forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
script_type_parser.cpp
488 lines (445 loc) · 17 KB
/
script_type_parser.cpp
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
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
#include <torch/csrc/jit/frontend/script_type_parser.h>
#include <ATen/core/type_factory.h>
#include <torch/csrc/jit/frontend/parser.h>
#include <torch/csrc/jit/ir/ir.h>
#include <torch/custom_class.h>
namespace torch {
namespace jit {
namespace {
bool isTorch(const Expr& expr) {
return expr.kind() == TK_VAR && Var(expr).name().name() == "torch";
}
std::string collectQualname(const Select& select) {
Expr base = select.value();
if (base.kind() == TK_VAR) {
return Var(base).name().name() + "." + select.selector().name();
}
std::string basename = collectQualname(Select(base));
return basename + "." + select.selector().name();
}
const std::unordered_map<std::string, c10::TypePtr>& string_to_type_lut() {
return c10::DefaultTypeFactory::basePythonTypes();
}
} // namespace
TypePtr ScriptTypeParser::subscriptToType(
const std::string& typeName,
const Subscript& subscript) const {
if (typeName == "Tuple" || typeName == "tuple") {
if (subscript.subscript_exprs().size() == 1 &&
subscript.subscript_exprs()[0].kind() == TK_TUPLE_LITERAL) {
// `typing.Tuple` special cases syntax for empty tuple annotations,
// i.e. `typing.Tuple[()]`. Allow for parsing an empty tuple literal
// here. See https://docs.python.org/3/library/typing.html#typing.Tuple
auto tup_literal = TupleLiteral(subscript.subscript_exprs()[0]);
if (tup_literal.inputs().size() > 0) {
throw ErrorReport(tup_literal.range())
<< "Tuple literal in Tuple type annotation must not "
<< "have any elements!";
}
return TupleType::create({});
}
std::vector<TypePtr> subscript_expr_types;
for (auto expr : subscript.subscript_exprs()) {
subscript_expr_types.emplace_back(parseTypeFromExprImpl(expr));
}
return TupleType::create(subscript_expr_types);
} else if (typeName == "List" || typeName == "list") {
if (subscript.subscript_exprs().size() != 1) {
throw ErrorReport(subscript)
<< " expected exactly one element type but found "
<< subscript.subscript_exprs().size();
}
auto elem_type =
parseTypeFromExprImpl(*subscript.subscript_exprs().begin());
return ListType::create(elem_type);
} else if (typeName == "Optional") {
if (subscript.subscript_exprs().size() != 1) {
throw ErrorReport(subscript)
<< " expected exactly one element type but found "
<< subscript.subscript_exprs().size();
}
auto elem_type =
parseTypeFromExprImpl(*subscript.subscript_exprs().begin());
return OptionalType::create(elem_type);
} else if (typeName == "Union") {
std::vector<TypePtr> subscript_expr_types;
subscript_expr_types.reserve(subscript.subscript_exprs().size());
for (auto expr : subscript.subscript_exprs()) {
subscript_expr_types.emplace_back(parseTypeFromExprImpl(expr));
}
return UnionType::create(subscript_expr_types);
} else if (typeName == "Future" || typeName == "torch.jit.Future") {
if (subscript.subscript_exprs().size() != 1) {
throw ErrorReport(subscript)
<< " expected exactly one element type but found "
<< subscript.subscript_exprs().size();
}
auto elem_type =
parseTypeFromExprImpl(*subscript.subscript_exprs().begin());
return FutureType::create(elem_type);
} else if (typeName == "RRef") {
if (subscript.subscript_exprs().size() != 1) {
throw ErrorReport(subscript)
<< " expected exactly one element type but found "
<< subscript.subscript_exprs().size();
}
auto elem_type =
parseTypeFromExprImpl(*subscript.subscript_exprs().begin());
return RRefType::create(elem_type);
} else if (typeName == "Dict" || typeName == "dict") {
if (subscript.subscript_exprs().size() != 2) {
throw ErrorReport(subscript)
<< " expected exactly 2 element types but found "
<< subscript.subscript_exprs().size();
}
auto key_type = parseTypeFromExprImpl(subscript.subscript_exprs()[0]);
auto value_type = parseTypeFromExprImpl(subscript.subscript_exprs()[1]);
return DictType::create(key_type, value_type);
} else {
throw ErrorReport(subscript.range())
<< "Unknown type constructor " << typeName;
}
}
c10::optional<std::pair<TypePtr, int32_t>> ScriptTypeParser::parseBroadcastList(
const Expr& expr) const {
// Alias torch.nn._common_types._size_?_t to BroadcastingList?[int]
if (expr.kind() == TK_VAR) {
auto var = Var(expr);
auto& name = var.name().name();
constexpr auto _size_prefix = "_size_";
constexpr auto _size_suffix = "_t";
constexpr auto _size_n_len = 9; // strlen("_size_X_t")
constexpr auto _size_prefix_len = 6; // strlen("_size_");
if (name.find(_size_prefix) == 0 && name.length() == _size_n_len &&
name.find(_size_suffix) == _size_prefix_len + 1 &&
::isdigit(name[_size_prefix_len])) {
int n = name[_size_prefix_len] - '0';
return std::pair<TypePtr, int32_t>(ListType::create(IntType::get()), n);
}
}
if (expr.kind() != TK_SUBSCRIPT)
return c10::nullopt;
auto subscript = Subscript(expr);
if (subscript.value().kind() != TK_VAR)
return c10::nullopt;
auto var = Var(subscript.value());
auto subscript_exprs = subscript.subscript_exprs();
// handle the case where the BroadcastingList is wrapped in a Optional type
if (var.name().name() == "Optional") {
auto broadcast_list = parseBroadcastList(subscript_exprs[0]);
if (broadcast_list) {
TypePtr opt_type = OptionalType::create(broadcast_list->first);
return std::pair<TypePtr, int32_t>(opt_type, broadcast_list->second);
} else {
return c10::nullopt;
}
} else if (var.name().name().find("BroadcastingList") != 0) {
return c10::nullopt;
}
if (subscript_exprs.size() != 1)
throw ErrorReport(subscript.subscript_exprs().range())
<< "BroadcastingList/Optional[BroadcastingList] "
"must be subscripted with a type";
auto typ = subscript_exprs[0];
auto len = var.name().name().substr(strlen("BroadcastingList"));
if (typ.kind() != TK_VAR)
throw ErrorReport(subscript.value().range())
<< "Subscripted type must be a type identifier";
auto value_name = Var(typ).name().name();
if (value_name != "float" && value_name != "int")
throw ErrorReport(subscript.value().range())
<< "Broadcastable lists only supported for int or float";
auto elem_ptr = string_to_type_lut().find(value_name);
AT_ASSERT(elem_ptr != string_to_type_lut().end());
TypePtr list_ptr = ListType::create(elem_ptr->second);
const char* len_c = len.c_str();
// NOLINTNEXTLINE(cppcoreguidelines-init-variables)
char* end;
size_t len_v = strtoull(len_c, &end, 10);
if (end != len_c + len.size()) {
throw ErrorReport(subscript.subscript_exprs().range())
<< "subscript of Broadcastable list must be a positive integer";
}
return std::pair<TypePtr, int32_t>(list_ptr, len_v);
}
// gets the base type name given namespaces where the types live
// turns torch.Tensor -> Tensor, X -> X
c10::optional<std::string> ScriptTypeParser::parseBaseTypeName(
const Expr& expr) const {
switch (expr.kind()) {
case TK_VAR: {
return Var(expr).name().name();
}
case TK_NONE: {
return "None";
}
case TK_NONE_TYPE: {
return "NoneType";
}
case '.': {
auto select = Select(expr);
const std::string& name = select.selector().name();
// Special case for torch.Tensor and its' subclasses
const std::unordered_set<std::string> tensor_subtypes = {
"Tensor",
"LongTensor",
"FloatTensor",
"DoubleTensor",
"IntTensor",
"ShortTensor",
"HalfTensor",
"CharTensor",
"ByteTensor",
"BoolTensor"};
if (isTorch(select.value()) && tensor_subtypes.count(name) == 1) {
return name;
} else {
// Otherwise, it's a fully qualified class name
return collectQualname(select);
}
} break;
}
return at::nullopt;
}
TypePtr ScriptTypeParser::parseTypeFromExpr(const Expr& expr) const {
// the resolver needs to recursively resolve the expression, so to avoid
// resolving all type expr subtrees we only use it for the top level
// expression and base type names.
if (resolver_) {
if (auto typePtr =
resolver_->resolveType(expr.range().text().str(), expr.range())) {
return typePtr;
}
}
return parseTypeFromExprImpl(expr);
}
TypePtr ScriptTypeParser::parseTypeFromExprImpl(const Expr& expr) const {
if (expr.kind() == TK_SUBSCRIPT) {
auto subscript = Subscript(expr);
auto value_name = parseBaseTypeName(subscript.value());
if (!value_name) {
throw ErrorReport(subscript.value().range())
<< "Subscripted type must be a type identifier";
}
return subscriptToType(*value_name, subscript);
} else if (expr.kind() == TK_STRINGLITERAL) {
const auto& type_name = StringLiteral(expr).text();
// Check if the type is a custom class. This is done by checking
// if type_name starts with "torch.classes."
if (type_name.find("torch.classes.") == 0) {
auto custom_class_type = getCustomClass("__torch__." + type_name);
return custom_class_type;
}
// `torch.cuda.Stream` and `torch.cuda.Event` are aliased as
// custom classes of type torch.classes.cuda.Stream and
// torch.classes.cuda.Event respectively. Return the respective
// custom class types for these two cases.
if (type_name.find("torch.cuda.Stream") == 0) {
auto custom_class_type =
getCustomClass("__torch__.torch.classes.cuda.Stream");
return custom_class_type;
}
if (type_name.find("torch.cuda.Event") == 0) {
auto custom_class_type =
getCustomClass("__torch__.torch.classes.cuda.Event");
return custom_class_type;
}
if (resolver_) {
if (auto typePtr = resolver_->resolveType(type_name, expr.range())) {
return typePtr;
}
}
throw ErrorReport(expr) << "Unknown type name '" << type_name << "'";
} else if (auto name = parseBaseTypeName(expr)) {
auto itr = string_to_type_lut().find(*name);
if (itr != string_to_type_lut().end()) {
return itr->second;
}
if (resolver_) {
if (auto typePtr = resolver_->resolveType(*name, expr.range())) {
return typePtr;
}
}
if (auto custom_class_type = getCustomClass(*name)) {
return custom_class_type;
}
throw ErrorReport(expr) << "Unknown type name '" << *name << "'";
}
throw ErrorReport(expr.range())
<< "Expression of type " << kindToString(expr.kind())
<< " cannot be used in a type expression";
}
TypePtr ScriptTypeParser::parseType(const std::string& str) {
Parser p(std::make_shared<Source>(str));
return parseTypeFromExpr(p.parseExp());
}
std::vector<IValue> ScriptTypeParser::evaluateDefaults(
const SourceRange& r,
const std::vector<Expr>& default_types,
const std::vector<Expr>& default_exprs) {
std::vector<IValue> default_values;
if (default_exprs.empty())
return default_values;
// To evaluate the default expressions, we create a graph with no inputs,
// and whose returns are the default values we need.
// We then run constant prop on this graph and check the results are
// constant. This approach avoids having to have separate handling of
// default arguments from standard expressions by piecing together existing
// machinery for graph generation, constant propgation, and constant
// extraction.
auto tuple_type = Subscript::create(
r,
Var::create(r, Ident::create(r, "Tuple")),
List<Expr>::create(r, default_types));
auto blank_decl = Decl::create(
r, List<Param>::create(r, {}), Maybe<Expr>::create(r, tuple_type));
auto tuple_expr =
TupleLiteral::create(r, List<Expr>::create(r, default_exprs));
auto ret = Return::create(r, tuple_expr);
auto def = Def::create(
r,
Ident::create(r, "defaults"),
blank_decl,
List<Stmt>::create(r, {ret}));
CompilationUnit cu;
cu.define(
c10::nullopt,
/*properties=*/{},
/*propResolvers=*/{},
{def},
{resolver_},
nullptr);
Stack stack;
// XXX: We need to turn optimization off here because otherwise we try to
// recursively initialize stuff in DecomposeOps.
GraphOptimizerEnabledGuard guard(false);
cu.get_function(def.name().name()).run(stack);
return stack.at(0).toTupleRef().elements().vec();
}
std::vector<Argument> ScriptTypeParser::parseArgsFromDecl(
const Decl& decl,
bool skip_self) {
auto params_begin = decl.params().begin();
auto params_end = decl.params().end();
if (skip_self) {
++params_begin;
}
std::vector<Argument> retval;
std::vector<Expr> default_types;
std::vector<Expr> default_exprs;
// gather any non-empty default arguments
for (auto it = params_begin; it != params_end; ++it) {
auto param = *it;
auto def = param.defaultValue();
if (def.present()) {
if (!param.type().present()) {
// We require explicit type-hints for default expressions.
// If param doesn't have a type, we could default to "Tensor",
// just like what happens in the Python frontend.
// However here things are a bit more complicated, because
// default expressions are evaluated using a custom-built
// graph, and error messages coming out of that in case
// the type doesn't match the value are quite obscure.
throw ErrorReport(param.range())
<< "Keyword arguments with defaults need to be type-hinted (TorchScript C++ frontend)";
}
default_types.emplace_back(param.type().get());
default_exprs.emplace_back(def.get());
}
}
auto default_values =
evaluateDefaults(decl.range(), default_types, default_exprs);
auto defaults_it = default_values.begin();
for (auto it = params_begin; it != params_end; ++it) {
auto decl_arg = *it;
TypePtr type;
c10::optional<int32_t> N = c10::nullopt;
if (!decl_arg.type().present()) {
// If this param doesn't have a type, default to "tensor"
type = TensorType::getInferred();
} else {
// BroadcastList list can only appear at the argument level
Expr type_expr = decl_arg.type().get();
if (auto maybe_broad_list = parseBroadcastList(type_expr)) {
type = maybe_broad_list->first;
N = maybe_broad_list->second;
} else {
type = parseTypeFromExpr(decl_arg.type().get());
}
}
c10::optional<IValue> default_value = c10::nullopt;
if (decl_arg.defaultValue().present()) {
default_value = *defaults_it++;
}
auto arg = Argument(
decl_arg.ident().name(),
type,
N,
default_value,
decl_arg.kwarg_only(),
/*alias_info=*/c10::nullopt);
retval.push_back(arg);
}
return retval;
}
std::vector<Argument> ScriptTypeParser::parseReturnFromDecl(const Decl& decl) {
// we represent no annoation on a return type as having no values in the
// schema's return() list
// in emitReturn we take the actual return value to be the value of the
// return statement if no one was provided here
if (!decl.return_type().present())
return {};
if (parseBroadcastList(decl.return_type().get()))
throw ErrorReport(decl.return_type().range())
<< "Broadcastable lists cannot appear as a return type";
TypePtr parsed_type;
Expr type_expr = decl.return_type().get();
parsed_type = parseTypeFromExpr(type_expr);
return {Argument(
"",
parsed_type,
/*N =*/c10::nullopt,
/*default_value =*/c10::nullopt,
/*kwarg_only =*/false)};
}
FunctionSchema ScriptTypeParser::parseSchemaFromDef(
const Def& def,
bool skip_self) {
const auto name = def.name().name();
std::vector<Argument> args = parseArgsFromDecl(def.decl(), skip_self);
std::vector<Argument> returns = parseReturnFromDecl(def.decl());
return FunctionSchema(
name, "", std::move(args), std::move(returns), false, false);
}
c10::IValue ScriptTypeParser::parseClassConstant(const Assign& assign) {
if (assign.lhs().kind() != TK_VAR) {
throw ErrorReport(assign.range())
<< "Expected to a variable for class constant";
}
const auto final_type = assign.type().get();
auto expr = assign.rhs().get();
if (final_type.kind() != TK_SUBSCRIPT) {
throw ErrorReport(assign.range())
<< "Expected subscripted type for class constant";
}
auto subscript = Subscript(final_type);
auto value_name = parseBaseTypeName(subscript.value());
if (!value_name) {
throw ErrorReport(subscript.value().range())
<< "Subscripted type must be a type identifier";
}
if (*value_name != "Final") {
throw ErrorReport(subscript.range())
<< "Base type must be Final for class constant";
}
if (subscript.subscript_exprs().size() != 1) {
throw ErrorReport(subscript)
<< " expected exactly one element type but found "
<< subscript.subscript_exprs().size();
}
auto type = *subscript.subscript_exprs().begin();
auto default_val = evaluateDefaults(expr.range(), {type}, {expr});
return *default_val.begin();
}
} // namespace jit
} // namespace torch