You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

2796 lines
109 KiB

import dataclasses
import itertools
import re
from dataclasses import dataclass
from enum import auto, Enum
from typing import Callable, Dict, Iterator, List, Optional, Sequence, Set, Tuple, Union
from torchgen.utils import assert_never, NamespaceHelper, OrderedSet
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
#
# DATA MODEL
#
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
#
# Some general principles for our data model.
#
# - Stop using C++ data types as the internal data representation
# format. Instead, the internal data structures are centered
# around JIT schema representation. This avoid a big problem
# with the old codegen where we read in all the types from
# native_functions.yaml and then immediately had to retranslate
# them into C++ types.
#
# - More semantic data representation. Instead of representing
# everything as dicts and strings, we define dataclasses for
# every interesting entity the code generation has to deal with.
# These dataclasses have strong semantic invariants: for example,
# we generally require them to roundtrip losslessly into the
# form they were parsed from. These structures are immutable
# and you're expected to populate information once during
# construction.
# Represent a source location; used for better error reporting
@dataclass(frozen=True)
class Location:
file: str
line: int
def __str__(self) -> str:
return f"{self.file}:{self.line}"
# Valid values of the 'variants' field in native_functions.yaml
class Variant(Enum):
function = auto()
method = auto()
# Default kernel namespace
DEFAULT_KERNEL_NAMESPACE = "at::native"
# NOTE: Keep the list in sync with `DispatchKey` in c10/core/DispatchKey.h
BACKEND_COMPONENTS = "CPU CUDA HIP XLA MTIA MPS IPU XPU HPU VE Lazy Meta PrivateUse1 PrivateUse2 PrivateUse3".split()
FUNCTIONALITY_KEYS = [
"",
"Quantized",
"Sparse",
"SparseCsr",
"NestedTensor",
"Autograd",
]
# This list guards dispatches that can be used in derivatives.yaml
# For now we omit AutogradFunctionality and AutogradOther
AUTOGRAD_KEYS = ["AutogradNestedTensor"] + [
"Autograd" + component for component in BACKEND_COMPONENTS
]
FRAGMENT_NAMESPACES = {"quantized", "quantized_decomposed"}
# This doesn't have to be in sync with the header, it only needs to contain
# entries that we actually use in the codegen or want pyi entries for
class DispatchKey(Enum):
Undefined = 0
CatchAll = Undefined
FPGA = auto()
ORT = auto()
Vulkan = auto()
Metal = auto()
MKLDNN = auto()
OpenGL = auto()
OpenCL = auto()
IDEEP = auto()
CustomRNGKeyId = auto()
MkldnnCPU = auto()
Sparse = auto()
SparseCsr = auto()
NestedTensor = auto()
Dense = auto()
PreDispatch = auto()
Python = auto()
FuncTorchDynamicLayerBackMode = auto()
ZeroTensor = auto()
Conjugate = auto()
Negative = auto()
BackendSelect = auto()
Named = auto()
AutogradOther = auto()
AutogradFunctionality = auto()
AutogradNestedTensor = auto()
Tracer = auto()
Autocast = auto()
Batched = auto()
VmapMode = auto()
FuncTorchGradWrapper = auto()
FuncTorchBatched = auto()
BatchedNestedTensor = auto()
FuncTorchVmapMode = auto()
FuncTorchDynamicLayerFrontMode = auto()
Functionalize = auto()
TESTING_ONLY_GenericWrapper = auto()
TESTING_ONLY_GenericMode = auto()
ADInplaceOrView = auto()
Autograd = auto()
CompositeImplicitAutograd = auto()
CompositeImplicitAutogradNestedTensor = auto()
CompositeExplicitAutograd = auto()
CompositeExplicitAutogradNonFunctional = auto()
FuncTorchBatchedDecomposition = auto()
# BEGIN autogenerated
CPU = auto()
CUDA = auto()
HIP = auto()
XLA = auto()
MTIA = auto()
MPS = auto()
IPU = auto()
XPU = auto()
HPU = auto()
VE = auto()
Lazy = auto()
Meta = auto()
PrivateUse1 = auto()
PrivateUse2 = auto()
PrivateUse3 = auto()
QuantizedCPU = auto()
QuantizedCUDA = auto()
QuantizedHIP = auto()
QuantizedXLA = auto()
QuantizedMTIA = auto()
QuantizedMPS = auto()
QuantizedIPU = auto()
QuantizedXPU = auto()
QuantizedHPU = auto()
QuantizedVE = auto()
QuantizedLazy = auto()
QuantizedMeta = auto()
QuantizedPrivateUse1 = auto()
QuantizedPrivateUse2 = auto()
QuantizedPrivateUse3 = auto()
SparseCPU = auto()
SparseCUDA = auto()
SparseHIP = auto()
SparseXLA = auto()
SparseMTIA = auto()
SparseMPS = auto()
SparseIPU = auto()
SparseXPU = auto()
SparseHPU = auto()
SparseVE = auto()
SparseLazy = auto()
SparseMeta = auto()
SparsePrivateUse1 = auto()
SparsePrivateUse2 = auto()
SparsePrivateUse3 = auto()
SparseCsrCPU = auto()
SparseCsrCUDA = auto()
SparseCsrHIP = auto()
SparseCsrXLA = auto()
SparseCsrMTIA = auto()
SparseCsrMPS = auto()
SparseCsrIPU = auto()
SparseCsrXPU = auto()
SparseCsrHPU = auto()
SparseCsrVE = auto()
SparseCsrLazy = auto()
SparseCsrMeta = auto()
SparseCsrPrivateUse1 = auto()
SparseCsrPrivateUse2 = auto()
SparseCsrPrivateUse3 = auto()
NestedTensorCPU = auto()
NestedTensorCUDA = auto()
NestedTensorHIP = auto()
NestedTensorXLA = auto()
NestedTensorMTIA = auto()
NestedTensorMPS = auto()
NestedTensorIPU = auto()
NestedTensorXPU = auto()
NestedTensorHPU = auto()
NestedTensorVE = auto()
NestedTensorLazy = auto()
NestedTensorMeta = auto()
NestedTensorPrivateUse1 = auto()
NestedTensorPrivateUse2 = auto()
NestedTensorPrivateUse3 = auto()
AutogradCPU = auto()
AutogradCUDA = auto()
AutogradHIP = auto()
AutogradXLA = auto()
AutogradMTIA = auto()
AutogradMPS = auto()
AutogradIPU = auto()
AutogradXPU = auto()
AutogradHPU = auto()
AutogradVE = auto()
AutogradLazy = auto()
AutogradMeta = auto()
AutogradPrivateUse1 = auto()
AutogradPrivateUse2 = auto()
AutogradPrivateUse3 = auto()
# END autogenerated
def __str__(self) -> str:
return self.name
def lower(self) -> str:
return str(self).lower()
@staticmethod
def parse(value: str) -> "DispatchKey":
for k, v in DispatchKey.__members__.items():
if k == value:
return v
raise AssertionError(f"unknown dispatch key {value}")
class _TorchDispatchModeKey(Enum):
FAKE = auto()
PROXY = auto()
FUNCTIONAL = auto()
def codegen_per_backend_entries() -> str:
r = []
for fk in FUNCTIONALITY_KEYS:
for bc in BACKEND_COMPONENTS:
r.append(f" {fk}{bc} = auto()")
return "\n".join(r)
for fk in FUNCTIONALITY_KEYS:
for bc in BACKEND_COMPONENTS:
if not hasattr(DispatchKey, fk + bc):
r = codegen_per_backend_entries()
print(r)
raise RuntimeError(
f"Missing {fk}{bc} from DispatchKey enum. Here is the autogenerated list we expect to have:\n\n{r}"
)
STRUCTURED_DISPATCH_KEYS = {DispatchKey.MPS, DispatchKey.CUDA, DispatchKey.CPU}
UFUNC_DISPATCH_KEYS = {DispatchKey.CUDA, DispatchKey.CPU}
# Set of supported dispatch keys
dispatch_keys = [
DispatchKey.CPU,
DispatchKey.SparseCPU,
DispatchKey.SparseCsrCPU,
DispatchKey.MkldnnCPU,
DispatchKey.CUDA,
DispatchKey.MPS,
DispatchKey.SparseCUDA,
DispatchKey.SparseCsrCUDA,
DispatchKey.QuantizedCPU,
DispatchKey.QuantizedCUDA,
DispatchKey.CompositeImplicitAutograd,
DispatchKey.CompositeImplicitAutogradNestedTensor,
DispatchKey.CompositeExplicitAutograd,
DispatchKey.CompositeExplicitAutogradNonFunctional,
DispatchKey.NestedTensorCPU,
DispatchKey.NestedTensorCUDA,
# Meta is a magic key: it is automatically generated for structured
# kernels
DispatchKey.Meta,
DispatchKey.SparseMeta,
DispatchKey.SparseCsrMeta,
DispatchKey.QuantizedMeta,
DispatchKey.NestedTensorMeta,
DispatchKey.ZeroTensor,
]
# Dispatch keys that "support all backends". These codegen slightly differently
# then backend specific keys.
def is_generic_dispatch_key(dk: DispatchKey) -> bool:
return dk in {
DispatchKey.CompositeExplicitAutograd,
DispatchKey.CompositeExplicitAutogradNonFunctional,
DispatchKey.CompositeImplicitAutograd,
DispatchKey.CompositeImplicitAutogradNestedTensor,
}
# CUDA specific dispatch keys
def is_cuda_dispatch_key(dk: DispatchKey) -> bool:
return dk in {
DispatchKey.CUDA,
DispatchKey.QuantizedCUDA,
DispatchKey.SparseCUDA,
DispatchKey.SparseCsrCUDA,
DispatchKey.NestedTensorCUDA,
DispatchKey.AutogradCUDA,
}
# Structured kernel generation is only supported for certain key types;
# otherwise use old-style
def is_structured_dispatch_key(dk: DispatchKey) -> bool:
return dk in STRUCTURED_DISPATCH_KEYS
def is_ufunc_dispatch_key(dk: DispatchKey) -> bool:
# For now, ufunc dispatch keys coincide with structured keys
return dk in UFUNC_DISPATCH_KEYS
# This is oddly named ScalarType and not DType for symmetry with C++
class ScalarType(Enum):
Byte = auto()
Char = auto()
Short = auto()
Int = auto()
Long = auto()
Half = auto()
Float = auto()
Double = auto()
ComplexHalf = auto()
ComplexFloat = auto()
ComplexDouble = auto()
Bool = auto()
BFloat16 = auto()
Float8_e5m2 = auto()
Float8_e5m2fnuz = auto()
Float8_e4m3fn = auto()
Float8_e4m3fnuz = auto()
def __str__(self) -> str:
return self.name
@staticmethod
def maybe_parse(value: str) -> Optional["ScalarType"]:
for k, v in ScalarType.__members__.items():
if k == value:
return v
return None
@staticmethod
def parse(value: str) -> "ScalarType":
mb_r = ScalarType.maybe_parse(value)
assert mb_r is not None, f"unknown dtype {value}"
return mb_r
@staticmethod
def parse_set(values: str) -> OrderedSet["ScalarType"]:
dtypes: OrderedSet[ScalarType] = OrderedSet()
for value in values.split(", "):
if value in DTYPE_CLASSES:
dtypes.update(DTYPE_CLASSES[value])
else:
dtypes.add(ScalarType.parse(value))
return dtypes
DTYPE_CLASSES: Dict[str, OrderedSet[ScalarType]] = {}
# NB: Integral doesn't include boolean
DTYPE_CLASSES["Integral"] = OrderedSet(
[
ScalarType.Byte,
ScalarType.Char,
ScalarType.Int,
ScalarType.Long,
ScalarType.Short,
]
)
# NB: Floating doesn't include low precision types
DTYPE_CLASSES["Floating"] = OrderedSet([ScalarType.Float, ScalarType.Double])
DTYPE_CLASSES["Complex"] = OrderedSet(
[ScalarType.ComplexFloat, ScalarType.ComplexDouble]
)
DTYPE_CLASSES["All"] = DTYPE_CLASSES["Integral"] | DTYPE_CLASSES["Floating"]
DTYPE_CLASSES["AllAndComplex"] = DTYPE_CLASSES["All"] | DTYPE_CLASSES["Complex"]
DTYPE_CLASSES["FloatingAndComplex"] = (
DTYPE_CLASSES["Floating"] | DTYPE_CLASSES["Complex"]
)
# Represents the valid entries for ufunc_inner_loop in native_functions.yaml.
# NB: if you add a new UfuncKey, you will teach torchgen.dest.ufunc how
# to process it. Most logic will ignore keys they don't understand, so your
# new key will get silently ignored until you hook in logic to deal with it.
class UfuncKey(Enum):
# These are low level keys that represent exactly one particular
# instantiation of the kernel produced by codegen
CUDAFunctor = auto()
CUDAFunctorOnOther = auto()
CUDAFunctorOnSelf = auto()
CPUScalar = auto()
CPUVector = auto()
# These are the ones users will usually specify, and
# implicitly "fill in" the low level keys
ScalarOnly = auto() # CUDA*, CPUScalar
Generic = auto() # CUDA*, CPU*
def __str__(self) -> str:
return self.name
@staticmethod
def parse(value: str) -> "UfuncKey":
for k, v in UfuncKey.__members__.items():
if k == value:
return v
raise AssertionError(f"unknown ufunc key {value}")
class DeviceCheckType(Enum):
NoCheck = 0
ExactSame = 1
class ViewSchemaKind(Enum):
aliasing = auto()
aliasing_inplace = auto()
non_aliasing = auto()
# The basic input to the code generation is native_functions.yaml.
# The name "native", BTW, comes from the distinction between native
# functions and legacy TH functions. The legacy TH functions are gone,
# but the "native" descriptor has stuck.
#
# NativeFunction models a single entry in native_functions.yaml. Its
# fields roughly correspond to what you would see in the YAML itself,
# but after canonicalization and parsing has occurred.
#
# You can see some of the overall design patterns for how we setup
# dataclasses in this class, but we will defer a complete discussion
# of this at FunctionSchema.
@dataclass(frozen=True)
class NativeFunction:
# The namespace for this operator. For example, if we have "at::add"
# then the namespace would be "at". This enables ops to be registered
# through the same DSL with a custom namespace. If not specified, the
# default namespace would be "at".
namespace: str
# The function schema of the operator in question. This schema
# has been parsed; see FunctionSchema for more about its structure.
# (This type is quoted as we are forward referencing a type
# defined later in the file. I opted for this ordering of the
# classes for expository clarity.)
func: "FunctionSchema"
# Whether or not to generate mutable tensor arguments like regular
# ones
use_const_ref_for_mutable_tensors: bool
# Whether or not to omit automatic generation of a DeviceGuard
device_guard: bool
# How to emit automatic generation of device check
device_check: DeviceCheckType
# What python module to put the function in
python_module: Optional[str]
# TODO: figure out what this does
category_override: Optional[str]
# If no variants are specified in native_functions.yaml, this is
# assumed to be {'function'}.
variants: Set[Variant]
# Whether or not we should skip generating registrations for
# this kernel. This is a bit of a double-edged sword, as manual
# registrations don't participate in codegen-based selective build!
manual_kernel_registration: bool
# Whether or not to skip generating TensorMethod/Functions bindings
# for this kernel. Technically, this doesn't actually skip generating
# the binding; instead, the binding gets generated to __dispatch_{funcname}
# so you can make use of the normal binding if you need it.
manual_cpp_binding: bool
# The location in the YAML file were this native function entry was
# defined. This is for conveniently reporting error messages!
loc: "Location"
# A list of operators that are expected to be auto-generated for this NativeFunction.
# Note: This list isn't actually directly used by the codegen to generate anything.
# Instead, the codegen figures out what operators to generate purely based off of
# function schema, and uses the autogen declarations to error check.
# We expect every NativeFunction that gets auto-generated be explicitly called out
# in native_functions.yaml
autogen: List["OperatorName"]
# If non-empty, this kernel is subject to ufunc codegen.
# Sorted by ufunc_key
ufunc_inner_loop: Dict[UfuncKey, "UfuncInnerLoop"]
# Whether or not this out functions is a "structured kernel". Structured
# kernels are defined a little differently from normal kernels; in
# particular, their shape checking logic is defined separately from
# the kernel. Only out functions can be structured; other functions
# delegate to the out function using the structured_delegate keyword.
# Every structured kernel must have at least an out and a functional
# variant.
structured: bool
# Whether or not this non-out function is a structured kernel, defined
# in terms of the out kernel referenced by the string here.
structured_delegate: Optional["OperatorName"]
# Only valid for structured kernels. Specifies alternative of what
# to inherit from when defining the meta class for the structured
# operator. This will usually be TensorIteratorBase. This also
# changes the semantics of set_output to call the parent class.
structured_inherits: Optional[str]
# Structured kernels can declare elements as "precomputed". These elements
# are returned by the meta function in one struct and passed to the impl
# function in lieu of certain kernel arguments that these precomputed
# elements supersede. Information about the names and types of these
# precomputed elements and how they correspond to kernel arguments is stored
# in this member, if applicable.
precomputed: Optional["Precompute"]
# Argument names whose default should be excluded from the C++ interface.
# Intended for resolving overload ambiguities between signatures.
cpp_no_default_args: Set[str]
# Note [Abstract ATen methods]
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# An abstract ATen method is one whose dispatch differs between
# types. These are implemented in derived types (with a
# standard (throwing) definition in Type). A concrete ATen
# method is one which has the same dispatch for all types;
# we just implement it in the base Type. This is exposed
# in Declarations.yaml via a field named 'abstract'.
is_abstract: bool
# Whether or not the NativeFunction contains a backend-agnostic kernel
has_composite_implicit_autograd_kernel: bool
has_composite_implicit_autograd_nested_tensor_kernel: bool
has_composite_explicit_autograd_kernel: bool
has_composite_explicit_autograd_non_functional_kernel: bool
# Tags are used to describe semantic information about (groups of) operators,
# That aren't easily inferrable directly from the operator's schema.
tags: Set[str]
# NB: The benefit of defining a dataclass is that we automatically get
# a constructor defined for all the fields we specify. No need
# to explicitly write it out.
# We parse both the NativeFunction + backend-specific information about it, which it stored in a corresponding BackendIndex.
@staticmethod
def from_yaml(
ei: Dict[str, object],
loc: "Location",
valid_tags: Set[str],
ignore_keys: Optional[Set[DispatchKey]] = None,
) -> Tuple[
"NativeFunction", Dict[DispatchKey, Dict["OperatorName", "BackendMetadata"]]
]:
"""
Parse a NativeFunction from a dictionary as directly parsed
from native_functions.yaml
"""
e = ei.copy()
funcs = e.pop("func")
assert isinstance(funcs, str), f"not a str: {funcs}"
# only support one level of namespace. E.g., aten::add
namespace_helper = NamespaceHelper.from_namespaced_entity(
namespaced_entity=funcs, max_level=1
)
namespace = namespace_helper.get_cpp_namespace(default="aten")
func = FunctionSchema.parse(namespace_helper.entity_name)
cpp_no_default_args_list = e.pop("cpp_no_default_args", [])
assert isinstance(cpp_no_default_args_list, list)
cpp_no_default_args = set(cpp_no_default_args_list)
use_const_ref_for_mutable_tensors = e.pop(
"use_const_ref_for_mutable_tensors", False
)
assert isinstance(use_const_ref_for_mutable_tensors, bool)
variants_s = e.pop("variants", "function")
assert isinstance(variants_s, str)
variants: Set[Variant] = set()
for v in variants_s.split(", "):
if v == "function":
variants.add(Variant.function)
elif v == "method":
variants.add(Variant.method)
else:
raise AssertionError(f"illegal variant {v}")
manual_kernel_registration = e.pop("manual_kernel_registration", False)
assert isinstance(
manual_kernel_registration, bool
), f"not a bool: {manual_kernel_registration}"
manual_cpp_binding = e.pop("manual_cpp_binding", False)
assert isinstance(manual_cpp_binding, bool), f"not a bool: {manual_cpp_binding}"
device_guard = e.pop("device_guard", True)
assert isinstance(device_guard, bool), f"not a bool: {device_guard}"
device_check_s = e.pop("device_check", None)
assert device_check_s is None or isinstance(
device_check_s, str
), f"not a str: {device_check_s}"
device_check: DeviceCheckType
if device_check_s is None:
device_check = DeviceCheckType.ExactSame
else:
device_check = DeviceCheckType[device_check_s]
structured = e.pop("structured", False)
assert isinstance(structured, bool), f"not a bool: {structured}"
structured_delegate_s = e.pop("structured_delegate", None)
assert structured_delegate_s is None or isinstance(
structured_delegate_s, str
), f"not a str: {structured_delegate_s}"
assert structured_delegate_s is None or "::" not in structured_delegate_s, (
"namespace is not supported in structured delegate,"
" using the same namespace as the native function"
)
structured_delegate: Optional[OperatorName] = None
if structured_delegate_s is not None:
structured_delegate = OperatorName.parse(structured_delegate_s)
structured_inherits = e.pop("structured_inherits", None)
assert structured_inherits is None or isinstance(
structured_inherits, str
), f"not a str: {structured_inherits}"
assert structured_inherits is None or "::" not in structured_inherits, (
"namespace is not supported in structured inherits,"
" using the same namespace as the native function"
)
python_module = e.pop("python_module", None)
assert python_module is None or isinstance(
python_module, str
), f"not a str: {python_module}"
assert (
python_module is None or Variant.method not in variants
), "functions in modules cannot be methods"
category_override = e.pop("category_override", None)
assert category_override is None or isinstance(
category_override, str
), f"not a str: {category_override}"
precomputed_dict = e.pop("precomputed", None)
assert precomputed_dict is None or structured is True
precomputed = Precompute.parse(precomputed_dict) if precomputed_dict else None
tags_inp = e.pop("tags", [])
if isinstance(tags_inp, str):
tags_inp = [tags_inp]
assert isinstance(tags_inp, list)
# All aten ops generated by torchgen receive the pt2_compliant tag.
if namespace == "aten" and "pt2_compliant_tag" in valid_tags:
tags_inp.append("pt2_compliant_tag")
tags: Set[str] = set()
for t in tags_inp:
assert len(valid_tags) > 0
# TODO: verify that the tag is valid and has an entry in tags.yaml
if t in valid_tags:
tags.add(t)
else:
raise AssertionError(f"illegal tag {t}")
from torchgen.api import cpp
raw_dispatch = e.pop("dispatch", None)
assert raw_dispatch is None or isinstance(raw_dispatch, dict), e
dispatch: Dict[DispatchKey, BackendMetadata] = {}
num_dispatch_keys: int = 0
if raw_dispatch is not None:
assert not manual_kernel_registration, (
"cannot specify both manual_kernel_registration and dispatch; with "
"manual registration, dispatch has no effect!"
)
redundant_composite_implicit_autograd = False
for ks, v in raw_dispatch.items():
if ks == "__line__":
continue # not worth tracking line numbers for dispatch entries
assert isinstance(ks, str), e
for k in ks.split(","):
dispatch_key = DispatchKey.parse(k.strip())
num_dispatch_keys += 1
if ignore_keys and dispatch_key in ignore_keys:
continue
assert dispatch_key in dispatch_keys, (
f"Dispatch key {dispatch_key} of kernel {v} "
"is not a supported dispatch key."
)
# We only allow at most 3 levels of namespace for kernels.
# We will append "native" to a custom kernel namespace.
namespace_helper = NamespaceHelper.from_namespaced_entity(
v, max_level=3
)
kernel_namespace = namespace_helper.get_cpp_namespace(default="at")
# Why is 'structured' included? External backends (e.g.
# XLA) opt into which ops are structured independently
# of which in-tree ops are structured
dispatch[dispatch_key] = BackendMetadata(
kernel=namespace_helper.entity_name,
structured=structured
and is_structured_dispatch_key(dispatch_key),
cpp_namespace=(kernel_namespace + "::native"),
)
if (
dispatch_key is DispatchKey.CompositeImplicitAutograd
and v == cpp.name(func)
):
redundant_composite_implicit_autograd = True
# We count the number of dispatch keys which have not been ignored to prevent a dispatch table
# in which all backend keys are ignored but necessarily kept, remaining compositeimplicit,
# from being treated as redundant.
assert not (
num_dispatch_keys == 1 and redundant_composite_implicit_autograd
), (
"unnecessary dispatch table for this function; just delete the dispatch "
"key entirely"
)
# if a function is a structured delegate, deleting the dispatch
# table is NOT semantics preserving
assert (
structured_delegate
or dispatch.keys() != {DispatchKey.CompositeImplicitAutograd}
or dispatch[DispatchKey.CompositeImplicitAutograd].supports_symint()
or num_dispatch_keys != 1
), (
f"unexpected name for singleton CompositeImplicitAutograd dispatch entry: expected {cpp.name(func)} "
f"but got {dispatch[DispatchKey.CompositeImplicitAutograd]}. Rename your implementation to the expected "
"name, then delete the dispatch table"
)
elif not structured and structured_delegate is None:
name = str(func.name.name)
assert not (
name.startswith("new_")
or name.endswith("_like")
# TODO: maybe it's better to test the return
or (
func.arguments.tensor_options
and not func.arguments.has_tensor_arg()
)
), (
f"expected {name} to have a CompositeExplicitAutograd "
"dispatch entry, but there was no dispatch table. Factory functions "
"should not have implicit dispatch as they should not be decomposed "
"for __torch_dispatch__"
)
dispatch[DispatchKey.CompositeImplicitAutograd] = BackendMetadata(
cpp.name(func), structured=False, cpp_namespace=DEFAULT_KERNEL_NAMESPACE
)
composites_in_dispatch = [
d
for d in dispatch
if d == DispatchKey.CompositeExplicitAutograd
or d == DispatchKey.CompositeExplicitAutogradNonFunctional
or d == DispatchKey.CompositeImplicitAutograd
or d == DispatchKey.CompositeImplicitAutogradNestedTensor
]
assert len(composites_in_dispatch) <= 1 or (
len(composites_in_dispatch) == 2
and (
DispatchKey.CompositeExplicitAutogradNonFunctional
not in composites_in_dispatch
)
and (
DispatchKey.CompositeImplicitAutogradNestedTensor
in composites_in_dispatch
)
), (
"cannot specify more than one of CompositeExplicitAutograd, CompositeExplicitAutogradNonFunctional, "
"or CompositeImplicitAutograd on a single kernel; each "
"strictly subsumes the other. If you wanted to provide an explicit autograd "
"implementation, specify CompositeExplicitAutograd; otherwise specify CompositeImplicitAutograd only"
)
autogen_str = e.pop("autogen", "")
assert isinstance(autogen_str, str)
autogen = (
[]
if autogen_str == ""
else [OperatorName.parse(x) for x in autogen_str.split(", ")]
)
raw_ufunc_inner_loop = e.pop("ufunc_inner_loop", {})
ufunc_inner_loop = {}
if isinstance(raw_ufunc_inner_loop, str):
ufunc_inner_loop[UfuncKey.Generic] = UfuncInnerLoop.parse(
raw_ufunc_inner_loop, UfuncKey.Generic
)
elif isinstance(raw_ufunc_inner_loop, dict):
for k, vo in raw_ufunc_inner_loop.items():
if k == "__line__":
continue
assert isinstance(k, str), f"ufunc_inner_loop key is not a str: {k}"
assert isinstance(vo, str), f"ufunc_inner_loop value is not a str: {v}"
ufunc_key = UfuncKey.parse(k)
ufunc_inner_loop[ufunc_key] = UfuncInnerLoop.parse(vo, ufunc_key)
else:
raise AssertionError(
f"ufunc_inner_loop not str or dict: {raw_ufunc_inner_loop}"
)
# Program the BackendIndex for the implicit dispatch entry from ufunc
if ufunc_inner_loop:
assert structured, "ufunc must be structured"
# Delay import ufunc here to avoid circular import issue
# See: https://github.com/pytorch/pytorch/issues/81294
import torchgen.api.ufunc as ufunc
for dispatch_key in UFUNC_DISPATCH_KEYS:
assert (
dispatch_key not in dispatch
), f"ufunc should not have explicit dispatch entry for {dispatch_key}"
dispatch[dispatch_key] = BackendMetadata(
kernel=ufunc.schema_kernel_name(func, dispatch_key),
structured=True,
cpp_namespace=DEFAULT_KERNEL_NAMESPACE,
)
if structured_delegate:
# Structured functions MUST have a dispatch table
is_abstract = True
else:
is_abstract = (
dispatch.keys() != {DispatchKey.CompositeImplicitAutograd}
and dispatch.keys()
!= {DispatchKey.CompositeImplicitAutogradNestedTensor}
and dispatch.keys()
!= {
DispatchKey.CompositeImplicitAutograd,
DispatchKey.CompositeImplicitAutogradNestedTensor,
}
)
has_composite_implicit_autograd_kernel = (
DispatchKey.CompositeImplicitAutograd in dispatch.keys()
)
has_composite_implicit_autograd_nested_tensor_kernel = (
DispatchKey.CompositeImplicitAutogradNestedTensor in dispatch.keys()
)
has_composite_explicit_autograd_kernel = (
DispatchKey.CompositeExplicitAutograd in dispatch.keys()
)
has_composite_explicit_autograd_non_functional_kernel = (
DispatchKey.CompositeExplicitAutogradNonFunctional in dispatch.keys()
)
# We aren't going to store dispatch metadata inline in NativeFunctions;
# instead it is separately indexed by backend (so other backends can
# add more dispatch entries after the fact). Reindex the individual
# metadata by OperatorName!
backend_metadata = {k: {func.name: v} for k, v in dispatch.items()}
# don't care if it exists or not; make it easier to use this function
# with other yaml parsers that aren't setting __line__ in the dict
e.pop("__line__", None)
assert not e, f"leftover entries: {e}"
# Asserts that we can't do in post_init, because they rely on backend-specific info
if structured_delegate is not None:
for key in STRUCTURED_DISPATCH_KEYS:
assert key not in dispatch, (
f"if structured_delegate, then must not have {key} in dispatch dictionary "
"(it is delegated!)"
)
return (
NativeFunction(
func=func,
use_const_ref_for_mutable_tensors=use_const_ref_for_mutable_tensors,
variants=variants,
structured=structured,
structured_delegate=structured_delegate,
structured_inherits=structured_inherits,
precomputed=precomputed,
autogen=autogen,
ufunc_inner_loop=ufunc_inner_loop,
manual_kernel_registration=manual_kernel_registration,
manual_cpp_binding=manual_cpp_binding,
python_module=python_module,
category_override=category_override,
device_guard=device_guard,
device_check=device_check,
loc=loc,
cpp_no_default_args=cpp_no_default_args,
is_abstract=is_abstract,
has_composite_implicit_autograd_kernel=has_composite_implicit_autograd_kernel,
has_composite_implicit_autograd_nested_tensor_kernel=has_composite_implicit_autograd_nested_tensor_kernel,
has_composite_explicit_autograd_kernel=has_composite_explicit_autograd_kernel,
has_composite_explicit_autograd_non_functional_kernel=has_composite_explicit_autograd_non_functional_kernel,
tags=tags,
namespace=namespace,
),
backend_metadata,
)
def validate_unstructured(self) -> None:
# TODO: probably better to accumulate these errors and report them all
# at once
assert not self.structured, (
"This function is structured, but there was "
"no valid functional variant of it."
)
assert self.structured_delegate, (
"This function delegates to another structured out function, "
"but no valid function was found (the delegate may not exist, or it has the wrong type)"
)
# __post_init__ functions in dataclasses can be used to do extra
# validation after construction.
#
# Notice that we don't do any type validation here. In fact, we
# rely exclusively on mypy to check if you've done types correctly!
# Validation is for nontrivial invariants that cannot be (conveniently)
# encoded in the type system.
def __post_init__(self) -> None:
if self.func.arguments.out:
assert self.variants == {Variant.function}, (
"Native functions with out arguments MUST "
"be declared with only function variant; e.g., variants: function; "
"otherwise you will tickle a Python argument binding bug "
"(which usually manifests itself as the result variable being undefined.)"
)
if self.structured:
assert self.func.kind() == SchemaKind.out, (
"Put structured field on the out= "
"variant of a function; did you mean structured_delegate?"
)
assert (
self.device_guard
), "device_guard: False is not respected by structured kernels"
if self.structured_delegate:
assert self.func.kind() != SchemaKind.out, (
"structured_delegate field not allowed "
"on out= functions; did you mean structured?"
)
assert (
self.device_guard
), "device_guard: False is not respected by structured kernels"
# Technically, with the asserts above, this assert is impossible to
# happen
assert not (
self.structured and self.structured_delegate
), "Cannot have both structured and structured_delegate on function"
defaulted_arguments = {
a.name for a in self.func.schema_order_arguments() if a.default is not None
}
invalid_args = set.difference(self.cpp_no_default_args, defaulted_arguments)
assert len(invalid_args) == 0, f"Invalid cpp_no_default_args: {invalid_args}"
if self.structured_inherits is not None:
assert (
self.structured
), "structured_inherits must also imply structured: True"
if str(self.func.name).startswith("_foreach"):
assert self.device_check == DeviceCheckType.NoCheck, (
"foreach kernels fall back to slow path when tensor are on different devices, "
"device_check not allowed to be enabled"
)
# NB: if your function accidentally has rand/dropout/... in its name
# but is not actually random, feel free to amend this to special case
if (
"rand" in str(self.func.name)
or (
(
"dropout" in str(self.func.name)
or any(
"dropout" in arg.name for arg in self.func.arguments.flat_all
)
)
# Backwards of dropout is typically deterministic
and "backward" not in str(self.func.name)
and str(self.func.name.name) not in ["_cudnn_init_dropout_state"]
)
or self.func.arguments.has_generator_arg()
):
assert "nondeterministic_seeded" in self.tags, str(self.func.name)
@property
def has_composite_kernel(self) -> bool:
return (
self.has_composite_implicit_autograd_kernel
or self.has_composite_explicit_autograd_kernel
or self.has_composite_explicit_autograd_non_functional_kernel
) or (
self.has_composite_implicit_autograd_kernel
and self.has_composite_implicit_autograd_nested_tensor_kernel
)
@property
def is_view_op(self) -> bool:
rets = self.func.returns
is_non_mutating_view = len(rets) > 0 and any(
r.annotation is not None and not r.annotation.is_write for r in rets
)
# See Note [resize_ in Functionalization] for more dtails
is_inplace_view = (
"inplace_view" in self.tags
and str(self.func.name) != "resize_"
and str(self.func.name) != "resize_as_"
)
is_wildcard_view = any(
inp.annotation is not None and "*" in inp.annotation.alias_set_after
for inp in self.func.schema_order_arguments()
)
return is_non_mutating_view or is_inplace_view or is_wildcard_view
@property
def view_schema_kind(self) -> ViewSchemaKind:
if self.is_view_op and self.func.name.name.inplace:
assert "inplace_view" in self.tags
return ViewSchemaKind.aliasing_inplace
if self.is_view_op:
return ViewSchemaKind.aliasing
else:
return ViewSchemaKind.non_aliasing
@property
def root_name(self) -> str:
return self.func.name.name.base
@property
def part_of_structured_group(self) -> bool:
return self.structured or self.structured_delegate is not None
class SchemaKind(Enum):
functional = auto()
inplace = auto()
out = auto()
mutable = auto()
scratch = auto()
# A structured kernel is guaranteed to have a functional and out variant, and
# optionally an inplace variant.
#
# NB: we create NativeFunctionsGroup *even if* the function is not
# actually annotated structured. Test the structured boolean to see if it
# actually is structured or not.
@dataclass(frozen=True)
class NativeFunctionsGroup:
functional: NativeFunction
inplace: Optional[NativeFunction]
mutable: Optional[NativeFunction]
out: NativeFunction
@property
def structured(self) -> bool:
# Whether or not the operator has a meta() function. This information is backend-agnostic.
return self.out.structured
def __post_init__(self) -> None:
test_sig: FunctionSchema = self.functional.func.signature()
for f in self.functions():
if test_sig != f.func.signature():
raise AssertionError(
"NativeFunctionsGroup constructed from two NativeFunctions "
f"that don't have matching signatures: {test_sig} != {f.func.signature()}"
)
if self.structured != f.part_of_structured_group:
raise AssertionError(
"NativeFunctionsGroup constructed from structured and unstructured "
f"functions: {self.out.func.name} and {f.func.name}"
)
assert self.functional.func.kind() == SchemaKind.functional
assert self.out.func.kind() == SchemaKind.out
assert self.functional.namespace == self.out.namespace
if self.inplace is not None:
assert self.inplace.func.kind() == SchemaKind.inplace
assert self.inplace.namespace == self.functional.namespace
if self.mutable is not None:
assert self.mutable.func.kind() == SchemaKind.mutable
assert self.mutable.namespace == self.functional.namespace
# See Note [Overload Ambiguity With Functional Variants]
assert self.functional.func.name.name.functional_overload
if self.structured:
# For now, structured composite kernels are not supported (need some
# design work to figure out how to make the composite case work)
assert (
not self.out.has_composite_implicit_autograd_kernel
and not self.out.has_composite_implicit_autograd_nested_tensor_kernel
)
assert self.functional.structured_delegate == self.out.func.name, (
f"{self.functional.func.name} delegates to {self.functional.structured_delegate} "
f"but its actual delegate is {self.out.func.name}"
)
if self.inplace is not None:
assert self.inplace.structured_delegate == self.out.func.name
generated_fns = sorted(
[str(f.func.name) for f in self.functions() if "generated" in f.tags]
)
generated_fns_str = ", ".join(str(x) for x in generated_fns)
expected_generated_fns: Set[str] = set()
for f in self.functions():
expected_generated_fns.update(str(op) for op in f.autogen)
expected_generated_fns_str = ", ".join(
str(x) for x in sorted(expected_generated_fns)
)
if len(expected_generated_fns) == 0 and len(generated_fns) > 0:
raise RuntimeError(
f"The codegen expects to be able to generate '{generated_fns_str}'."
" In order to generate them however, we expect them to be called out explicitly in the yaml."
f" Please add an 'autogen: {generated_fns_str}' line to the entry for {str(f.func.name)}"
)
if expected_generated_fns_str != generated_fns_str:
raise RuntimeError(
f"The codegen expects to be able to generate '{generated_fns_str}'."
f" To do so, it expects a line: 'autogen: {generated_fns_str}'."
f" Instead, it found 'autogen: {expected_generated_fns_str}'"
)
def signature(self) -> "FunctionSchema":
return self.out.func.signature()
def functions(self) -> Iterator[NativeFunction]:
yield self.functional
yield self.out
if self.inplace is not None:
yield self.inplace
if self.mutable is not None:
yield self.mutable
@property
def root_name(self) -> str:
return self.functional.root_name
@staticmethod
def from_dict(
d: Dict[SchemaKind, NativeFunction]
) -> Optional["NativeFunctionsGroup"]:
assert d
if len(d) == 1:
return None
d = dict(d) # non-destructive updates please
functional = d.pop(SchemaKind.functional, None)
inplace = d.pop(SchemaKind.inplace, None)
mutable = d.pop(SchemaKind.mutable, None)
out = d.pop(SchemaKind.out, None)
assert not d
assert functional is not None
# There are a few operators which only have functional/inplace variants;
# these don't count as structured for our purposes here
if out is None:
return None
# assuming all variants have the same namespace
return NativeFunctionsGroup(
functional=functional,
inplace=inplace,
mutable=mutable,
out=out,
)
@dataclass(frozen=True)
class BackendMetadata:
# The name of the backend kernel, for a given operator
# for in-tree backends. These names come directly from the 'dispatch" field
# in native_functions.yaml. The dispatch entry is optional; in that
# case, that is equivalent to having written:
#
# dispatch:
# CompositeImplicitAutograd: $operator_name
kernel: str
# Whether or not the operator has a structured kernel implemented, for this particular backend.
# For in-tree backends, they all have the same value for structured- this is listed
# in native_functions.yaml.
# However, external backends like XLA can indendently toggle which ops are structured.
structured: bool
# The namespace for kernels, default value: DEFAULT_KERNEL_NAMESPACE
cpp_namespace: str
def supports_symint(self) -> bool:
return "_symint" in self.kernel
@dataclass(frozen=True)
class UfuncInnerLoop:
name: str
supported_dtypes: OrderedSet[ScalarType]
# key is stored here because it affects the semantics of name,
# so its helpful to have them together for further processing
ufunc_key: UfuncKey
@staticmethod
def parse(value: str, ufunc_key: UfuncKey) -> "UfuncInnerLoop":
name, supported_dtypes_str = value.split(" ", 1)
assert supported_dtypes_str[0] == "("
assert supported_dtypes_str[-1] == ")"
supported_dtypes: OrderedSet[ScalarType] = OrderedSet()
for k in supported_dtypes_str[1:-1].split(", "):
supported_dtypes |= ScalarType.parse_set(k)
return UfuncInnerLoop(
name=name, supported_dtypes=supported_dtypes, ufunc_key=ufunc_key
)
# BackendIndex represents a backend.
# The BackendIndex encodes per-operator information that is potentially different
# for each backend. The most obvious example is the name of the kernel
# (the 'dispatch' entry in native_functions.yaml).
# However, there can be other examples of different backends having different information.
# External backends can choose to opt their kernels to be structured independently from in-tree backends,
# which means that this information isn't inherently tied to a NativeFunction- it's different per backend.
@dataclass(frozen=True)
class BackendIndex:
dispatch_key: DispatchKey
# Mainly important for structured kernels, this determines which variant in the operator group is used to implement the others.
# All in-tree ops use out kernels, while XLA uses functional kernels.
use_out_as_primary: bool
# Whether the backend requires a device guard, and device checks.
# For in-tree backends, this is currently just CUDA/HIP
# For out-of-tree backends, this is currently just Intel XPU
device_guard: bool
# Whether the backend is in-tree (CPU/CUDA) or out-of-tree (XLA)
external: bool
# Other backend-specific information that is on a per-operator basis
index: Dict["OperatorName", BackendMetadata]
@staticmethod
def grow_index(
parent_index: Dict[DispatchKey, Dict["OperatorName", BackendMetadata]],
child_index: Dict[DispatchKey, Dict["OperatorName", BackendMetadata]],
) -> None:
for k, v in child_index.items():
for op_name, metadata in v.items():
assert (
op_name not in parent_index[k]
), f"duplicate operator {op_name} for dispatch key {k}"
parent_index[k][op_name] = metadata
def primary(self, g: NativeFunctionsGroup) -> NativeFunction:
if self.use_out_as_primary:
return g.out
else:
return g.functional
def has_kernel(self, g: Union[NativeFunction, NativeFunctionsGroup]) -> bool:
m = self.get_kernel(g)
return m is not None
def get_kernel(
self, g: Union[NativeFunction, NativeFunctionsGroup]
) -> Optional[BackendMetadata]:
if isinstance(g, NativeFunction):
f = g
elif isinstance(g, NativeFunctionsGroup):
f = self.primary(g)
else:
assert_never(g)
if f.func.name not in self.index:
return None
return self.index[f.func.name]
def native_function_class_name(self) -> Optional[str]:
if self.external:
return f"{str(self.dispatch_key)}NativeFunctions"
else:
# TODO: This discrepancy isn't required; we could also generated
# a class for in-tree kernels. It'll just require carefully
# updating every kernel definition + callsite of every in-tree aten kernel.
return None
# The function schema is undoubtedly the most important data structure
# in all of the codegen, as it defines the type signature for operators,
# and most of the code generation we do is type directed (e.g., look at
# the types, decide what to do. Think about how we code generate
# C++ function stubs!)
#
# We will also see in this class the general structure for how we model
# data in this code generation. A few notable properties to point out
# ahead of time:
#
# - These dataclasses are a *lossless* representation of the strings
# they are parsed from. In fact, we assert that given the
# information stored in the dataclass, we can exactly reconstruct
# the string we parsed from (and assert this inside the parse
# definition). There are a few reasons for this:
#
# - If you find that it is difficult to reconstruct the string
# given a dataclass, that is a clue that you are data
# representation is wrong.
#
# - It helps ensure that all relevant information is present
# in the dataclass, so that downstream users aren't tempted
# to reparse the original string to get some information
# that was omitted.
#
# - It forces you to represent the data in-memory in the same way
# it is recorded textually, which makes the dataclasses easier
# to understand for someone who is familiar with the
# textual format. (As a tradeoff, it means you have to model
# the syntax, even when it is inconvenient. But maybe that means
# the syntax is bad!) If you don't understand the internal
# representation, go look at the printing code to see how
# it maps onto the surface syntax!
#
# - It makes it easy to test the parsing code, as parsing code
# that is inconsistent with the string code will fail early
# and loudly. (As a tradeoff, it makes the parsing code a bit
# brittle (in particular, with trivial whitespace changes you
# are likely to trigger an assert error).
#
# In general, try to make the __str__ code as simple as possible
# (even at the cost of more complex parsing logic.) Additionally,
# try to minimize redundancy in data representation. (Precomputed
# fields are OK though: they are defined as a simple function on
# the canonical representation in question.)
#
# - These dataclasses are all frozen; once constructed their
# values never change. This makes it easy to tell where any
# given data came from: just look to the constructor. As a
# tradeoff, you can't easily "decorate" a schema with extra
# information from a post-facto analysis. We impose this
# restriction to make these structures more understandable.
#
@dataclass(frozen=True)
class FunctionSchema:
# The name of the operator this function schema describes.
name: "OperatorName"
arguments: "Arguments"
# TODO: Need to handle collisions with argument names at some point
returns: Tuple["Return", ...]
def schema_order_arguments(self) -> Iterator["Argument"]:
return itertools.chain(
self.arguments.flat_positional,
self.arguments.flat_kwarg_only,
self.arguments.out,
)
decl_re = re.compile(r"(?P<name>[^\(]+)\((?P<args>.*)\) -> (?P<returns>.*)")
@staticmethod
def parse(func: str) -> "FunctionSchema":
# We should probably get a proper parser here
decls = FunctionSchema.decl_re.findall(func)
assert len(decls) == 1, f"Invalid function schema: {func}"
ops, args, return_decl = decls[0]
name = OperatorName.parse(ops)
arguments = Arguments.parse(args)
returns = parse_returns(return_decl)
r = FunctionSchema(name=name, arguments=arguments, returns=returns)
assert str(r) == func, f"{str(r)} != {func}"
return r
def returns_are_aliased(self) -> bool:
# We assert earlier that schemas can't have a mix of aliased and non-aliased returns
return any(
r
for r in self.returns
if r.annotation is not None and r.annotation.is_write
)
def __post_init__(self) -> None:
for arg, ret in zip(self.arguments.out, self.returns):
assert arg.annotation == ret.annotation, (
"Out arguments must have matching return Tensor; furthermore, "
"the ith-argument needs to correspond to the ith return"
)
# We also enforce that if you have any mutable, positional args, then they are not returned.
# This makes it easier to group these functions properly with their functional/out= counterparts.
for a in self.arguments.post_self_positional_mutable:
assert not any(
a.annotation == r.annotation for r in self.returns
), f"If you have a schema with mutable positional args, we expect them to not be returned. schema: {str(self)}"
# Invariant: we expect out arguments to appear as keyword arguments in the schema.
# This means that all mutable returns should be aliased to a keyword argument
# (except for "self", which we explicitly don't treat as an out argument because of its use in methods)
# See Note [is_out_fn]
out_and_self = list(self.arguments.out) + [
arg for arg in self.arguments.flat_positional if arg.name == "self"
]
mutable_returns = [
ret
for ret in self.returns
if ret.annotation is not None and ret.annotation.is_write
]
immutable_returns = [
ret
for ret in self.returns
if ret.annotation is None or not ret.annotation.is_write
]
# Some assertions: We don't want any functions with a return type of "-> (Tensor(a!), Tensor)",
# because:
# (1) It's more annoying to handle properly
# (2) It's unnecessary - you can't method-chain on the first (mutated) output because it's part of a tuple.
# Instead, we expect the (a!) argument to not be returned.
assert (
len(mutable_returns) == 0 or len(immutable_returns) == 0
), f"NativeFunctions must have either only mutable returns, or only immutable returns. Found: {str(self)}"
for ret in mutable_returns:
assert any(ret.annotation == arg.annotation for arg in out_and_self), (
'All mutable returns must be aliased either to a keyword argument, or to "self". '
"Did you forget to mark an out argument as keyword-only?"
)
if self.arguments.out:
# out= ops that return their mutable inputs are only really useful for method chaining.
# And method chaining is only really useful if the thing you're returning is a plain Tensor.
# So ideally, we'd enforce that out= ops with a single plain mutable tensor should return the tensor,
# and all other types of out= op schemas should return void.
# There are a bunch of existing out= ops that return tuples of tensors though, so we're stuck with allowing that.
if any(a.type != BaseType(BaseTy.Tensor) for a in self.arguments.out):
assert (
len(self.returns) == 0
), "out= ops that accept tensor lists as out arguments "
"are expected to have no return type (since you can't do method chaining on them)"
else:
# mutable keyword arguments whose name has _scratch_ prefix are
# scratch tensors for memory planning and should not be returned
assert len(
[
arg
for arg in self.arguments.out
if not arg.name.startswith("_scratch_")
]
) == len(
self.returns
), "Must return as many arguments as there are out arguments, or no return at all"
if self.name.name.inplace:
self_a = self.arguments.self_arg
assert (
self_a
and self_a.argument.annotation
and self_a.argument.annotation.is_write
)
if self_a.argument.type == BaseType(BaseTy.Tensor):
# All inplace ops with an ordinary `Tensor self` argument should return self,
# to allow for method chaining.
assert (
len(self.returns) == 1
and self.returns[0].annotation == self_a.argument.annotation
)
else:
# You can't method chain on non-tensor self arguments though (like a List[Tensor])
# so in all other cases we expect the return type to be none.
assert len(self.returns) == 0
if self.arguments.tensor_options is not None:
assert self.kind() == SchemaKind.functional, (
"Found an operator that is not functional or out variant, but has tensor options arguments."
"This is not allowed- tensor options arguments are only allowed for factory functions."
f"schema: {str(self)}"
)
if self.is_functional_fn():
assert self.kind() == SchemaKind.functional, (
"Found an operator that is not functional, but its overload contains the string 'functional'."
"This is a special keyword in the codegen, please use a different overload name."
f"schema: {str(self)}"
)
def is_functional_fn(self) -> bool:
return "functional" in self.name.overload_name
def is_out_fn(self) -> bool:
# Note [is_out_fn]
#
# out functions are the variants which take an explicit out= argument
# to populate into. We need to know if a schema corresponds to an
# out function for several reasons:
#
# - They codegen differently in C++ API
# - codegen to at::add_out rather than at::add
# - out argument is moved to front of C++ argument list
#
# out functions are DEFINED to be any function with a keyword-only
# argument that is mutable. In principle, this could lead to a
# false positive if you define a function that mutates a
# kwarg only argument, but this isn't the "true" output of this
# function. A more robust definition that would work in this
# case would also look at:
#
# - The output types. Out functions take in the arguments
# they mutate and then return them again; this is sort
# of "definitionally" what makes something an out function.
# Historically, we DO check this for consistency.
# - Correspondence with pure variant. An out function
# should have a signature equivalent to its pure variant,
# but just with extra kwargs for the output elements. This
# is difficult to actually check for and historically
# we only do this check in tools/
return bool(self.arguments.out)
def kind(self) -> SchemaKind:
"""
What kind of schema is this? A functional schema is one
that returns a newly allocated output; an inplace schema
modifies the self argument inplace; an out schema writes
the result into an explicitly provided out argument.
"""
is_out = bool(self.arguments.out)
is_scratch = bool(
[arg for arg in self.arguments.out if arg.name.startswith("_scratch_")]
)
is_inplace = self.name.name.inplace
is_mutable = any(
a.annotation is not None and a.annotation.is_write
for a in self.arguments.post_self_positional
)
assert not (is_out and is_inplace)
# out= and inplace schemas can also have post_self_positional mutable args,
# but we give precedence to out= and inplace when deciding the schema kind.
# Tradeoff: we probably don't want to have to teach codegen that looks at inplace ops
# to also worry about mutable post_self_positional arguments,
# but it seems like a much bigger lift to classify them has having a new schema kind.
# The number of ops that fit in this strange category is small enough that
# we can probably manually write code for them instead of forcing the codegen to handle them.
if is_inplace:
return SchemaKind.inplace
elif is_scratch:
assert (
is_out
), "invariant: all scratch operators are expected to be out= operators too"
return SchemaKind.scratch
elif is_out:
assert (
not is_scratch
), "We should not categorize a scratch op as an out variant. Check if the order of if statements are expected!"
return SchemaKind.out
elif is_mutable:
return SchemaKind.mutable
else:
return SchemaKind.functional
# For every return:
# - If the return aliases an input, we return the input name
# - Otherwise, we return None.
# If return names were enforced to be consistent with aliasing information, then we wouldn't need this.
def aliased_return_names(self) -> List[Optional[str]]:
outs: List[Optional[str]] = []
for r in self.returns:
aliased_args = [
a
for a in self.arguments.flat_all
if a.annotation is not None and a.annotation == r.annotation
]
if len(aliased_args) == 0:
outs.append(None)
elif len(aliased_args) == 1:
outs.append(aliased_args[0].name)
else:
aliased_names = ", ".join(a.name for a in aliased_args)
raise AssertionError(
f"Found a return ({r.name})that aliases multiple inputs ({aliased_names})"
)
return outs
def signature(
self,
*,
strip_default: bool = False,
strip_view_copy_name: bool = False,
keep_return_names: bool = False,
) -> "FunctionSchema":
"""
Certain schemas are 'related', in that they are simply
inplace/out/functional versions of the same function. This method
factors these schemas into the "core" functional signature which
is equal across all versions.
Here is what normalization happens to the schema to convert
it to a signature:
- The overload name is stripped (name is retained, since
it expresses semantic content about what the function does)
- Inplace is set False
- Out arguments are stripped
- Mutable post_self_positional args are converted to returns
- Mutability annotations are stripped (this is sound
because you cannot overload on mutability annotation)
- Return names are stripped since they are not overloadable and
some variants have return names but some not
- TensorOptions are dropped
because out= variants of factory functions don't include them
(and we want to be able to pair up factory functions with their out variants)
Finally, we want to be able to pair up related "view" and their
corresponding "view_copy" operators. We do this by optionally
stripping the trailing "_copy" from the base name.
Example of a mutable op before and after:
f.func (Mutable operator):
_fused_moving_avg_obs_fq_helper(Tensor self, Tensor observer_on, Tensor fake_quant_on, Tensor(a!) running_min, Tensor(b!) running_max, Tensor(c!) scale, Tensor(d!) zero_point, float averaging_const, int quant_min, int quant_max, int ch_axis, bool per_row_fake_quant=False, bool symmetric_quant=False) -> (Tensor output, Tensor mask) # noqa: B950
f.func (Corresponding functional operator):
_fused_moving_avg_obs_fq_helper.functional(Tensor self, Tensor observer_on, Tensor fake_quant_on, Tensor running_min, Tensor running_max, Tensor scale, Tensor zero_point, float averaging_const, int quant_min, int quant_max, int ch_axis, bool per_row_fake_quant=False, bool symmetric_quant=False) -> (Tensor output, Tensor mask, Tensor running_min_out, Tensor running_max_out, Tensor scale_out, Tensor zero_point_out) # noqa: B950
f.func.signature() output:
_fused_moving_avg_obs_fq_helper(Tensor self, Tensor observer_on, Tensor fake_quant_on, Tensor running_min, Tensor running_max, Tensor scale, Tensor zero_point, float averaging_const, int quant_min, int quant_max, int ch_axis, bool per_row_fake_quant=False, bool symmetric_quant=False) -> (Tensor, Tensor, Tensor, Tensor, Tensor, Tensor) # noqa: B950
"""
def strip_ret_annotation(r: Return) -> Return:
return Return(
name=r.name if keep_return_names else None,
type=r.type,
annotation=None,
)
base_name = self.name.name.base
if strip_view_copy_name:
if base_name.endswith("_copy"):
base_name = base_name.replace("_copy", "")
elif base_name.endswith("_scatter"):
base_name = base_name.replace("scatter", "inverse")
# find mutable inputs that are not originally returned, and convert them to returns
returns_from_mutable_inputs = tuple(
# When we're grouping functions we strip the return names,
# but when we're generating the actual functional variants then we follow
# a convention for what to name the returns
Return(
name=f"{a.name}_out" if keep_return_names else None,
type=a.type,
annotation=None,
)
for a in itertools.chain(
# Order is important here (otherwise e.g. inplace with mutable args
# and out= with mutable args won't have the same signature)
[self.arguments.self_arg.argument]
if self.arguments.self_arg is not None
else [],
self.arguments.out,
self.arguments.post_self_positional,
)
if a.annotation is not None
and a.annotation.is_write
and not any(a.annotation == r.annotation for r in self.returns)
)
original_returns = tuple(map(strip_ret_annotation, self.returns))
# Ordering is important here. We expect the "mutable input" returns to come last.
returns = original_returns + returns_from_mutable_inputs
args_sig = self.arguments.signature(strip_default=strip_default)
# See Note [bernoulli.p schema]
if str(self.name) == "bernoulli.p":
args_sig = Arguments.parse(str(args_sig).replace("float p", "float p=0.5"))
return FunctionSchema(
name=OperatorName(
name=BaseOperatorName(
base=base_name,
inplace=False,
dunder_method=self.name.name.dunder_method,
),
overload_name="", # stripped
),
arguments=args_sig,
returns=returns,
)
def view_signature(self) -> "FunctionSchema":
return self.signature(strip_view_copy_name=True)
def with_name(self, name: "OperatorName") -> "FunctionSchema":
return FunctionSchema(
name=name,
arguments=self.arguments,
returns=self.returns,
)
@property
def modifies_arguments(self) -> bool:
return self.kind() in [SchemaKind.inplace, SchemaKind.out, SchemaKind.mutable]
def has_symint(self) -> bool:
return self.arguments.has_symint_arg()
def __str__(self) -> str:
all_arguments_str = str(self.arguments)
if len(self.returns) == 1:
returns = str(self.returns[0]) # omit parentheses
else:
returns = "(" + ", ".join(map(str, self.returns)) + ")"
return f"{self.name}({all_arguments_str}) -> {returns}"
# Here is the rest of the data model, described more briefly.
# Simplified version for what actually shows up in built-ins.
# Look at alias_info.h for expanded syntax. If you need the structure,
# you also need to make this structure recursive so it can be lined
# up with the type components too. For primitives this isn't really
# necessary
@dataclass(frozen=True)
class Annotation:
# Typically only has one element. Not actually a set so
# we can conveniently assume it is canonically ordered
alias_set: Tuple[str, ...]
is_write: bool
alias_set_after: Tuple[str, ...]
@staticmethod
def parse(ann: str) -> "Annotation":
# TODO: implement a proper parser if this gets more ugly
# Regex Explanation:
# Example: "a! -> a|b"
# Group #1: alias before optional '|', required. Matches the first
# character 'a' in the example
# Group #2: optional alias set after optional '|', matches empty string
# in the example
# Group #3: optional "is write" flag, matches '!' in the example.
# Group #4: optional section containing arrow, matches " -> a|b" in the
# example.
# Group #5: optional alias after set, supports wildcard, matches "a|b"
# in the example.
# Group #6: optional sub-section of alias after set, matches "|b" in the
# example.
m = re.match(r"^([a-z])(\|[a-z])*(!?)( -> (\*|[a-z](\|[a-z])*))?$", ann)
assert m is not None, f"unrecognized alias annotation {ann}"
before_alias = m.group(1) + (m.group(2) if m.group(2) else "")
alias_set = tuple(before_alias.split("|"))
is_write = m.group(3) == "!"
assert not (
is_write and len(alias_set) > 1
), f"alias set larger than 1 is not mutable, got {ann} instead."
after_set = tuple(m.group(5).split("|")) if m.group(5) else tuple()
assert not (
len(before_alias) > 1 and len(after_set) > 1
), f"before alias set and after alias set cannot be larger than 1 at the same time, got {ann} instead."
r = Annotation(
alias_set=alias_set, is_write=is_write, alias_set_after=after_set
)
assert str(r) == ann, f"{r} != {ann}"
return r
def __str__(self) -> str:
alias_set = "|".join(self.alias_set)
if self.is_write:
alias_set = f"{alias_set}!"
alias_set_after = "|".join(self.alias_set_after)
if alias_set_after:
alias_set = f'{alias_set}{" -> "}{alias_set_after}'
return alias_set
# The base class for the type system. This is also loosely modeled
# off of jit_type.h, but we've simplified the hierarchy to focus
# in on the aspects of the type system that matter for code generation
# (for example, there's no SingleElementType subclass anymore).
# You never actually construct a Type; usually it's going to be one
# of the subclasses. If Python had ADTs this would be one!
@dataclass(frozen=True)
class Type:
@staticmethod
def parse(t: str) -> "Type":
r = Type._parse(t)
assert str(r) == t, f"{r} != {t}"
return r
@staticmethod
def _parse(t: str) -> "Type":
m = re.match(r"^(.+)\?$", t)
if m is not None:
return OptionalType(Type.parse(m.group(1)))
m = re.match(r"^(.+)\[([0-9]+)?\]$", t)
if m is not None:
size = int(m.group(2)) if m.group(2) is not None else None
return ListType(elem=Type.parse(m.group(1)), size=size)
# '__torch__.torch.classes.' is the prefix for custom class
m = re.match(r"^__torch__\.torch\.classes\.([a-zA-Z0-9_.]+)$", t)
if m is not None:
return CustomClassType(m.group(1))
try:
return BaseType(BaseTy[t])
except KeyError as e:
raise RuntimeError(f"unrecognized type {t}") from e
def __str__(self) -> str:
raise NotImplementedError
# WARNING: These concepts are not very well-defined. For example,
# is "int?" nullable? How about "int?[]". They are defined
# so we can conveniently generate legacy Declarations.yaml but
# really we should probably just remove these at some point
def is_base_ty_like(self, base_ty: "BaseTy") -> bool:
raise NotImplementedError
def is_tensor_like(self) -> bool:
return self.is_base_ty_like(BaseTy.Tensor)
def is_generator_like(self) -> bool:
return self.is_base_ty_like(BaseTy.Generator)
def is_symint_like(self) -> bool:
return self.is_base_ty_like(BaseTy.SymInt)
def is_nullable(self) -> bool:
raise NotImplementedError
def is_list_like(self) -> Optional["ListType"]:
raise NotImplementedError
# Base types are simple, atomic types with no further structure
class BaseTy(Enum):
Generator = auto()
ScalarType = auto()
Tensor = auto()
int = auto()
Dimname = auto()
DimVector = auto()
float = auto()
str = auto()
bool = auto()
Layout = auto()
Device = auto()
DeviceIndex = auto()
Scalar = auto()
MemoryFormat = auto()
QScheme = auto()
Storage = auto()
Stream = auto()
SymInt = auto()
ConstQuantizerPtr = auto() # TODO: rename
@dataclass(frozen=True)
class BaseType(Type):
name: BaseTy
def __str__(self) -> str:
return f"{self.name.name}"
def is_base_ty_like(self, base_ty: BaseTy) -> bool:
return self.name == base_ty
def is_nullable(self) -> bool:
return False
def is_list_like(self) -> Optional["ListType"]:
return None
def is_symint_like(self) -> bool:
return self.name == BaseTy.SymInt
# Optional types may be specified, or may also be validly given None
@dataclass(frozen=True)
class OptionalType(Type):
elem: Type
def __str__(self) -> str:
return f"{self.elem}?"
def is_base_ty_like(self, base_ty: BaseTy) -> bool:
return self.elem.is_base_ty_like(base_ty)
def is_symint_like(self) -> bool:
return self.elem.is_symint_like()
def is_nullable(self) -> bool:
return True
def is_list_like(self) -> Optional["ListType"]:
return self.elem.is_list_like()
# A type representing a PyTorch custom class
@dataclass(frozen=True)
class CustomClassType(Type):
class_name: str
def __str__(self) -> str:
"""
Return the class name will prefix __torch__.torch.classes
"""
return f"__torch__.torch.classes.{self.class_name}"
def is_base_ty_like(self, base_ty: BaseTy) -> bool:
return False
def is_symint_like(self) -> bool:
return False
def is_nullable(self) -> bool:
"""
Assume a custom class is not nullable.
"""
return False
def is_list_like(self) -> Optional["ListType"]:
return None
# List types specify that we may have multiples of an element. We
# also support explicit sizes on list types, but these have
# some nontrivial semantics! (However, for C++ API purposes, explicit
# sizes are mostly erased from the type system.)
#
# DANGER WILL ROBINSON: C++ elaboration depends on elem type; e.g.,
# int[] elaborates differently than bool[3]!
@dataclass(frozen=True)
class ListType(Type):
elem: Type
size: Optional[int]
def __str__(self) -> str:
size = f"{self.size}" if self.size else ""
return f"{self.elem}[{size}]"
def is_base_ty_like(self, base_ty: BaseTy) -> bool:
return self.elem.is_base_ty_like(base_ty)
def is_symint_like(self) -> bool:
return self.elem.is_symint_like()
def is_nullable(self) -> bool:
return self.elem.is_nullable()
def is_list_like(self) -> Optional["ListType"]:
return self
@dataclass(frozen=True)
class Argument:
# NB: I didn't put kwarg_only as a boolean field here, unlike
# c10::Argument, so that printing works correctly
name: str
type: Type
default: Optional[str]
# The semantics of the annotation field are a little strange.
#
# Alias annotations parametrize Tensors (since Tensors are the only things
# that can alias.) This motivates why I write Tensor(a!)? (and not, for
# example, Tensor?(a!)), because the (a!) describes aliasing on the tensor,
# which may be optional (i.e., the alias annotation should bind first to
# Tensor, before the optional postfix annotation).
#
# However, despite being a property of Tensor, we (and c10::Argument)
# store the annotation at the top level of the Argument, rather than
# inside the embedded Tensor type. In the C++ version of this
# class, we then go through great lengths to mimic the type
# structure in the annotation structure so we can correlate
# annotations with types.
#
# Now, it turns out, in all applications in code generation, the
# structure of annotated types is very simple. So we just hard
# code it here. But if we ever do get anything more complex, this
# model will have to change!
annotation: Optional[Annotation]
@staticmethod
def parse(arg: str) -> "Argument":
name: str
default: Optional[str]
type_and_annot, name_and_default = arg.rsplit(" ", 1)
if "=" in name_and_default:
name, default = name_and_default.split("=")
else:
name = name_and_default
default = None
# TODO: deduplicate annotation matching with Return
match = re.match(r"Tensor\((.+)\)(.*)", type_and_annot)
annotation: Optional[Annotation]
if match:
# If you update this, make sure the __str__ still works too
assert match.group(2) in [
"",
"?",
"[]",
], "unrecognized alias analysis form with Tensor"
type_s = "Tensor" + match.group(2)
annotation = Annotation.parse(match.group(1))
else:
type_s = type_and_annot
annotation = None
type = Type.parse(type_s)
r = Argument(
name=name,
type=type,
default=default,
annotation=annotation,
)
assert str(r) == arg, f"{str(r)} != {arg}"
return r
@property
def is_write(self) -> bool:
return self.annotation is not None and self.annotation.is_write
def __str__(self) -> str:
type = f"{self.type}"
if self.annotation:
assert type in ["Tensor", "Tensor?", "Tensor[]"]
type = type.replace("Tensor", f"Tensor({self.annotation})")
if self.name is None:
return type
else:
mb_default = ""
if self.default:
mb_default = f"={self.default}"
return f"{type} {self.name}{mb_default}"
@dataclass(frozen=True)
class Return:
name: Optional[str]
type: Type
annotation: Optional[Annotation]
@staticmethod
def parse(arg: str) -> "Return":
name: Optional[str]
if " " in arg:
type_and_annot, name = arg.rsplit(" ", 1)
else:
type_and_annot = arg
name = None
match = re.match(r"Tensor\((.+)\)(.*)", type_and_annot)
annotation: Optional[Annotation]
if match:
# If you update this, make sure the __str__ still works too
assert match.group(2) in [
"",
"?",
"[]",
], "unrecognized alias analysis form with Tensor"
type_s = "Tensor" + match.group(2)
annotation = Annotation.parse(match.group(1))
else:
type_s = type_and_annot
annotation = None
type = Type.parse(type_s)
r = Return(
name=name,
type=type,
annotation=annotation,
)
assert str(r) == arg, f"{str(r)} != {arg}"
return r
@property
def is_write(self) -> bool:
return self.annotation is not None and self.annotation.is_write
def __str__(self) -> str:
type = f"{self.type}"
if self.annotation:
assert type in ["Tensor", "Tensor?", "Tensor[]"]
type = type.replace("Tensor", f"Tensor({self.annotation})")
if self.name is None:
return type
else:
return f"{type} {self.name}"
# Represents the self argument for functions that may be methods
@dataclass(frozen=True)
class SelfArgument:
argument: Argument
# Bundle of arguments that represent a TensorOptions. This is mostly
# relevant for the public C++ API but we bake it into the core data
# model because other APIs often have to interact with it
@dataclass(frozen=True)
class TensorOptionsArguments:
dtype: Argument
layout: Argument
device: Argument
pin_memory: Argument
def all(self) -> Sequence[Argument]:
return [self.dtype, self.layout, self.device, self.pin_memory]
@dataclass(frozen=True)
class Arguments:
# pre_self_positional is usually empty, but is notably non-empty
# for where.self, where the condition argument comes before the
# self argument
pre_self_positional: Tuple[Argument, ...]
self_arg: Optional[SelfArgument]
post_self_positional: Tuple[Argument, ...]
pre_tensor_options_kwarg_only: Tuple[Argument, ...]
tensor_options: Optional[TensorOptionsArguments]
# post_tensor_options is typically memory format, which should be
# part of tensor options but isn't right now, and is usually
# placed after the tensor options arguments
post_tensor_options_kwarg_only: Tuple[Argument, ...]
# Unlike in the previous codegen, we have factored out 'out' arguments
# in the canonical representation, removing them from kwarg
# arguments. This choice is justified by numerous downstream
# transformations which treat out arguments specially; additionally,
# you can see that canonicity is not violated!
out: Tuple[Argument, ...] # these are also kwarg-only
@property
def flat_non_out(self) -> Sequence[Argument]:
ret: List[Argument] = []
ret.extend(self.flat_positional)
ret.extend(self.flat_kwarg_only)
return ret
@property
def flat_positional(self) -> Sequence[Argument]:
ret: List[Argument] = []
ret.extend(self.pre_self_positional)
if self.self_arg is not None:
ret.append(self.self_arg.argument)
ret.extend(self.post_self_positional)
return ret
@property
def post_self_positional_mutable(self) -> Sequence[Argument]:
return [a for a in self.post_self_positional if a.is_write]
# NB: doesn't contain out arguments
@property
def flat_kwarg_only(self) -> Sequence[Argument]:
ret: List[Argument] = []
ret.extend(self.pre_tensor_options_kwarg_only)
if self.tensor_options is not None:
ret.extend(self.tensor_options.all())
ret.extend(self.post_tensor_options_kwarg_only)
return ret
@property
def flat_all(self) -> Sequence[Argument]:
ret: List[Argument] = []
ret.extend(self.flat_positional)
ret.extend(self.flat_kwarg_only)
ret.extend(self.out)
return ret
@property
def non_out(
self,
) -> Sequence[Union[Argument, SelfArgument, TensorOptionsArguments]]:
ret: List[Union[Argument, SelfArgument, TensorOptionsArguments]] = []
ret.extend(self.positional)
ret.extend(self.kwarg_only)
return ret
@property
def positional(self) -> Sequence[Union[Argument, SelfArgument]]:
ret: List[Union[Argument, SelfArgument]] = []
ret.extend(self.pre_self_positional)
if self.self_arg is not None:
ret.append(self.self_arg)
ret.extend(self.post_self_positional)
return ret
@property
def kwarg_only(self) -> Sequence[Union[Argument, TensorOptionsArguments]]:
ret: List[Union[Argument, TensorOptionsArguments]] = []
ret.extend(self.pre_tensor_options_kwarg_only)
if self.tensor_options is not None:
ret.append(self.tensor_options)
ret.extend(self.post_tensor_options_kwarg_only)
return ret
@property
def all(self) -> Sequence[Union[Argument, SelfArgument, TensorOptionsArguments]]:
ret: List[Union[Argument, SelfArgument, TensorOptionsArguments]] = []
ret.extend(self.positional)
ret.extend(self.kwarg_only)
ret.extend(self.out)
return ret
def mutable_arg_names(self) -> List[str]:
return [
a.name
for a in self.flat_all
if a.annotation is not None and a.annotation.is_write
]
def has_tensor_arg(self) -> bool:
return any(a.type.is_tensor_like() for a in self.flat_non_out)
def has_symint_arg(self) -> bool:
return any(a.type.is_symint_like() for a in self.flat_non_out)
def has_generator_arg(self) -> bool:
return any(a.type.is_generator_like() for a in self.flat_non_out)
def signature(self, *, strip_default: bool = False) -> "Arguments":
# dataclasses.replace could be used here, but it is less
# type safe so for now I've opted to type everything out
def strip_arg_annotation(a: Argument) -> Argument:
return Argument(
name=a.name,
type=a.type,
default=a.default if not strip_default else None,
annotation=None,
)
return Arguments(
pre_self_positional=tuple(
map(strip_arg_annotation, self.pre_self_positional)
),
self_arg=SelfArgument(strip_arg_annotation(self.self_arg.argument))
if self.self_arg is not None
else None,
post_self_positional=tuple(
map(strip_arg_annotation, self.post_self_positional)
),
# Since TensorOptions are dropped, the post_tensor_options_kwargs are
# converted to pre_tensor_options_kwargs
pre_tensor_options_kwarg_only=tuple(
map(strip_arg_annotation, self.pre_tensor_options_kwarg_only)
)
+ tuple(map(strip_arg_annotation, self.post_tensor_options_kwarg_only)),
# TensorOptions are dropped in signature,
# so we can pair factory functions with their out= variants.
tensor_options=None,
post_tensor_options_kwarg_only=tuple(),
# out arguments are dropped in signature
out=(),
)
def remove_self_annotation(self) -> "Arguments":
assert self.self_arg is not None
return dataclasses.replace(
self,
self_arg=SelfArgument(
dataclasses.replace(self.self_arg.argument, annotation=None)
),
)
def with_out_args(self, outs: List[Argument]) -> "Arguments":
assert len(self.out) == 0
return dataclasses.replace(
self,
out=tuple(outs),
)
@staticmethod
def _preparse(args: str) -> Tuple[List[Argument], List[Argument], List[Argument]]:
positional: List[Argument] = []
kwarg_only: List[Argument] = []
out: List[Argument] = []
arguments_acc = positional
# TODO: Use a real parser here; this will get bamboozled
# by signatures that contain things like std::array<bool, 2> (note the space)
for arg in args.split(", "):
if not arg:
continue
if arg == "*":
assert (
arguments_acc is positional
), "invalid syntax: kwarg-only specifier * can only occur once"
arguments_acc = kwarg_only
continue
parg = Argument.parse(arg)
# Currently, we rely directly on the invariant that there are NO
# kwarg-only mutating arguments. If you want to relax this,
# we will need a more semantic way of matching that takes
# into account return arguments. In that case, you will have
# to manage out computation a level up, in FunctionSchema. See Note
# [is_out_fn]
if parg.annotation is not None and parg.annotation.is_write:
if arguments_acc is positional:
pass # do nothing
elif arguments_acc is kwarg_only:
arguments_acc = out
else:
assert arguments_acc is not out
arguments_acc.append(parg)
return positional, kwarg_only, out
@staticmethod
def parse(args: str) -> "Arguments":
"""
Input: 'int x, int y, int z'
"""
# We do this in two phases. First we parse into three
# main categories: positional, kwarg_only, out.
# Then, we reparse positional and kwarg_only to separate
# out the self argument and tensor options arguments.
positional, kwarg_only, out = Arguments._preparse(args)
# Split self argument
self_ix = None
for i, a in enumerate(positional):
if a.name == "self":
self_ix = i
break
pre_self_positional: List[Argument]
self_arg: Optional[SelfArgument]
post_self_positional: List[Argument]
if self_ix is not None:
pre_self_positional = positional[:self_ix]
self_arg = SelfArgument(positional[self_ix])
post_self_positional = positional[self_ix + 1 :]
else:
pre_self_positional = []
self_arg = None
post_self_positional = positional
# Group tensor options arguments
pre_tensor_options_kwarg_only: List[Argument] = []
tensor_options: Optional[TensorOptionsArguments] = None
post_tensor_options_kwarg_only: List[Argument] = []
kwarg_only_acc = pre_tensor_options_kwarg_only
def pred(name: str, ty: Type) -> Callable[[Argument], bool]:
return lambda a: a.name == name and a.type in [ty, OptionalType(ty)]
predicates = [ # order matters
pred("dtype", Type.parse("ScalarType")),
pred("layout", Type.parse("Layout")),
pred("device", Type.parse("Device")),
pred("pin_memory", Type.parse("bool")),
]
i = 0
while i < len(kwarg_only):
# If there is enough space...
if i <= len(kwarg_only) - len(predicates):
# And the next len(predicates) arguments look like TensorOptions arguments
if all(
p(a)
for p, a in zip(predicates, kwarg_only[i : i + len(predicates)])
):
assert kwarg_only_acc is pre_tensor_options_kwarg_only
# Group them together as one argument
tensor_options = TensorOptionsArguments(
dtype=kwarg_only[i],
layout=kwarg_only[i + 1],
device=kwarg_only[i + 2],
pin_memory=kwarg_only[i + 3],
)
i += len(predicates)
kwarg_only_acc = post_tensor_options_kwarg_only
continue
kwarg_only_acc.append(kwarg_only[i])
i += 1
return Arguments(
pre_self_positional=tuple(pre_self_positional),
self_arg=self_arg,
post_self_positional=tuple(post_self_positional),
pre_tensor_options_kwarg_only=tuple(pre_tensor_options_kwarg_only),
tensor_options=tensor_options,
post_tensor_options_kwarg_only=tuple(post_tensor_options_kwarg_only),
out=tuple(out),
)
def __str__(self) -> str:
all_arguments: List[str] = []
all_arguments.extend(map(str, self.flat_positional))
if self.flat_kwarg_only or self.out:
all_arguments.append("*")
all_arguments.extend(map(str, self.flat_kwarg_only))
all_arguments.extend(map(str, self.out))
return ", ".join(all_arguments)
def __post_init__(self) -> None:
# TODO: These invariants are weirdly asymmetric?
# TODO: Fancier types?
if self.self_arg is None:
assert not self.pre_self_positional
if self.tensor_options is None:
assert not self.post_tensor_options_kwarg_only
# We don't allow any of the following to have argument annotations,
# to keep things simple.
mutable_pre_self_positionals = [
a
for a in self.pre_self_positional
if a.annotation is not None and a.annotation.is_write
]
assert (
len(mutable_pre_self_positionals) == 0
), "mutable pre_self_positional arguments are not currently supported in the schema"
# Names that validly are __iXXX__ indicating inplace operations.
# Taken from https://www.python.org/dev/peps/pep-0203/#new-methods
# NB: PyTorch hasn't actually implemented all of these
AUGMENTED_ASSIGNMENT_NAMES = [
"add",
"sub",
"mul",
"div",
"mod",
"pow",
"lshift",
"rshift",
"and",
"xor",
"or",
]
# A BaseOperatorName is what we think of the operator name, without
# the overload name. Unusually, we don't represent this as just a
# string; instead, we directly represent a few important semantic
# bits of information we derive from the string: namely whether
# or not it's inplace (add_) and whether or not it's a double-underscore
# method (__add__)
@dataclass(frozen=True)
class BaseOperatorName:
base: str
inplace: bool
dunder_method: bool
# Note [Overload Ambiguity With Functional Variants]
# A handful of operators have both a "mutable" and a "functional" variant.
# (native_batch_norm is a good example, although this isn't the case today).
# For those operators, the mutable and functional variant take in the same set of
# arguments, but have different alias annotations.
# this makes it ambiguous when you try to resolve an OverloadPacket into an overload,
# given a set of input arguments.
#
# So instead of making the "functional" variant in this case a real overload, e.g:
# native_batch_norm (mutable variant)
# native_batch_norm.functional (functional variant)
# we make it a new base operator,
# native_batch_norm_functional (functional variant)
#
# In an ideal world, we would probably invert this so the operators were:
# native_batch_norm.mutable (mutable variant)
# native_batch_norm (functional variant)
#
# Doing that is BC-breaking though, so we're stuck with the above modeling.
functional_overload: bool = False
@staticmethod
def parse(op: str) -> "BaseOperatorName":
assert op != ""
assert not op.endswith("_out"), (
"_out suffix is reserved and not permitted for operator names; "
"did you mean to specify an out overload name instead?"
)
m = re.match(r"^__([^_]+)__$", op)
if m is not None:
dunder_method = True
base = m.group(1)
if any(base == f"i{n}" for n in AUGMENTED_ASSIGNMENT_NAMES):
inplace = True
base = base[1:]
else:
inplace = False
# temporary, this is not intrinsically true but
# has been historically true for dunder methods
# we support (but, if we ever got, say, __int__, this would
# be wrong!)
assert base[0] != "i"
else:
dunder_method = False
base = op
if base[-1] == "_":
inplace = True
base = base[:-1]
else:
inplace = False
# See Note [Overload Ambiguity With Functional Variants]
functional_suffix = "_functional"
if base.endswith(functional_suffix):
functional_overload = True
base = base[: -len(functional_suffix)]
# This seems complicated and unnecessary, so banning dunder methods
# for now on ops that have a functional + mutable variant (like native_batch_norm).
assert not dunder_method and not inplace
else:
functional_overload = False
r = BaseOperatorName(
base=base,
inplace=inplace,
dunder_method=dunder_method,
functional_overload=functional_overload,
)
assert str(r) == op, f"{str(r)} != {op}"
return r
def __str__(self) -> str:
if self.dunder_method:
i = "i" if self.inplace else ""
return f"__{i}{self.base}__"
else:
i = (
"_"
if self.inplace
else "_functional"
if self.functional_overload
else ""
)
return f"{self.base}{i}"
# Operator name is the base operator name along with the (typically not
# user visible) overload string.
@dataclass(frozen=True)
class OperatorName:
name: BaseOperatorName
overload_name: str
@staticmethod
def parse(op_name: str) -> "OperatorName":
if "." in op_name:
name, overload_name = op_name.split(".", 1)
else:
name = op_name
overload_name = ""
r = OperatorName(name=BaseOperatorName.parse(name), overload_name=overload_name)
assert str(r) == op_name, f"{str(r)} != {op_name}"
return r
def __str__(self) -> str:
if self.overload_name:
return f"{self.name}.{self.overload_name}"
else:
return f"{self.name}"
# NB: This must be synchronized with the naming scheme in
# aten/src/ATen/templates/Operators.h
# Given a function schema "aten::op.overload(...)",
# If there is no overload name, this returns f"{op}"
# If there is an overload name, this returns f"{op}_{overload}"
def unambiguous_name(self) -> str:
if self.overload_name:
return f"{self.name}_{self.overload_name}"
else:
return f"{self.name}"
def remove_inplace(self) -> "OperatorName":
return OperatorName(
name=BaseOperatorName(
base=self.name.base,
inplace=False,
dunder_method=self.name.dunder_method,
),
overload_name=self.overload_name,
)
def with_overload(self, overload: str) -> "OperatorName":
return OperatorName(
name=BaseOperatorName(
base=self.name.base,
inplace=False,
dunder_method=self.name.dunder_method,
),
overload_name=overload,
)
def gets_generated_out_inplace_wrapper(
f: NativeFunction, g: NativeFunctionsGroup, b: BackendIndex
) -> bool:
return (
f.func.kind() is not SchemaKind.functional
and not b.has_kernel(f)
and b.has_kernel(g.functional)
)
# NativeFunction objects that are views (f.is_view_op returns True)
# are added into a `NativeFunctionsViewGroup`, which we can use to
# easily access the generated (optional) view_copy NativeFunction.
# It's convenient to group them together, so we pair them up in NativeFunctionsViewGroup.
# See Note [Codegen'd {view}_copy Operators]
#
# One property of this representation is that in order for a view-like op to be part of
# a NativeFunctionsViewGroup, the "aliasing" version of that view op must exist.
# There's one case where that doesn't happen: we have a non-aliasing `narrow_copy.out` op,
# but don't have corresponding aliasing `narrow.out` op.
# This means that `narrow_copy.out` won't appear as a NativeFunctionsViewGroup.
@dataclass(frozen=True)
class NativeFunctionsViewGroup:
view: NativeFunction
# Note: the {view}_copy operator is optional because we currently don't generate copy variants
# for all view ops. Notably, we don't generate them for CompositeImplicitAutograd views
# (we already get them "for free" through decomposition)
view_copy: Optional[NativeFunction]
# view_inplace ops are also optional, but every view_inplace op should have out-of-place variant.
view_inplace: Optional[NativeFunction]
def __post_init__(self) -> None:
assert self.view.is_view_op
if self.view_copy is None:
assert not gets_generated_view_copy(self.view), (
f"{str(self.view.func.name)} appears to be a new operator that aliases its inputs."
" The codegen expects you to add a corresponding operator to native_functions.yaml:"
f" {get_view_copy_name(self.view)!s}."
" See Note [view_copy NativeFunctions] for details."
)
else:
assert self.view_copy.func.name.name.base.endswith(("_copy", "_scatter"))
assert self.view.func.signature() == self.view_copy.func.signature(
strip_view_copy_name=True,
)
assert "view_copy" in self.view_copy.tags, (
f"{str(self.view_copy.func.name), str(self.view.tags)} appears to be a view_copy operator. The codegen expects"
" view_copy operators to be annotated with the 'view_copy' tag in native_functions.yaml."
" See Note [view_copy NativeFunction] for details."
)
if self.view_inplace is not None:
assert self.view.func.signature() == self.view_inplace.func.signature()
if self.view.has_composite_implicit_autograd_kernel:
if self.view_inplace is not None:
assert self.view_inplace.has_composite_implicit_autograd_kernel, (
f"{str(self.view.func.name)} and {str(self.view_inplace.func.name)} must either"
" both have CompositeImplicitAutograd kernels, or both not have composite kernels."
)
if self.view.has_composite_implicit_autograd_nested_tensor_kernel:
if self.view_inplace is not None:
assert (
self.view_inplace.has_composite_implicit_autograd_nested_tensor_kernel
), (
f"{str(self.view.func.name)} and {str(self.view_inplace.func.name)} must either"
" both have CompositeImplicitAutogradNestedTensor kernels, or both not have composite kernels."
)
def functions(self, *, include_copy: bool = True) -> Iterator[NativeFunction]:
yield self.view
if self.view_inplace is not None:
yield self.view_inplace
if self.view_copy is not None and include_copy:
yield self.view_copy
@property
def root_name(self) -> str:
return self.view.root_name
@property
def composite(self) -> bool:
# We currently assert that the "group" is consistent.
# If the view op is composite, then its view_inplace op is too.
return self.view.has_composite_implicit_autograd_kernel
def gets_generated_view_copy(f: NativeFunction) -> bool:
# Only aliasing (view) operators get a copy variant.
if not f.is_view_op:
return False
# We don't need to bother generating copy variants for CompositeImplicitAutograd ops,
# because we can let them decompose into base view ops.
if f.has_composite_implicit_autograd_kernel:
return False
# We also don't need to generate copy variants for inplace views.
if "inplace_view" in f.tags:
return False
# Assume ops ending in _inverse have manually-defined copy variants
# (e.g. slice_inverse() has the copy variant slice_scatter()).
# We -could- probably generate these as well, but the codegen will be
# slightly different, and hand-writing these few kernels keeps codegen
# complexity lower.
if f.func.name.name.base.endswith("_inverse"):
return False
return True
# Given a NativeFunction that corresponds to a view op,
# returns the OperatorName of the corresponding "copy" variant of the op.
def get_view_copy_name(f: NativeFunction) -> "OperatorName":
# Right now, when asking for a view op's corresponding "view_copy" name
# we assert for sanity that the op is allowed to have a generated view_copy variant.
# (We can do this because "gets_generated_view_copy()" tell us which ops get a generated view_copy op).
# However, narrow_copy() already exists as an op directly in native_functions.yaml.
# I'm hardcoding narrow_copy here for now to maintain the assert,
# But we could also just get rid of the assert.
list_of_ops_with_explicit_view_copy_operators = ["narrow"]
if str(f.func.name) not in list_of_ops_with_explicit_view_copy_operators:
assert gets_generated_view_copy(f)
base_name = f"{f.func.name.name.base}_copy"
view_copy_name = OperatorName(
name=BaseOperatorName(
base=base_name, inplace=False, dunder_method=f.func.name.name.dunder_method
),
overload_name=f.func.name.overload_name,
)
return view_copy_name
# Helper functions for parsing argument lists (both inputs and returns)
def parse_returns(return_decl: str) -> Tuple[Return, ...]:
"""
Input: '()'
Output: []
"""
if return_decl == "()":
return ()
if return_decl[0] == "(" and return_decl[-1] == ")":
return_decl = return_decl[1:-1]
return tuple(Return.parse(arg) for arg in return_decl.split(", "))
# A Precompute instance consists of a map from kernel argument name
# to the list of Argument instances that should replace that
# kernel argument in the impl function.
@dataclass(frozen=True)
class Precompute:
# A map from kernel argument name -> a list of precomputed
# elements that replaces/supersedes it.
replace: Dict[str, List[Argument]]
# List of precomputed args added without replacement
add: List[Argument]
@staticmethod
def parse(src: object) -> "Precompute":
assert isinstance(src, list)
# src is a list of strings of the format:
# {kernel param name} -> {replacement decl}[, {replacement decl}, ...]
# [{add decl}[, {add decl}, ...]]
# The last line is optional and contains the precomputed parameters that are
# added without replacement.
# The other lines are parsed to get the names of which precomputed elements
# should replace which kernel arguments.
add_args = []
if " -> " not in src[-1]:
add_list = src[-1].split(",")
add_args = [Argument.parse(name.strip()) for name in add_list]
src = src[:-1]
replace = {}
for raw_replace_item in src:
assert isinstance(raw_replace_item, str)
assert " -> " in raw_replace_item, (
"precomputed parameters without replacement"
" are allowed only in the last line"
)
arg, with_list_raw = raw_replace_item.split(" -> ")
with_list = with_list_raw.split(",")
with_list_args = [Argument.parse(name.strip()) for name in with_list]
replace[arg] = with_list_args
r = Precompute(replace=replace, add=add_args)
assert r.to_list() == src, "r.to_list() != src"
return r
def __post_init__(self) -> None:
# the template parameters are upper so if these are the
# same then it is ambiguous
for a in self.add:
assert a.name.upper() != a.name
for args in self.replace.values():
for a in args:
assert a.name.upper() != a.name
def to_list(self) -> List[str]:
replace_list = []
for kernel_param, replacement_params in self.replace.items():
replacements = ", ".join(str(param) for param in replacement_params)
replace_list.append(f"{kernel_param} -> {replacements}")
return replace_list