def process_function(f: NativeFunction) -> Optional[str]:
name = cpp.name(f.func)
has_tensor_options = python.has_tensor_options(f)
is_factory = has_tensor_options or name.endswith("_like")
if Variant.function not in f.variants or not is_factory:
return None
sig = CppSignatureGroup.from_native_function(f, method=False).signature
formals: List[str] = []
exprs: List[str] = []
requires_grad = "false"
for arg in sig.arguments():
qualified_type = fully_qualified_type(arg.type)
if arg.default:
formals.append(f"{qualified_type} {arg.name} = {arg.default}")
else:
formals.append(f"{qualified_type} {arg.name}")
if isinstance(arg.argument, TensorOptionsArguments):
# note: we remove the requires_grad setting from the TensorOptions because
# it is ignored anyways (and we actually have an assertion that it isn't set
# which would fail otherwise). We handle requires_grad explicitly here
# instead of passing it through to the kernel.
exprs.append(
f"at::TensorOptions({arg.name}).requires_grad(c10::nullopt)")
# Manually set the requires_grad bit on the result tensor.
requires_grad = f"{arg.name}.requires_grad()"
else:
exprs.append(arg.name)
return f"""\
def __call__(self, f: NativeFunction) -> str:
if not self.selector.is_root_operator(f"aten::{f.func.name}"):
return ""
# We unconditionally generate function wrappers,
sig_group = CppSignatureGroup.from_native_function(f, method=False)
sig = sig_group.most_faithful_signature()
# escape double quote in schema, get rid of extra double quotes
schema = cpp_string(str(sig.func))[1:-1]
# arguments
args = sig.arguments()
connector = ",\n\t\t"
args_code = []
for arg in args:
if not arg.default:
arg_cpp = "c10::IValue(c10::nullopt)"
elif arg.default.startswith("{"):
arg_cpp = f"c10::IntArrayRef({arg.default})"
else:
arg_cpp = f"c10::IValue({arg.default})"
args_code.append(
f"""c10::Argument("{arg.name}", nullptr, c10::nullopt, {arg_cpp})"""
)
returns = f.func.returns
returns_code = []
for ret in returns:
returns_code.append(f"""c10::Argument("{ret.name if ret.name else ""}")""")
return f"""
def signature_original(f: NativeFunction) -> str:
# remove inplace suffix but keep outplace suffix
opname = str(f.func.name.name.base)
if f.func.is_out_fn():
opname += "_out"
if f.func.name.name.inplace and pyi:
opname += "_"
args = CppSignatureGroup.from_native_function(
f, method=False).signature.arguments()
# Simply ignore TensorOptionsArguments as it does not exist in deprecated.yaml.
types = ", ".join(
argument_type_str(a.argument.type) for a in args
if isinstance(a.argument, Argument))
return f"{opname}({types})"
def __call__(self, f: NativeFunction) -> str:
if not self.selector.is_root_operator(f"aten::{f.func.name}"):
return ""
if self.target is Target.DECLARATION:
# Note [The ATen Codegen Unboxing API]
# Similar to the ATen Operators API, ATen Codegen Unboxing API lives in the at::unboxing namespace, and
# will be used by codegen unboxing wrappers (CodegenUnboxingWrappers.cpp).
# The Wrappers will be registered into torch::jit::OperatorRegistry using RegisterOperators API.
#
# Important characteristics about the Codegen Unboxing API:
# (1) It follows the OperatorRegistry API.
# This is kind of necessary to avoid overhead.
# For example: if it followed the C++ API, then all of the faithful C++ factory functions
# would need to wrap their arguments into TensorOptions only to unwrap them again.
# (2) Under the hood it calls C++ API.
return f"""
// aten::{f.func}
TORCH_API void {f.func.name.unambiguous_name()}(Stack & stack);
"""
else:
sig_group = CppSignatureGroup.from_native_function(
f, method=(Variant.method in f.variants)
)
sig = sig_group.most_faithful_signature()
# parse arguments into C++ code
binding_list, code_list = convert_arguments(f)
# for each C++ argument, generate the conversion code
code_connector = "\n\t"
arg_connector = ", "
# function call and push back to stack
prefix = "self_base." if sig.method else "at::"
translated_args = translate(
binding_list, sig.arguments(), method=sig.method
)
args_str = f"{arg_connector.join(e.expr for e in translated_args)}"
if len(f.func.returns) == 0:
ret_str = ""
push_str = ""
else:
ret_str = "auto result_ = "
push_str = """
pack(stack, std::move(result_));
"""
return f"""
def process_function(f: NativeFunction) -> Optional[str]:
name = cpp.name(f.func)
has_tensor_options = python.has_tensor_options(f)
is_factory = has_tensor_options or name.endswith("_like")
if Variant.function not in f.variants or not is_factory:
return None
cpp_sigs = CppSignatureGroup.from_native_function(f, method=False)
sigs = [cpp_sigs.signature]
if cpp_sigs.symint_signature is not None:
sigs.append(cpp_sigs.symint_signature)
r = ""
for sig in sigs:
formals: List[str] = []
exprs: List[str] = []
requires_grad = "false"
for arg in sig.arguments():
qualified_type = fully_qualified_type(arg.type)
if arg.default:
formals.append(f"{qualified_type} {arg.name} = {arg.default}")
else:
formals.append(f"{qualified_type} {arg.name}")
if isinstance(arg.argument, TensorOptionsArguments):
# note: we remove the requires_grad setting from the TensorOptions because
# it is ignored anyways (and we actually have an assertion that it isn't set
# which would fail otherwise). We handle requires_grad explicitly here
# instead of passing it through to the kernel.
exprs.append(
f"at::TensorOptions({arg.name}).requires_grad(c10::nullopt)"
)
# Manually set the requires_grad bit on the result tensor.
requires_grad = f"{arg.name}.requires_grad()"
else:
exprs.append(arg.name)
r += f"""\
inline at::Tensor {sig.name()}({', '.join(formals)}) {{
at::AutoDispatchBelowADInplaceOrView guard;
return autograd::make_variable(at::{sig.name()}({', '.join(exprs)}), /*requires_grad=*/{requires_grad});
}}
"""
return r
def __call__(self, f: NativeFunction) -> str:
if not self.selector.is_root_operator(f"aten::{f.func.name}"):
return ""
# We unconditionally generate function wrappers,
sig_group = CppSignatureGroup.from_native_function(f, method=False)
sig = sig_group.most_faithful_signature()
# escape double quote in schema, get rid of extra double quotes
schema = cpp_string(str(sig.func))[1:-1]
# arguments
args = sig.arguments()
connector = ",\n\t\t"
args_code = []
for arg in args:
# Using method=False faithful C++ API, so we should not see SelfArgument/TensorOptionsArgument
assert isinstance(arg.argument, Argument)
if not arg.argument.default:
arg_cpp = "c10::IValue(c10::nullopt)"
else:
# The unboxing code uses the faithful C++ API to avoid the overhead
# from wrapping/unwrapping TensorOptios.
# However, we would look to include default args for schema parsing.
# Default args only show up in the nonfaithful C++ API,
arg_default = cpp.default_expr(arg.argument.default,
arg.argument.type)
if arg_default.startswith("{"):
arg_cpp = f"c10::IntArrayRef({arg_default})"
else:
arg_cpp = f"c10::IValue({arg_default})"
args_code.append(
f"""c10::Argument("{arg.name}", nullptr, c10::nullopt, {arg_cpp})"""
)
returns = f.func.returns
returns_code = []
for ret in returns:
returns_code.append(
f"""c10::Argument("{ret.name if ret.name else ""}")""")
return f"""
def convert_arguments(f: NativeFunction) -> Tuple[List[Binding], List[str]]:
# we need the 'self' argument so method needs to be False
args = (CppSignatureGroup.from_native_function(
f, method=False).most_faithful_signature().arguments())
code_list = [
f"c10::IValue {args[i].name} = std::move(peek(stack, {i}, {len(args)}));"
for i in range(len(args))
] + [""]
binding_list = []
for i, arg in enumerate(args):
# expecting only Argument
if not isinstance(arg.argument, Argument):
raise Exception(
f"Unexpected argument type, expecting `Argument` but got {arg}"
)
argument: Argument = arg.argument
unboxed_name, _, code, decl = argumenttype_ivalue_convert(
argument.type, argument.name, mutable=argument.is_write)
code_list.extend(decl)
code_list.extend(code)
binding_list.append(arg.with_name(unboxed_name))
return binding_list, code_list
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 gen_unstructured(
self,
f: NativeFunction,
g: Optional[NativeFunctionsGroup] = None) -> Optional[str]:
with native_function_manager(f):
inplace_meta = False
gets_out_inplace_wrapper = False
if not self.backend_index.has_kernel(f):
if (self.backend_index.dispatch_key == DispatchKey.Meta
and f.func.kind() is SchemaKind.inplace and
# Defer to composites for meta implementation
not f.has_composite_kernel and
# Inplace list operations are not supported
len(f.func.returns) == 1):
inplace_meta = True
elif (not self.backend_index.use_out_as_primary
and g is not None and gets_generated_out_inplace_wrapper(
f, g, self.backend_index)):
# We want to generate inplace/out wrappers, that don't have a kernel for the backend.
gets_out_inplace_wrapper = True
else:
return None
if f.manual_kernel_registration:
return None
if (self.target is Target.REGISTRATION
and not self.selector.is_native_function_selected(f)):
return None
sig = self.wrapper_kernel_sig(f)
name = sig.name()
returns_type = sig.returns_type().cpp_type()
args = sig.arguments()
args_str = ", ".join(a.defn() for a in args)
# See Note [Direct dispatch bindings]
cpp_sig_group = CppSignatureGroup.from_native_function(
f, method=False, fallback_binding=False)
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:
# short circuit for inplace_meta
if inplace_meta:
assert f.func.arguments.self_arg is not None
self_arg_name = f.func.arguments.self_arg.argument.name
# TODO: handle in place on tensor list
return f"""
{returns_type} {name}({args_str}) {{
TORCH_CHECK_NOT_IMPLEMENTED({self_arg_name}.is_meta(),
"Cannot inplace into non-meta tensor with meta tensor argument");
return {self_arg_name};
}}
"""
# short circuit for generated inplace/out wrappers
if gets_out_inplace_wrapper:
return self.gen_out_inplace_wrapper(f, g)
metadata = self.backend_index.get_kernel(f)
if metadata is None:
return None
if self.class_method_name is None:
impl_name = f"{self.cpp_namespace}::{metadata.kernel}"
else:
impl_name = f"{self.cpp_namespace}::{self.class_method_name}::{metadata.kernel}"
args_exprs_str = ", ".join(a.name for a in args)
device_check = " // No device check\n"
# Backends that require device guards presumably also require device checks.
if self.backend_index.device_guard:
device_check_args = itertools.chain(
f.func.arguments.out, f.func.arguments.flat_positional)
device_check = RegisterDispatchKey.gen_device_check(
f.device_check, list(device_check_args), name)
device_guard = "// DeviceGuard omitted" # default
if f.device_guard and self.backend_index.device_guard:
has_tensor_options = any(
isinstance(a, TensorOptionsArguments)
for a in f.func.arguments.non_out)
if has_tensor_options:
# kernel is creating a tensor
device_guard = """
const DeviceGuard device_guard(device_or_default(device));"""
# CUDA requires special handling
if is_cuda_dispatch_key(
self.backend_index.dispatch_key):
device_guard = (
f"globalContext().lazyInitCUDA();\n{device_guard}"
)
else:
# kernel is operating on existing tensors
# There is precedence for which argument we use to do
# device guard. This describes the precedence order.
self_arg = ([
f.func.arguments.self_arg.argument
] if f.func.arguments.self_arg is not None else [])
candidate_args = itertools.chain(
self_arg,
f.func.arguments.out,
f.func.arguments.flat_positional,
)
# Only tensor like arguments are eligible
device_of = next(
(f"{a.name}" for a in candidate_args
if a.type.is_tensor_like()),
None,
)
if device_of is not None:
device_guard = f"const OptionalDeviceGuard device_guard(device_of({device_of}));"
return f"""\
namespace {{
{returns_type} {name}({args_str}) {{
{device_check}
{device_guard}
return {impl_name}({args_exprs_str});
}}
}} // anonymous namespace
"""
elif self.target is Target.REGISTRATION:
if f.manual_kernel_registration or self.skip_dispatcher_op_registration:
return None
else:
payload = f"TORCH_FN({name})"
return f'm.impl("{f.func.name}",\n{payload});\n'
else:
assert_never(self.target)
def cpp_arguments(f: NativeFunction) -> Sequence[Binding]:
return CppSignatureGroup.from_native_function(
f, method=False).signature.arguments()
