
            for parent in worklist.pop().parents:
                if parent not in rv:
                    assert isinstance(parent, ParentsTracker)
                    rv[parent] = True
                    worklist.append(parent)
        return rv.keys()


class MutableLocalBase:
    """

    Base class for Variable.mutable_local

    """

    def __init__(self, typ: MutableLocalSource):
        # In HigherOrderOperator tracing, we need to distinguish
        # between MutableLocals inside the HigherOrderOperator and
        # ones outside it. For example, it is not safe to mutate
        # `a` in the following example because it was constructed
        # in a different scope.
        #
        # def f(x):
        #     a = 1
        #     def g(x):
        #         nonlocal a
        #         a = 2
        #         return x
        #     return wrap(g, x) + a
        #
        # We use self.scope to distinguish this.
        # scope == 0: The object was an existing variable
        # scope == 1: The object was created while Dynamo
        #             was introspecting a function
        #             (and no HigherOrderOps were involved)
        # scope >= 2: The object was created through
        #             Dynamo introspection of a HigherOrderOp.
        #             The exact number corresponds to the level
        #             of nested HigherOrderOps.
        if typ is MutableLocalSource.Existing:
            self.scope = 0
        elif typ is MutableLocalSource.Local:
            self.scope = current_scope_id()
        else:
            unimplemented(f"Unsupported MutableLocalSource: {typ}")


class MutableLocal(MutableLocalBase):
    """

    Marker used to indicate this (list, iter, etc) was constructed in

    local scope and can be mutated safely in analysis without leaking

    state.

    """

    def __init__(self):
        super().__init__(MutableLocalSource.Local)

    def __hash__(self):
        return id(self)

    def __eq__(self, other):
        return self is other


def _is_top_level_scope(scope_id):
    return scope_id == 1


def is_side_effect_safe(m: MutableLocalBase):
    scope_id = current_scope_id()

    # In the top-level scope (if no HigherOrderOperators are involved),
    # we are allowed to modify variables created in this scope as well
    # as existing variables.
    if _is_top_level_scope(scope_id):
        return True
    # Otherwise, only allow local mutation of variables created in the current scope
    return m.scope == scope_id


class VariableTrackerMeta(type):
    def __call__(cls, *args, **kwargs):
        """Call __post_init__"""
        obj = type.__call__(cls, *args, **kwargs)
        obj.__post_init__(*args, **kwargs)
        return obj

    def __instancecheck__(cls, instance) -> bool:
        """Make isinstance work with LazyVariableTracker"""
        if type.__instancecheck__(
            variables.LazyVariableTracker, instance
        ) and cls not in (
            VariableTracker,
            variables.LazyVariableTracker,
        ):
            instance = instance.realize()
        return type.__instancecheck__(cls, instance)


class VariableTracker(metaclass=VariableTrackerMeta):
    """

    Base class for tracked locals and stack values



    VariableTracker instances are immutable and should be copied in

    order to change them.

    """

    # fields to leave unmodified in apply()
    _nonvar_fields = {
        "value",
        "guards",
        "source",
        "mutable_local",
        "parents_tracker",
        "user_code_variable_name",
    }

    def clone(self, **kwargs):
        """Shallow copy with some (optional) changes"""
        args = dict(self.__dict__)
        args.update(kwargs)
        return self.__class__(**args)

    @classmethod
    def copy(cls, value):
        """Deeper (but not full) copy, leaving FX and user objects alone"""
        return cls.apply(identity, value)

    @classmethod
    def apply(

        cls,

        fn: Callable[["VariableTracker"], "VariableTracker"],

        value,

        cache=None,

        skip_fn=lambda _: False,  # Whether we should skip applying to this var

    ):
        """

        Walk this object and call fn on all the VariableTracker

        instances

        """
        if cache is None:
            cache = dict()

        idx = id(value)
        if idx in cache:
            return cache[idx][0]

        if isinstance(value, VariableTracker):
            if not skip_fn(value):

                def update_object_dict(v):
                    changed = False
                    rv = v.__dict__
                    for key in rv.keys():
                        if key not in v._nonvar_fields:
                            prior = rv[key]
                            rv[key] = cls.apply(fn, prior, cache, skip_fn)
                            changed = changed or prior is not rv[key]

                    return v

                value = value.unwrap()
                was_realized = value.is_realized()
                result = fn(update_object_dict(value))
                if not was_realized and value.is_realized():
                    # running fn() resulted in value getting realized,
                    # which means we missed updating the contents of result
                    result = update_object_dict(result.unwrap())
            else:
                result = fn(value)
                if result is not None:
                    result = result.unwrap()
        elif istype(value, list):
            result = [cls.apply(fn, v, cache, skip_fn) for v in value]
        elif istype(value, tuple):
            result = tuple(cls.apply(fn, v, cache, skip_fn) for v in value)
        elif istype(value, (dict, collections.OrderedDict)):
            result = {
                k: cls.apply(fn, v, cache, skip_fn) for k, v in list(value.items())
            }
        else:
            result = value

        # save `value` to keep it alive and ensure id() isn't reused
        cache[idx] = (result, value)
        return result

    def __repr__(self):
        return f"{self.__class__.__name__}()"

    def python_type(self):
        """

        Abstract method to be implemented by subclasses of VariableTracker.



        This method should return the type represented by the instance of the subclass.

        The purpose is to provide a standardized way to retrieve the Python type information

        of the variable being tracked.



        Returns:

            type: The Python type (such as int, str, list, etc.) of the variable tracked by

                the subclass. If the type cannot be determined or is not relevant,

                leaving it undefined or invoking super() is always sound.



        Note:

            This is an abstract method and may be overridden in subclasses.



        Example:

            class SetVariable(VariableTracker):

                def python_type(self):

                    return set



        Raises:

            NotImplementedError: If the method is not implemented in a subclass.

        """
        raise NotImplementedError(f"{self} has no type")

    def as_python_constant(self):
        """For constants"""
        raise NotImplementedError(f"{self} is not a constant")

    def guard_as_python_constant(self):
        """Similar to as_python_constant(), but add ID_MATCH guards to try to force things to become constants"""
        try:
            return self.as_python_constant()
        except NotImplementedError as e:
            unimplemented(str(e))

    def is_python_constant(self):
        try:
            self.as_python_constant()
            return True
        except NotImplementedError:
            return False

    def make_guard(self, fn):
        if self.source:
            return self.source.make_guard(fn)
        raise NotImplementedError()

    def const_getattr(self, tx, name: str) -> Any:
        """getattr(self, name) returning a python constant"""
        raise NotImplementedError()

    def var_getattr(self, tx, name: str) -> "VariableTracker":
        """getattr(self, name) returning a new variable"""
        value = self.const_getattr(tx, name)
        if not variables.ConstantVariable.is_literal(value):
            raise NotImplementedError()
        source = None
        if self.source:
            source = AttrSource(self.source, name)
        return variables.ConstantVariable.create(value, source=source)

    def is_proxy(self):
        try:
            self.as_proxy()
            return True
        except NotImplementedError:
            return False

    def as_proxy(self):
        raise NotImplementedError(str(self))

    def maybe_fx_node(self):
        try:
            proxy = self.as_proxy()
            import torch.fx

            if isinstance(proxy, torch.fx.Proxy):
                return proxy.node
            return None
        except NotImplementedError:
            return None

    def reconstruct(self, codegen):
        raise NotImplementedError()

    def can_reconstruct(self, tx):
        """If it is possible to reconstruct the Python object this

        VariableTracker represents."""
        assert tx is tx.output.root_tx, "Only root tx can reconstruct"
        try:
            from ..codegen import PyCodegen

            cg = PyCodegen(tx)
            self.reconstruct(cg)
            return True
        except NotImplementedError:
            return False

    def unpack_var_sequence(self, tx) -> List["VariableTracker"]:
        raise NotImplementedError()

    def has_unpack_var_sequence(self, tx) -> bool:
        try:
            self.unpack_var_sequence(tx)
            return True
        except NotImplementedError:
            return False

    def inspect_parameter_names(self) -> List[str]:
        unimplemented(f"inspect_parameter_names: {self}")

    def call_hasattr(self, tx, name: str) -> "VariableTracker":
        unimplemented(f"hasattr {self.__class__.__name__} {name}")

    def call_function(

        self, tx, args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]"

    ) -> "VariableTracker":
        unimplemented(f"call_function {self} {args} {kwargs}")

    def call_method(

        self,

        tx,

        name,

        args: "List[VariableTracker]",

        kwargs: "Dict[str, VariableTracker]",

    ) -> "VariableTracker":
        if name == "__len__" and self.has_unpack_var_sequence(tx):
            assert not (args or kwargs)
            return variables.ConstantVariable.create(len(self.unpack_var_sequence(tx)))
        elif (
            name == "__getattr__"
            and len(args) == 1
            and args[0].is_python_constant()
            and not kwargs
        ):
            return self.var_getattr(tx, args[0].as_python_constant())
        raise unimplemented(f"call_method {self} {name} {args} {kwargs}")

    def rename(self, tx, name):
        return self

    def realize(self) -> "VariableTracker":
        """Used by LazyVariableTracker to build the real VariableTracker"""
        return self

    def recursive_realize(self):
        """Realize all objects under this"""
        return VariableTracker.apply(lambda x: x.realize(), self)

    def unwrap(self) -> "VariableTracker":
        """Used by LazyVariableTracker to return the real VariableTracker if it already exists"""
        return self

    def is_realized(self):
        """Used by LazyVariableTracker to indicate an unrealized node"""
        return True

    def __init__(

        self,

        *,

        source: Source = None,

        mutable_local: MutableLocal = None,

        parents_tracker: ParentsTracker = None,

    ):
        super().__init__()
        self.source = source
        self.mutable_local = mutable_local
        self.parents_tracker = parents_tracker

    def __post_init__(self, *args, **kwargs):
        if self.parents_tracker is None:
            self.parents_tracker = ParentsTracker()
        # visit children 1 level deep and ensure parent is set properly
        VariableTracker.apply(
            lambda node: node.parents_tracker.add(self.parents_tracker),
            [v for k, v in self.__dict__.items() if k not in self._nonvar_fields],
            skip_fn=lambda _: True,
        )


def typestr(*objs):
    if len(objs) == 1:
        (obj,) = objs
        if isinstance(obj, VariableTracker):
            return str(obj)
        else:
            return type(obj).__name__
    else:
        return " ".join(map(typestr, objs))