def gen_composite_view_copy_kernel(
g: NativeFunctionsViewGroup) -> Optional[str]:
if g.view_copy is None:
return None
# For view_copy.SymInt overloads,
# See gen_symint_view_copy_kernel.
if g.view_copy.func.name.overload_name == "SymInt":
return None
# We can make view_copy work in more cases by using reshape()
# when a normal view call would ordinarily fail.
# This also makes LTC more efficient, because they don't need to include
# clone() calls in their graph (which is normally needed by reshape).
if str(g.view_copy.func.name) == "view_copy":
return """\
at::Tensor view_copy(const at::Tensor & self, at::IntArrayRef size) {
DimVector shape = infer_size_dv(size, self.numel());
if (!at::detail::computeStride(self.sizes(), self.strides(), shape).has_value()) {
return self.reshape(size);
} else {
auto output = at::_ops::view::call(self, size);
return output.clone();
}
}
"""
# 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_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_symint_view_copy_kernel(view_copy: NativeFunction,
view_copy_symint: NativeFunction) -> str:
# view_copy.symint is a native signature, since we're generating an at::native:: kernel
view_copy_symint_sig = NativeSignature(view_copy_symint.func)
# view_copy is a dispatcher signature, since we're calling into the at::_ops API
view_copy_sig = DispatcherSignature(view_copy.func)
exprs = ", ".join([
e.expr for e in translate(view_copy_symint_sig.arguments(),
view_copy_sig.arguments())
])
return f"""
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