def __call__(self, f: NativeFunction) -> Optional[str]:
if str(f.func.name.name).endswith('_like') or str(
f.func.name.name).startswith('new_'):
return None
name = native.name(f.func)
native_sig = NativeSignature(f.func)
if not any(
isinstance(a.argument, TensorOptionsArguments)
for a in native_sig.arguments()):
return None
native_tensor_args = [
a for a in native_sig.arguments()
if isinstance(a.argument, Argument)
and a.argument.type.is_tensor_like()
]
dispatcher_sig = DispatcherSignature.from_schema(f.func)
sig: Union[NativeSignature, DispatcherSignature]
sig = dispatcher_sig
dispatcher_exprs = dispatcher_sig.exprs()
dispatch_key = "c10::computeDispatchKey(dtype, layout, device)"
if self.target is Target.DEFINITION:
# I don't think there's actually a good reason to generate
# these two cases differently
# The first case could probably be improved though- it calls computeDispatchKeySet(),
# which looks at TLS dispatch keys- there should not be any by the time we reach backend select.
if native_tensor_args:
tensor_args = ', '.join(a.name for a in native_tensor_args)
compute_dk = f"""\
DispatchKeySet _dk_set = c10::DispatchKeySet({dispatch_key}) | c10::detail::multi_dispatch_key_set({tensor_args});
DispatchKeySet _dk_mask = c10::DispatchKeySet(DispatchKeySet::FULL_AFTER, DispatchKey::BackendSelect);
DispatchKeySet _dk = c10::impl::computeDispatchKeySet(_dk_set, _dk_mask);"""
else:
compute_dk = f"DispatchKeySet _dk = c10::DispatchKeySet({dispatch_key});"
return f"""\
// aten::{f.func}
C10_ALWAYS_INLINE
{sig.defn(name)} {{
static auto op = c10::Dispatcher::singleton()
.findSchemaOrThrow("aten::{f.func.name.name}", "{f.func.name.overload_name}")
.typed<{dispatcher_sig.type()}>();
{compute_dk}
return op.redispatch(_dk, {', '.join(a.expr for a in dispatcher_exprs)});
}}
"""
elif self.target is Target.REGISTRATION:
return f"""m.impl("aten::{f.func.name}", TORCH_FN({name}));"""
else:
assert_never(self.target)
def gen_composite_view_copy_kernel(
g: NativeFunctionsViewGroup) -> Optional[str]:
if g.view_copy is None:
return None
# view_copy is a native signature, since we're generating an at::native:: kernel
view_copy_sig = NativeSignature(g.view_copy.func)
# view is a dispatcher signature, since we're calling into the at::_ops API
view_sig = DispatcherSignature(g.view.func)
view_api_name = g.view.func.name.unambiguous_name()
exprs = ', '.join([
e.expr
for e in translate(view_copy_sig.arguments(), view_sig.arguments())
])
# view ops today always return either a Tensor or a list of Tensors
assert len(g.view.func.returns) == 1
assert g.view.func.returns[0].type == BaseType(BaseTy.Tensor) \
or g.view.func.returns[0].type == ListType(BaseType(BaseTy.Tensor), None)
if g.view.func.returns[0].type == BaseType(BaseTy.Tensor):
return_cloned_output = '''\
return output.clone();'''
else:
# If the return type is a list, we need to clone each tensor in the list.
return_cloned_output = f'''\
{view_copy_sig.returns_type().cpp_type()} out_clone;
for (const auto i : c10::irange(output.size())) {{
out_clone.push_back(output[i].clone());
}}
return out_clone;'''
# The default generated composite kernel for {view}_copy() operators just clones
# the input tensor, and runs the underlying view on the clone.
return f"""
def gen_unstructured(f: NativeFunction) -> List[str]:
ns = list(f.dispatch.values())
native_sig = NativeSignature(f.func)
rs = []
# Sometimes a function name shows up multiple times; only generate
# it once!
seen = set()
for n in ns:
if n in seen:
continue
if "legacy::" in n:
continue
seen.add(n)
rs.append(f"TORCH_API {native_sig.decl(name=n)};")
return rs
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 is_cuda_dispatch_key(self.backend_index.dispatch_key):
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.
for precomputed_elems in self.g.out.precomputed.replace.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) -> Optional[str]:
inplace_meta = False
if self.dispatch_key not in f.dispatch:
if (self.dispatch_key == DispatchKey.Meta
and f.func.kind() is SchemaKind.inplace and
# Defer to composites for meta implementation
DispatchKey.CompositeImplicitAutograd not in f.dispatch
and DispatchKey.CompositeExplicitAutograd not in f.dispatch
and
# Inplace list operations are not supported
len(f.func.returns) == 1):
inplace_meta = 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 = NativeSignature(f.func, prefix='wrapper_')
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};
}}
"""
impl_name = f"at::native::{f.dispatch[self.dispatch_key]}"
args_exprs_str = ', '.join(a.name for a in args)
device_guard = "// DeviceGuard omitted" # default
if f.device_guard and is_cuda_dispatch_key(self.dispatch_key):
has_tensor_options = any(
isinstance(a.argument, TensorOptionsArguments)
for a in args)
if has_tensor_options:
# kernel is creating a tensor
device_guard = """globalContext().lazyInitCUDA();
const DeviceGuard device_guard(device_or_default(device));"""
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_guard}
return {impl_name}({args_exprs_str});
}}
}} // anonymous namespace
"""
elif self.target is Target.REGISTRATION:
if f.manual_kernel_registration:
return None
else:
dispatcher_sig = DispatcherSignature.from_schema(f.func)
payload = f"TORCH_FN({name})"
return f'm.impl("{f.func.name}",\n{payload});\n'
else:
assert_never(self.target)