def argumenttype_type(t: Type,
*,
mutable: bool,
binds: ArgName,
remove_non_owning_ref_types: bool = False) -> NamedCType:
# If it's a value type, do the value type translation
r = valuetype_type(t,
binds=binds,
remove_non_owning_ref_types=remove_non_owning_ref_types)
if r is not None:
return r
if isinstance(t, BaseType):
if t.name == BaseTy.Tensor:
if mutable and not local.use_const_ref_for_mutable_tensors():
return NamedCType(binds, MutRefCType(BaseCType(tensorT)))
else:
return NamedCType(binds, ConstRefCType(BaseCType(tensorT)))
elif t.name == BaseTy.Scalar:
return NamedCType(binds, ConstRefCType(BaseCType(scalarT)))
else:
raise AssertionError(f"base type should have been value type {t}")
elif isinstance(t, OptionalType):
if str(t.elem) == "Tensor":
if mutable and not local.use_const_ref_for_mutable_tensors():
return NamedCType(binds, MutRefCType(
BaseCType(tensorT))) # TODO: fix this discrepancy
else:
return NamedCType(
binds, ConstRefCType(OptionalCType(BaseCType(tensorT))))
elif str(t.elem) == "Scalar":
return NamedCType(binds,
ConstRefCType(OptionalCType(BaseCType(scalarT))))
elif isinstance(t.elem, ListType) and str(t.elem.elem) == "int":
return NamedCType(binds, BaseCType(optionalIntArrayRefT))
elem = argumenttype_type(t.elem, mutable=mutable, binds=binds)
return NamedCType(binds, OptionalCType(elem.type))
elif isinstance(t, ListType):
# TODO: remove these special cases, ArrayRef fallthrough works fine
if str(t.elem) == "int":
if remove_non_owning_ref_types:
return NamedCType(binds, VectorCType(BaseCType(longT)))
else:
return NamedCType(binds, BaseCType(intArrayRefT))
elif str(t.elem) == "Tensor":
return NamedCType(binds, BaseCType(tensorListT))
elif str(t.elem) == "Scalar":
return NamedCType(binds, ArrayRefCType(BaseCType(scalarT)))
elif str(t.elem) == "SymInt":
return NamedCType(binds, BaseCType(symIntArrayRefT))
elif str(t.elem) == "Dimname":
return NamedCType(binds, BaseCType(dimnameListT))
elif str(t.elem) == "Tensor?":
return NamedCType(
binds,
ConstRefCType(ListCType(OptionalCType(BaseCType(tensorT)))))
elem = argumenttype_type(t.elem, mutable=mutable, binds=binds)
return NamedCType(binds, ArrayRefCType(elem.type))
else:
raise AssertionError(f"unrecognized type {repr(t)}")
def returntype_type(t: Type, *, mutable: bool) -> CType:
# placeholder is ignored
r = valuetype_type(t, binds="__placeholder__")
if r is not None:
return r.type
if isinstance(t, BaseType):
if t.name == BaseTy.Tensor:
if mutable:
if local.use_const_ref_for_mutable_tensors():
return ConstRefCType(BaseCType(tensorT))
else:
return MutRefCType(BaseCType(tensorT))
else:
# Note [Tensor Copy Returns]
# Currently, we use "Argument.is_write" to determine
# whether or not Tensor return types should be copies or references.
# If that ever changes, take a look at other locations of this note!
return BaseCType(tensorT)
elif t.name == BaseTy.Scalar:
return BaseCType(scalarT)
elif isinstance(t, ListType):
assert (
not mutable
), "Native functions should never return a mutable tensor list. They should return void."
elem = returntype_type(t.elem, mutable=False)
assert t.size is None, f"fixed size list returns not supported: {t}"
return VectorCType(elem)
raise AssertionError(f"unrecognized return type {t}")
def argumenttype_type(t: Type, *, mutable: bool, binds: ArgName) -> NamedCType:
if str(t) == "Tensor?":
tensor_type: OptionalCType = OptionalCType(BaseCType(tensorT))
if mutable and not local.use_const_ref_for_mutable_tensors():
return NamedCType(binds, MutRefCType(tensor_type))
else:
return NamedCType(binds, ConstRefCType(tensor_type))
elif str(t) == "Tensor?[]":
return NamedCType(
binds, ConstRefCType(ListCType(OptionalCType(BaseCType(tensorT)))))
elif str(t) == "Scalar":
return NamedCType(binds, ConstRefCType(BaseCType(scalarT)))
elif str(t) == "Scalar?":
return NamedCType(binds,
ConstRefCType(OptionalCType(BaseCType(scalarT))))
return cpp.argumenttype_type(t, mutable=mutable, binds=binds)
def translate(
bindings: Sequence[Union[Expr, Binding]],
goals: Sequence[Union[NamedCType, Binding]],
*,
method: bool = False,
allow_expensive_conversions: bool = False,
) -> List[Expr]:
binding_exprs: List[Expr] = []
for b in bindings:
if isinstance(b, Binding):
binding_exprs.append(Expr(
expr=b.name,
type=b.nctype,
))
else:
binding_exprs.append(b)
goal_ctypes: List[NamedCType] = []
for g in goals:
if isinstance(g, Binding):
goal_ctypes.append(g.nctype)
else:
goal_ctypes.append(g)
# Add all the bindings to the context
ctx: Dict[NamedCType, str] = {}
for b in binding_exprs:
ctx[b.type] = b.expr
# While we're at it, do some simple forward inference, looking through
# constructors.
#
# NB: When should you do forward inference versus backward inference?
# The general idea:
#
# - Backward inference WHEN the goal gets smaller
# - Forward inference WHEN the hypothesis gets smaller
#
# This helps ensure termination: backward inference starts with a goal
# and tries to make it simpler and simpler until it's trivial; if the
# goal can grow in size, we blow up to a really huge goal size.
# Similarly, with forward inference we take hypotheses and decompose
# them into simpler hypotheses; if hypotheses could expand in size,
# we also have potential nontermination. (In the code below, forward
# inference is only ever carried out at a single step, but you could
# imagine repeated application of forward inference being profitable.)
#
# A good starting point in the literature for exploring more about proof
# search are these lecture notes
# https://www.cs.cmu.edu/~fp/courses/oregon-m10/04-focusing.pdf
#
# TODO: My kingdom for a pattern matcher
# https://www.python.org/dev/peps/pep-0634/
#
# TODO: This could get us in recomputation trouble if b.expr is nontrivial.
# Fix this by implementing some sort of sharing so that if multiple
# goals share the same expression, we only compute it once. This seems
# to matter in practice as compiler is often unwilling to CSE nontrivial
# expressions like scalar.to<scalar_t>()
t = b.type
if (isinstance(t, ConstRefCType) and isinstance(t.elem, OptionalCType)
and isinstance(t.elem.elem, BaseCType)
and str(t.elem.elem.type) == "at::Tensor"):
ctx[NamedCType(
t.elem.elem.name, ConstRefCType(BaseCType(tensorT))
)] = f"({b.expr}.has_value() ? *{b.expr} : at::Tensor())"
if t.type == ConstRefCType(OptionalCType(BaseCType(tensorT))):
ctx[NamedCType(
t.name, BaseCType(optionalTensorRefT)
)] = f"(({b.expr}.has_value() && (*{b.expr}).defined()) ? at::OptionalTensorRef(*{b.expr}) : at::OptionalTensorRef())"
if t.type == ConstRefCType(BaseCType(scalarT)):
ctx[NamedCType(t.name,
BaseCType(opmath_t))] = f"({b.expr}).to<opmath_t>()"
if t.type == ConstRefCType(OptionalCType(BaseCType(scalarT))):
ctx[NamedCType(
t.name, BaseCType(optionalScalarRefT)
)] = f"({b.expr}.has_value() ? at::OptionalScalarRef(&({b.expr}.value())) : at::OptionalScalarRef())"
if t.type == BaseCType(scalar_t):
ctx[NamedCType(
t.name,
BaseCType(opmath_t))] = f"static_cast<opmath_t>({b.expr})"
# [Note: ITensorListRef]
if t.type == BaseCType(tensorListT):
ctx[NamedCType(
t.name,
BaseCType(iTensorListRefT))] = f"at::ITensorListRef({b.expr})"
# [Note: IOptTensorListRef]
if t.type == ConstRefCType(ListCType(OptionalCType(
BaseCType(tensorT)))):
ctx[NamedCType(t.name, BaseCType(
iOptTensorListRefT))] = f"at::IOptTensorListRef({b.expr})"
# Add implicit bindings if the generated code is inside a Tensor method
if method:
ctx[NamedCType("self", MutRefCType(
BaseCType(tensorT)))] = "const_cast<Tensor&>(*this)"
ctx[NamedCType("self", ConstRefCType(
BaseCType(tensorT)))] = "const_cast<Tensor&>(*this)"
# This is better! Byte-for-byte compat
# ctx[NamedCType("self", ConstRefCType(BaseCType(tensorT)))] = "*this"
def unsat(goal: NamedCType) -> NoReturn:
ctx_desc = "\n".join(f" {t.cpp_type()} {t.name}; // {e}"
for t, e in ctx.items())
raise UnsatError(f"""
Failed to synthesize the expression "{goal.cpp_type()} {goal.name}".
When I failed, the following bindings were available in the context:
{ctx_desc}
This probably means there is a missing rule in the rules of torchgen.api.translate.
Check this module for more information.
""")
# A shitty backtracking search implementation. It's shitty because it
# does backtracking via stack (bad idea!) and for the most part tries to
# avoid backtracking. In particular, if
# direct=True, we won't try to do any fancy synthesis, just trivial
# conversions (e.g., "T a" is OK for "const T& a"). So all of the
# existing rules in this function simply try to solve immediately,
# and bail if things don't work out.
def solve(goal: NamedCType, *, direct: bool) -> str:
def direct_solve(goal: NamedCType) -> str:
return solve(goal, direct=True)
if goal in ctx:
# Trivial
return ctx[goal]
# const & is satisfied with mutable &
if isinstance(goal.type, ConstRefCType):
try:
# WARNING: not strictly decreasing; be careful not
# to add a direct conversion that goes satisfies
# mutable& with const&
return solve(NamedCType(goal.name,
MutRefCType(goal.type.elem)),
direct=direct)
except UnsatError:
pass
# mutable & is satisfied with value
if isinstance(goal.type, MutRefCType):
try:
return solve(NamedCType(goal.name, goal.type.elem),
direct=direct)
except UnsatError:
pass
if direct:
unsat(goal)
# For now, all of these rules are mutually exclusive.
if goal == NamedCType("memory_format",
OptionalCType(BaseCType(memoryFormatT))):
memory_format = direct_solve(
NamedCType(
SpecialArgName.possibly_redundant_memory_format,
OptionalCType(BaseCType(memoryFormatT)),
))
# No need to join "memory_format" and "options" if the target API takes "options" directly.
# Otherwise it will cause the redundant memory_format error.
if options_ctype in goal_ctypes:
return memory_format
try:
options = direct_solve(options_ctype)
return f"c10::impl::check_tensor_options_and_extract_memory_format({options}, {memory_format})"
except UnsatError:
return memory_format
elif goal == NamedCType("options", BaseCType(tensorOptionsT)):
dtype = direct_solve(
NamedCType("dtype", OptionalCType(BaseCType(scalarTypeT))))
pin_memory = direct_solve(
NamedCType("pin_memory", OptionalCType(BaseCType(boolT))))
device = direct_solve(
NamedCType("device", OptionalCType(BaseCType(deviceT))))
layout = direct_solve(
NamedCType("layout", OptionalCType(BaseCType(layoutT))))
return f"TensorOptions().dtype({dtype}).layout({layout}).device({device}).pinned_memory({pin_memory})"
elif goal == NamedCType("dtype",
OptionalCType(BaseCType(scalarTypeT))):
try:
options = direct_solve(options_ctype)
return f"optTypeMetaToScalarType({options}.dtype_opt())"
except UnsatError:
out_tensor = direct_solve(out_tensor_ctype)
return f"{out_tensor}.scalar_type()"
elif goal == NamedCType("layout", OptionalCType(BaseCType(layoutT))):
try:
options = direct_solve(options_ctype)
return f"{options}.layout_opt()"
except UnsatError:
out_tensor = direct_solve(out_tensor_ctype)
return f"{out_tensor}.layout()"
elif goal == NamedCType("device", OptionalCType(BaseCType(deviceT))):
try:
options = direct_solve(options_ctype)
return f"{options}.device_opt()"
except UnsatError:
out_tensor = direct_solve(out_tensor_ctype)
return f"{out_tensor}.device()"
elif goal == NamedCType("pin_memory", OptionalCType(BaseCType(boolT))):
try:
options = direct_solve(options_ctype)
return f"{options}.pinned_memory_opt()"
except UnsatError:
# If we're calling a factory op from its out= variant,
# We don't actually care about the value of pin_memory.
out_tensor = direct_solve(out_tensor_ctype)
return "c10::nullopt"
# We can always do translations from value types to reference types, like vector<int> -> IntArrayRef
elif goal.type == BaseCType(intArrayRefT):
try:
return direct_solve(NamedCType(goal.name, longVec_ctype))
except UnsatError:
# We can also go SymIntArrayRef -> IntArrayRef
symIntArrayRef_type = direct_solve(
NamedCType(goal.name, BaseCType(symIntArrayRefT)))
return f"c10::asIntArrayRefSlow({symIntArrayRef_type})"
elif goal.type == BaseCType(symIntArrayRefT):
return direct_solve(NamedCType(goal.name, longSymVec_ctype))
elif goal.type == BaseCType(longT):
symInt_type = direct_solve(
NamedCType(goal.name, BaseCType(SymIntT)))
return f"{symInt_type}.expectInt()"
elif goal.type == BaseCType(optionalIntArrayRefT):
return direct_solve(NamedCType(goal.name, optionalLongVec_ctype))
elif goal.type == BaseCType(optionalScalarRefT):
return direct_solve(NamedCType(goal.name, optionalScalar_ctype))
elif goal.type == BaseCType(optionalTensorRefT):
return direct_solve(NamedCType(goal.name, optionalTensor_ctype))
# Note [translation from C++ reference to value types]
# The below cases are all for when we have an argument with a reference type,
# and a corresponding goal with a value type.
# These are needed when we populate the inputs to a lambda capture and we need
# to guarantee the lifetime of each captured argument.
# We guard it with an explicit kwarg because converting to a value type is expensive
# (O(n)) to convert from IntArrayRef to vector<int>),
# so the caller of translate() should be explicit that they need it.
if allow_expensive_conversions:
if goal.type == VectorCType(BaseCType(longT)):
intArrayRef_ctype = NamedCType(goal.name,
BaseCType(intArrayRefT))
argname = direct_solve(intArrayRef_ctype)
return f"{argname}.vec()"
if goal.type == VectorCType(BaseCType(SymIntT)):
symIntArrayRef_ctype = NamedCType(goal.name,
BaseCType(symIntArrayRefT))
argname = direct_solve(symIntArrayRef_ctype)
return f"{argname}.vec()"
elif goal.type == OptionalCType(VectorCType(BaseCType(longT))):
optionalIntArrayRef_ctype = NamedCType(
goal.name, BaseCType(optionalIntArrayRefT))
argname = direct_solve(optionalIntArrayRef_ctype)
return f"{argname}.has_value() ? c10::make_optional({argname}->vec()) : c10::nullopt"
elif goal.type == OptionalCType(BaseCType(scalarT)):
optionalScalarRef_ctype = NamedCType(
goal.name, BaseCType(optionalScalarRefT))
argname = direct_solve(optionalScalarRef_ctype)
return f"{argname}.has_value() ? c10::make_optional({argname}) : c10::nullopt"
elif goal.type == OptionalCType(BaseCType(scalarT)):
optionalTensorRef_ctype = NamedCType(
goal.name, BaseCType(optionalTensorRefT))
argname = direct_solve(optionalTensorRef_ctype)
return f"{argname}.has_value() ? c10::make_optional({argname}) : c10::nullopt"
# Technically, we also need to handle cases of C++ containers holding reference types.
# But there currently aren't any ops that require lambda capture codegen
# With arguments like std::vector<IntArrayRef>.
# If that changes, we'll have to add the translation here.
# We allow const casting on tensors, since const-correctness is a bit broken for at::Tensor.
# We could probably generalize this to non-tensor types too.
if goal.type == MutRefCType(BaseCType(tensorT)):
const_ref_tensor_ctype = NamedCType(
goal.name, ConstRefCType(BaseCType(tensorT)))
argname = direct_solve(const_ref_tensor_ctype)
return f"const_cast<Tensor&>({argname})"
unsat(goal)
return [Expr(solve(g, direct=False), g) for g in goal_ctypes]
def solve(goal: NamedCType, *, direct: bool) -> str:
def direct_solve(goal: NamedCType) -> str:
return solve(goal, direct=True)
if goal in ctx:
# Trivial
return ctx[goal]
# const & is satisfied with mutable &
if isinstance(goal.type, ConstRefCType):
try:
# WARNING: not strictly decreasing; be careful not
# to add a direct conversion that goes satisfies
# mutable& with const&
return solve(NamedCType(goal.name,
MutRefCType(goal.type.elem)),
direct=direct)
except UnsatError:
pass
# mutable & is satisfied with value
if isinstance(goal.type, MutRefCType):
try:
return solve(NamedCType(goal.name, goal.type.elem),
direct=direct)
except UnsatError:
pass
if direct:
unsat(goal)
# For now, all of these rules are mutually exclusive.
if goal == NamedCType("memory_format",
OptionalCType(BaseCType(memoryFormatT))):
memory_format = direct_solve(
NamedCType(
SpecialArgName.possibly_redundant_memory_format,
OptionalCType(BaseCType(memoryFormatT)),
))
# No need to join "memory_format" and "options" if the target API takes "options" directly.
# Otherwise it will cause the redundant memory_format error.
if options_ctype in goal_ctypes:
return memory_format
try:
options = direct_solve(options_ctype)
return f"c10::impl::check_tensor_options_and_extract_memory_format({options}, {memory_format})"
except UnsatError:
return memory_format
elif goal == NamedCType("options", BaseCType(tensorOptionsT)):
dtype = direct_solve(
NamedCType("dtype", OptionalCType(BaseCType(scalarTypeT))))
pin_memory = direct_solve(
NamedCType("pin_memory", OptionalCType(BaseCType(boolT))))
device = direct_solve(
NamedCType("device", OptionalCType(BaseCType(deviceT))))
layout = direct_solve(
NamedCType("layout", OptionalCType(BaseCType(layoutT))))
return f"TensorOptions().dtype({dtype}).layout({layout}).device({device}).pinned_memory({pin_memory})"
elif goal == NamedCType("dtype",
OptionalCType(BaseCType(scalarTypeT))):
try:
options = direct_solve(options_ctype)
return f"optTypeMetaToScalarType({options}.dtype_opt())"
except UnsatError:
out_tensor = direct_solve(out_tensor_ctype)
return f"{out_tensor}.scalar_type()"
elif goal == NamedCType("layout", OptionalCType(BaseCType(layoutT))):
try:
options = direct_solve(options_ctype)
return f"{options}.layout_opt()"
except UnsatError:
out_tensor = direct_solve(out_tensor_ctype)
return f"{out_tensor}.layout()"
elif goal == NamedCType("device", OptionalCType(BaseCType(deviceT))):
try:
options = direct_solve(options_ctype)
return f"{options}.device_opt()"
except UnsatError:
out_tensor = direct_solve(out_tensor_ctype)
return f"{out_tensor}.device()"
elif goal == NamedCType("pin_memory", OptionalCType(BaseCType(boolT))):
try:
options = direct_solve(options_ctype)
return f"{options}.pinned_memory_opt()"
except UnsatError:
# If we're calling a factory op from its out= variant,
# We don't actually care about the value of pin_memory.
out_tensor = direct_solve(out_tensor_ctype)
return "c10::nullopt"
# We can always do translations from value types to reference types, like vector<int> -> IntArrayRef
elif goal.type == BaseCType(intArrayRefT):
try:
return direct_solve(NamedCType(goal.name, longVec_ctype))
except UnsatError:
# We can also go SymIntArrayRef -> IntArrayRef
symIntArrayRef_type = direct_solve(
NamedCType(goal.name, BaseCType(symIntArrayRefT)))
return f"c10::asIntArrayRefSlow({symIntArrayRef_type})"
elif goal.type == BaseCType(symIntArrayRefT):
return direct_solve(NamedCType(goal.name, longSymVec_ctype))
elif goal.type == BaseCType(longT):
symInt_type = direct_solve(
NamedCType(goal.name, BaseCType(SymIntT)))
return f"{symInt_type}.expectInt()"
elif goal.type == BaseCType(optionalIntArrayRefT):
return direct_solve(NamedCType(goal.name, optionalLongVec_ctype))
elif goal.type == BaseCType(optionalScalarRefT):
return direct_solve(NamedCType(goal.name, optionalScalar_ctype))
elif goal.type == BaseCType(optionalTensorRefT):
return direct_solve(NamedCType(goal.name, optionalTensor_ctype))
# Note [translation from C++ reference to value types]
# The below cases are all for when we have an argument with a reference type,
# and a corresponding goal with a value type.
# These are needed when we populate the inputs to a lambda capture and we need
# to guarantee the lifetime of each captured argument.
# We guard it with an explicit kwarg because converting to a value type is expensive
# (O(n)) to convert from IntArrayRef to vector<int>),
# so the caller of translate() should be explicit that they need it.
if allow_expensive_conversions:
if goal.type == VectorCType(BaseCType(longT)):
intArrayRef_ctype = NamedCType(goal.name,
BaseCType(intArrayRefT))
argname = direct_solve(intArrayRef_ctype)
return f"{argname}.vec()"
if goal.type == VectorCType(BaseCType(SymIntT)):
symIntArrayRef_ctype = NamedCType(goal.name,
BaseCType(symIntArrayRefT))
argname = direct_solve(symIntArrayRef_ctype)
return f"{argname}.vec()"
elif goal.type == OptionalCType(VectorCType(BaseCType(longT))):
optionalIntArrayRef_ctype = NamedCType(
goal.name, BaseCType(optionalIntArrayRefT))
argname = direct_solve(optionalIntArrayRef_ctype)
return f"{argname}.has_value() ? c10::make_optional({argname}->vec()) : c10::nullopt"
elif goal.type == OptionalCType(BaseCType(scalarT)):
optionalScalarRef_ctype = NamedCType(
goal.name, BaseCType(optionalScalarRefT))
argname = direct_solve(optionalScalarRef_ctype)
return f"{argname}.has_value() ? c10::make_optional({argname}) : c10::nullopt"
elif goal.type == OptionalCType(BaseCType(scalarT)):
optionalTensorRef_ctype = NamedCType(
goal.name, BaseCType(optionalTensorRefT))
argname = direct_solve(optionalTensorRef_ctype)
return f"{argname}.has_value() ? c10::make_optional({argname}) : c10::nullopt"
# Technically, we also need to handle cases of C++ containers holding reference types.
# But there currently aren't any ops that require lambda capture codegen
# With arguments like std::vector<IntArrayRef>.
# If that changes, we'll have to add the translation here.
# We allow const casting on tensors, since const-correctness is a bit broken for at::Tensor.
# We could probably generalize this to non-tensor types too.
if goal.type == MutRefCType(BaseCType(tensorT)):
const_ref_tensor_ctype = NamedCType(
goal.name, ConstRefCType(BaseCType(tensorT)))
argname = direct_solve(const_ref_tensor_ctype)
return f"const_cast<Tensor&>({argname})"
unsat(goal)
def gen_one(self, f: NativeFunction) -> Optional[str]:
assert not f.manual_kernel_registration
if (self.target is Target.REGISTRATION
and not self.selector.is_native_function_selected(f)):
return None
# TODO: Now, there is something interesting going on here. In the code below,
# we generate CompositeExplicitAutograd implementations of functional and inplace
# based on the out implementation. But in fact, out is definable by
# functional too (just not very efficiently), and this is honestly the
# MORE likely situation for a backend implementor. How do we pick?
# Well, taking a page from Haskell type classes and default methods,
# we could conceivably register a circular definition (out in terms
# of functional, and functional in terms of out) and just require
# someone to implement one or the other. We'd have to do a little bit
# of work to not register one of these "weak" definitions unless there
# is a strong definition somewhere in the DAG! So it's not implemented yet.
if (self.backend_index.dispatch_key
== DispatchKey.CompositeExplicitAutograd
and f.func.kind() is SchemaKind.out):
# Never generate a default implementation for out, that's what you
# have to define as a backend implementor
return None
# Note [Direct dispatch bindings]
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Signature of the non-dispatched function we'll expose in a header
# (e.g., at::cpu::add). We don't generate methods (TODO: do this
# when CPUTensor class is a thing); nor do we generate fallback
# bindings for manual_cpp_binding functions.
cpp_sig_group = CppSignatureGroup.from_native_function(
f, method=False, fallback_binding=False)
# Signature of the wrapper function we'll register to the dispatcher
sig = NativeSignature(f.func, prefix="wrapper_")
if self.target is Target.NAMESPACED_DECLARATION:
result = f"TORCH_API {cpp_sig_group.signature.decl()};\n"
if cpp_sig_group.faithful_signature is not None:
result += f"TORCH_API {cpp_sig_group.faithful_signature.decl()};\n"
return result
elif self.target is Target.NAMESPACED_DEFINITION:
def generate_defn(cpp_sig: CppSignature) -> str:
return f"""
{cpp_sig.defn()} {{
return {sig.name()}({', '.join(e.expr for e in translate(cpp_sig.arguments(), sig.arguments()))});
}}
"""
result = generate_defn(cpp_sig_group.signature)
if cpp_sig_group.faithful_signature is not None:
result += generate_defn(cpp_sig_group.faithful_signature)
return result
elif self.target is Target.ANONYMOUS_DEFINITION:
k = f.func.kind()
# Construct the body of the wrapper function with signature sig
sig_body = []
# We'll use context to keep track of any variables we've brought
# into scope while generating code
context: List[Union[Binding, Expr]] = list(sig.arguments())
# Initialize the class corresponding to this structured
# operator; feeding it the output argument(s) if it is known
if self.backend_index.dispatch_key is DispatchKey.Meta:
class_name = f"structured_{meta.name(self.g)}_meta_{k.name}"
parent_class = f"at::meta::structured_{meta.name(self.g)}"
elif (self.backend_index.dispatch_key is
DispatchKey.CompositeExplicitAutograd):
# TODO: dedup this branch
class_name = f"structured_{meta.name(self.g)}_default_backend_{k.name}"
parent_class = f"at::meta::structured_{meta.name(self.g)}"
else:
metadata = self.backend_index.get_kernel(self.g)
assert metadata is not None
class_name = f"structured_{metadata.kernel}_{k.name}"
parent_class = f"{self.cpp_namespace}::structured_{metadata.kernel}"
if self.backend_index.device_guard:
device_check_args = itertools.chain(
f.func.arguments.out, f.func.arguments.flat_positional)
sig_body.append(
RegisterDispatchKey.gen_device_check(
f.device_check, list(device_check_args), sig.name()))
if k is SchemaKind.functional:
sig_body.append(f"{class_name} op;")
elif k is SchemaKind.inplace:
sig_body.append(f"{class_name} op(self);")
elif k is SchemaKind.out:
out_args_str = ", ".join(a.name for a in f.func.arguments.out)
sig_body.append(f"{class_name} op({out_args_str});")
# Translate the input native arguments into structured
# arguments for the meta call
meta_exprs = ", ".join(e.expr for e in translate(
context, structured.meta_arguments(self.g), method=False))
if self.g.out.precomputed:
# If this function group has precomputed elements, the meta function
# returns a struct containing them which must be saved so that it
# can be unpacked when generating code to call the impl.
sig_body.append(f"auto precompute = op.meta({meta_exprs});")
# Put all of the contents of the precompute struct into the context
# so that translate will be able to return the correct args for the
# call to the impl.
precomputed_values = [
*self.g.out.precomputed.replace.values(),
self.g.out.precomputed.add,
]
for precomputed_elems in precomputed_values:
for arg in precomputed_elems:
context.append(
Expr(
expr=f"precompute.{arg.name}",
type=structured.argument_type(arg,
binds=arg.name),
))
# Add a use of the precompute struct so FB internal compilers don't
# complain that there is an unused variable.
sig_body.append("(void)precompute;")
else:
sig_body.append(f"op.meta({meta_exprs});")
# After running meta, op.outputs_ is guaranteed to be valid;
# add it to the context
out_args = structured.out_arguments(self.g)
maybe_star = "*" if k is SchemaKind.functional else ""
for i, out_arg in enumerate(out_args):
assert ConstRefCType(BaseCType(tensorT)) == out_arg.nctype.type
context.append(
Expr(
expr=f"{maybe_star}op.outputs_[{i}]",
# TODO: Stop hardcoding that the output type is a Tensor. Note
# that for the codegen here this is fine because outputs_ is
# hardcoded to be tensor already
type=NamedCType(out_arg.nctype.name,
MutRefCType(BaseCType(tensorT))),
))
# With the expanded context, do the impl call (if not a meta
# function)
if self.backend_index.dispatch_key == DispatchKey.CompositeExplicitAutograd:
# TODO: https://github.com/pytorch/pytorch/issues/53023
out_sig_group = CppSignatureGroup.from_native_function(
self.g.out,
method=False,
fallback_binding=f.manual_cpp_binding)
out_sig = out_sig_group.most_faithful_signature()
api_name = out_sig.name()
out_exprs = ", ".join(e.expr for e in translate(
context, out_sig.arguments(), method=False))
# TODO: I think this means structured won't work with method
# only functions (but maybe you're saved by faithful? iunno.)
# NB: Originally I wrote this as an at::redispatch call, but
# I got in trouble because that meant I needed a DispatchKeySet
# in the wrapper function, which meant I needed a DispatchKeySet
# in the DispatchKeyFunctions declarations, but the defined API
# there does NOT permit a dispatch key set. I think you can
# probably unwind this by calling some function to do the TLS
# fetch and get the DispatchKeySet when you don't have it, but
# I didn't do it for this version
sig_body.append(f"at::{api_name}({out_exprs});")
elif self.backend_index.dispatch_key != DispatchKey.Meta:
impl_exprs = ", ".join(e.expr for e in translate(
context, structured.impl_arguments(self.g), method=False))
sig_body.append(f"op.impl({impl_exprs});")
# Destructively return the final tensors
# TODO: Do this in translate instead
if k is SchemaKind.functional:
if len(f.func.returns) == 1:
ret_expr = "std::move(op.outputs_[0]).take()" # small optimization
else:
moved = ", ".join(f"std::move(op.outputs_[{i}]).take()"
for i in range(len(f.func.returns)))
ret_expr = f"std::make_tuple({moved})"
elif k is SchemaKind.inplace:
ret_expr = "self"
elif k is SchemaKind.out:
if len(f.func.returns) == 1:
ret_expr = f.func.arguments.out[0].name
else:
refs = ", ".join(a.name for a in f.func.arguments.out)
ret_expr = f"std::forward_as_tuple({refs})"
sig_body.append(f"return {ret_expr};")
sig_body_str = "\n".join(sig_body)
# For an overview of what this template code looks like, see
# https://github.com/pytorch/rfcs/pull/9
return f"""\
{self.gen_class(
f, k,
class_name=class_name,
parent_class=parent_class,
generate_super=self.g.out.structured_inherits is not None
)}
{sig.defn()} {{
{sig_body_str}
}}
"""
elif self.target is Target.REGISTRATION:
return f'm.impl("{f.func.name}", TORCH_FN({sig.name()}));'
else:
assert_never(self.target)
# Silence mypy's "Missing return statement" error
return None
def save_var(var: SavedAttribute, is_output: bool) -> None:
name = var.nctype.name
type = var.nctype.type
should_append_getsetdef = True
should_append_raw_getsetdef = False
if (
type == BaseCType(tensorT)
or type == OptionalCType(BaseCType(tensorT))
or type == MutRefCType(OptionalCType(BaseCType(tensorT)))
or (type == BaseCType(scalarT) and is_output)
):
saved_variables.append(f"SavedVariable {name}_;")
release_variables.append(f"{name}_.reset_data();")
ptr = "shared_from_this()" if is_output else ""
unpack.append(f"auto {name} = {name}_.unpack({ptr});")
getter_definitions.append(
GETTER_DEFINITION_SAVEDVAR.substitute(
op=info.op, name=name, body=GETTER_BODY_SAVEDVAR
)
)
getter_definitions.append(
GETTER_DEFINITION_RAW_SAVEDVAR.substitute(
op=info.op, name=name, body=GETTER_BODY_RAW_SAVEDVAR
)
)
should_append_raw_getsetdef = True
elif type == BaseCType(tensorListT):
saved_variables.append(f"std::vector<SavedVariable> {name}_;")
saved_variables.append(f"bool {name}_released_ = false;")
# Just clear() is sufficient, we don't need to loop and clear each variable.
# Because the SavedVariable owns a tensor and a grad_fn, removing the SavedVariable makes them go away as well.
release_variables.append(f"{name}_.clear();")
release_variables.append(f"{name}_released_ = true;")
unpack.append(f"auto {name} = unpack_list({name}_);")
asserts.append(f"TORCH_CHECK(!{name}_released_, ERR_BACKWARD_TWICE);")
getter_definitions.append(
GETTER_DEFINITION_VEC_SAVEDVAR.substitute(
op=info.op, name=name, body=GETTER_BODY_VEC_SAVEDVAR
)
)
getter_definitions.append(
GETTER_DEFINITION_RAW_VEC_SAVEDVAR.substitute(
op=info.op, name=name, body=GETTER_BODY_RAW_VEC_SAVEDVAR
)
)
should_append_raw_getsetdef = True
elif type == ListCType(OptionalCType(BaseCType(tensorT))):
saved_variables.append(f"std::vector<SavedVariable> {name}_;")
saved_variables.append(f"bool {name}_released_ = false;")
# Just clear() is sufficient, we don't need to loop and clear each variable.
# Because the SavedVariable owns a tensor and a grad_fn, removing the SavedVariable makes them go away as well.
release_variables.append(f"{name}_.clear();")
release_variables.append(f"{name}_released_ = true;")
unpack.append(f"auto {name} = unpack_opt_list({name}_);")
asserts.append(f"TORCH_CHECK(!{name}_released_, ERR_BACKWARD_TWICE);")
getter_definitions.append(
GETTER_DEFINITION_VEC_SAVEDVAR.substitute(
op=info.op, name=name, body=GETTER_BODY_VEC_SAVEDVAR
)
)
getter_definitions.append(
GETTER_DEFINITION_RAW_VEC_SAVEDVAR.substitute(
op=info.op, name=name, body=GETTER_BODY_RAW_VEC_SAVEDVAR
)
)
should_append_raw_getsetdef = True
elif type == BaseCType(intArrayRefT):
saved_variables.append(f"std::vector<int64_t> {name};")
getter_definitions.append(
GETTER_DEFINITION.substitute(
op=info.op, name=name, body=GETTER_BODY_ARRAYREF_LONG
)
)
elif type == BaseCType(optionalIntArrayRefT):
saved_variables.append(f"c10::OptionalArray<int64_t> {name};")
getter_definitions.append(
GETTER_DEFINITION_OPT_ARRAYREF.substitute(
op=info.op, name=name, body=GETTER_BODY_ARRAYREF_LONG
)
)
elif type == OptionalCType(BaseCType(intArrayRefT)):
saved_variables.append(f"c10::OptionalArray<int64_t> {name};")
getter_definitions.append(
GETTER_DEFINITION_OPT_ARRAYREF.substitute(
op=info.op, name=name, body=GETTER_BODY_ARRAYREF_LONG
)
)
elif type == OptionalCType(ArrayRefCType(BaseCType(doubleT))):
saved_variables.append(f"c10::OptionalArray<double> {name};")
getter_definitions.append(
GETTER_DEFINITION_OPT_ARRAYREF.substitute(
op=info.op, name=name, body=GETTER_BODY_ARRAYREF_DOUBLE
)
)
elif type == BaseCType(longT):
saved_variables.append(f"{type.cpp_type()} {name} = 0;")
getter_definitions.append(
GETTER_DEFINITION.substitute(
op=info.op, name=name, body=GETTER_BODY_INT64_T
)
)
elif type == BaseCType(stringT):
saved_variables.append(f"std::string {name};")
getter_definitions.append(
GETTER_DEFINITION.substitute(
op=info.op, name=name, body=GETTER_BODY_STRING
)
)
elif type == OptionalCType(BaseCType(stringT)):
saved_variables.append(f"c10::optional<std::string> {name};")
getter_definitions.append(
GETTER_DEFINITION_OPT.substitute(
op=info.op, name=name, body=GETTER_BODY_STRING
)
)
else:
saved_variables.append(f"{type.cpp_type()} {name};")
if type in MISC_GETTER_DEFS:
getter_def, body = MISC_GETTER_DEFS[type]
getter_definitions.append(
getter_def.substitute(op=info.op, name=name, body=body)
)
else:
# Types we don't expose python bindings to yet:
# TypeAndSize, at::ScalarType, TensorOptions, TensorGeometry,
# std::vector<std::vector<int64_t>>, std::vector<at::ScalarType>
should_append_getsetdef = False
if should_append_getsetdef:
py_getsetdef_structs.append(
PY_GETSETDEF_STRUCT.substitute(op=info.op, name=name)
)
if should_append_raw_getsetdef:
py_getsetdef_structs.append(
PY_RAW_GETSETDEF_STRUCT.substitute(op=info.op, name=name)
)